Port SAC WIP (#581)

Co-authored-by: KeWang1017 <ke.wang@helloleap.ai>

.cache/calibration/aloha_default/left_follower.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2015,
+        3058,
+        3061,
+        1071,
+        1071,
+        2035,
+        2152,
+        2029,
+        2499
+    ],
+    "end_pos": [
+        -1008,
+        -1963,
+        -1966,
+        2141,
+        2143,
+        -971,
+        3043,
+        -1077,
+        3144
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/left_leader.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -1024
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2035,
+        3024,
+        3019,
+        979,
+        981,
+        1982,
+        2166,
+        2124,
+        1968
+    ],
+    "end_pos": [
+        -990,
+        -2017,
+        -2015,
+        2078,
+        2076,
+        -1030,
+        3117,
+        -1016,
+        2556
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/right_follower.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2056,
+        2895,
+        2896,
+        1191,
+        1190,
+        2018,
+        2051,
+        2056,
+        2509
+    ],
+    "end_pos": [
+        -1040,
+        -2004,
+        -2006,
+        2126,
+        2127,
+        -1010,
+        3050,
+        -1117,
+        3143
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.cache/calibration/aloha_default/right_leader.json

@@ -0,0 +1,68 @@
+{
+    "homing_offset": [
+        2048,
+        3072,
+        3072,
+        -1024,
+        -1024,
+        2048,
+        -2048,
+        2048,
+        -2048
+    ],
+    "drive_mode": [
+        1,
+        1,
+        1,
+        0,
+        0,
+        1,
+        0,
+        1,
+        0
+    ],
+    "start_pos": [
+        2068,
+        3034,
+        3030,
+        1038,
+        1041,
+        1991,
+        1948,
+        2090,
+        1985
+    ],
+    "end_pos": [
+        -1025,
+        -2014,
+        -2015,
+        2058,
+        2060,
+        -955,
+        3091,
+        -940,
+        2576
+    ],
+    "calib_mode": [
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "DEGREE",
+        "LINEAR"
+    ],
+    "motor_names": [
+        "waist",
+        "shoulder",
+        "shoulder_shadow",
+        "elbow",
+        "elbow_shadow",
+        "forearm_roll",
+        "wrist_angle",
+        "wrist_rotate",
+        "gripper"
+    ]
+}

.dockerignore

@@ -65,7 +65,6 @@ htmlcov/
 .nox/
 .coverage
 .coverage.*
-.cache
 nosetests.xml
 coverage.xml
 *.cover
@@ -73,6 +72,11 @@ coverage.xml
 .hypothesis/
 .pytest_cache/
 
+# Ignore .cache except calibration
+.cache/*
+!.cache/calibration/
+!.cache/calibration/**
+
 # Translations
 *.mo
 *.pot

.gitattributes

@@ -1,2 +1,6 @@
 *.memmap filter=lfs diff=lfs merge=lfs -text
 *.stl filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.json !text !filter !merge !diff

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,11 +1,15 @@
-# What does this PR do?
+## What this does
+Explain what this PR does. Feel free to tag your PR with the appropriate label(s).
 
 Examples:
-- Fixes # (issue)
-- Adds new dataset
-- Optimizes something
+|  Title               | Label           |
+|----------------------|-----------------|
+| Fixes #[issue]       | (🐛 Bug)        |
+| Adds new dataset     | (🗃️ Dataset)    |
+| Optimizes something  | (⚡️ Performance) |
 
-## How was it tested?
+## How it was tested
+Explain/show how you tested your changes.
 
 Examples:
 - Added `test_something` in `tests/test_stuff.py`.
@@ -13,20 +17,18 @@ Examples:
 - Optimized `some_function`, it now runs X times faster than previously.
 
 ## How to checkout & try? (for the reviewer)
+Provide a simple way for the reviewer to try out your changes.
 
 Examples:
 ```bash
-DATA_DIR=tests/data pytest -sx tests/test_stuff.py::test_something
+pytest -sx tests/test_stuff.py::test_something
 ```
 ```bash
 python lerobot/scripts/train.py --some.option=true
 ```
 
-## Before submitting
-Please read the [contributor guideline](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md#submitting-a-pull-request-pr).
-
-
-## Who can review?
-
-Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
+## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR
+**Note**: Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
 members/contributors who may be interested in your PR. Try to avoid tagging more than 3 people.
+
+**Note**: Before submitting this PR, please read the [contributor guideline](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md#submitting-a-pull-request-pr).

.github/scripts/dep_build.py

@@ -1,30 +0,0 @@
-PYPROJECT = "pyproject.toml"
-DEPS = {
-    "gym-pusht": '{ git = "git@github.com:huggingface/gym-pusht.git", optional = true}',
-    "gym-xarm": '{ git = "git@github.com:huggingface/gym-xarm.git", optional = true}',
-    "gym-aloha": '{ git = "git@github.com:huggingface/gym-aloha.git", optional = true}',
-}
-
-
-def update_envs_as_path_dependencies():
-    with open(PYPROJECT) as file:
-        lines = file.readlines()
-
-    new_lines = []
-    for line in lines:
-        if any(dep in line for dep in DEPS.values()):
-            for dep in DEPS:
-                if dep in line:
-                    new_line = f'{dep} = {{ path = "envs/{dep}/", optional = true}}\n'
-                    new_lines.append(new_line)
-                    break
-
-        else:
-            new_lines.append(line)
-
-    with open(PYPROJECT, "w") as file:
-        file.writelines(new_lines)
-
-
-if __name__ == "__main__":
-    update_envs_as_path_dependencies()

.github/workflows/build-docker-images.yml

@@ -10,63 +10,26 @@ on:
 
 env:
   PYTHON_VERSION: "3.10"
-#   CI_SLACK_CHANNEL: ${{ secrets.CI_DOCKER_CHANNEL }}
 
 jobs:
   latest-cpu:
-    name: "Build CPU"
-    runs-on: ubuntu-latest
+    name: CPU
+    runs-on:
+      group: aws-general-8-plus
     steps:
-      - name: Cleanup disk
+      - name: Install Git LFS
         run: |
-          sudo df -h
-          # sudo ls -l /usr/local/lib/
-          # sudo ls -l /usr/share/
-          sudo du -sh /usr/local/lib/
-          sudo du -sh /usr/share/
-          sudo rm -rf /usr/local/lib/android
-          sudo rm -rf /usr/share/dotnet
-          sudo du -sh /usr/local/lib/
-          sudo du -sh /usr/share/
-          sudo df -h
+          sudo apt-get update
+          sudo apt-get install git-lfs
+          git lfs install
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
 
       - name: Check out code
         uses: actions/checkout@v4
-
-      # HACK(aliberts): to be removed for release
-      # -----------------------------------------
-      - name: Checkout gym-aloha
-        uses: actions/checkout@v4
         with:
-          repository: huggingface/gym-aloha
-          path: envs/gym-aloha
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-xarm
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-xarm
-          path: envs/gym-xarm
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-pusht
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-pusht
-          path: envs/gym-pusht
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-
-      - name: Change envs dependencies as local path
-        run: python .github/scripts/dep_build.py
-      # -----------------------------------------
+          lfs: true
 
       - name: Login to DockerHub
         uses: docker/login-action@v3
@@ -83,84 +46,25 @@ jobs:
           tags: huggingface/lerobot-cpu
           build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
 
-    #   - name: Post to a Slack channel
-    #     id: slack
-    #     #uses: slackapi/slack-github-action@v1.25.0
-    #     uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
-    #     with:
-    #       # Slack channel id, channel name, or user id to post message.
-    #       # See also: https://api.slack.com/methods/chat.postMessage#channels
-    #       channel-id: ${{ env.CI_SLACK_CHANNEL }}
-    #       # For posting a rich message using Block Kit
-    #       payload: |
-    #         {
-    #           "text": "lerobot-cpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}",
-    #           "blocks": [
-    #             {
-    #               "type": "section",
-    #               "text": {
-    #                 "type": "mrkdwn",
-    #                 "text": "lerobot-cpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}"
-    #               }
-    #             }
-    #           ]
-    #         }
-    #     env:
-    #       SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
 
   latest-cuda:
-    name: "Build GPU"
-    runs-on: ubuntu-latest
+    name: GPU
+    runs-on:
+      group: aws-general-8-plus
     steps:
-      - name: Cleanup disk
+      - name: Install Git LFS
         run: |
-          sudo df -h
-          # sudo ls -l /usr/local/lib/
-          # sudo ls -l /usr/share/
-          sudo du -sh /usr/local/lib/
-          sudo du -sh /usr/share/
-          sudo rm -rf /usr/local/lib/android
-          sudo rm -rf /usr/share/dotnet
-          sudo du -sh /usr/local/lib/
-          sudo du -sh /usr/share/
-          sudo df -h
+          sudo apt-get update
+          sudo apt-get install git-lfs
+          git lfs install
+
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
 
       - name: Check out code
         uses: actions/checkout@v4
-
-      # HACK(aliberts): to be removed for release
-      # -----------------------------------------
-      - name: Checkout gym-aloha
-        uses: actions/checkout@v4
         with:
-          repository: huggingface/gym-aloha
-          path: envs/gym-aloha
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-xarm
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-xarm
-          path: envs/gym-xarm
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-pusht
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-pusht
-          path: envs/gym-pusht
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-
-      - name: Change envs dependencies as local path
-        run: python .github/scripts/dep_build.py
-      # -----------------------------------------
+          lfs: true
 
       - name: Login to DockerHub
         uses: docker/login-action@v3
@@ -177,27 +81,29 @@ jobs:
           tags: huggingface/lerobot-gpu
           build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
 
-      # - name: Post to a Slack channel
-      #   id: slack
-      #   #uses: slackapi/slack-github-action@v1.25.0
-      #   uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
-      #   with:
-      #     # Slack channel id, channel name, or user id to post message.
-      #     # See also: https://api.slack.com/methods/chat.postMessage#channels
-      #     channel-id: ${{ env.CI_SLACK_CHANNEL }}
-      #     # For posting a rich message using Block Kit
-      #     payload: |
-      #       {
-      #         "text": "lerobot-gpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}",
-      #         "blocks": [
-      #           {
-      #             "type": "section",
-      #             "text": {
-      #               "type": "mrkdwn",
-      #               "text": "lerobot-gpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}"
-      #             }
-      #           }
-      #         ]
-      #       }
-      #   env:
-      #     SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
+
+  latest-cuda-dev:
+    name: GPU Dev
+    runs-on:
+      group: aws-general-8-plus
+    steps:
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU dev
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-gpu-dev/Dockerfile
+          push: true
+          tags: huggingface/lerobot-gpu:dev
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}

.github/workflows/nightly-tests.yml

@@ -7,16 +7,15 @@ on:
   schedule:
     - cron: "0 2 * * *"
 
-env:
-  DATA_DIR: tests/data
+# env:
   # SLACK_API_TOKEN: ${{ secrets.SLACK_API_TOKEN }}
-
 jobs:
   run_all_tests_cpu:
-    name: "Test CPU"
+    name: CPU
     strategy:
       fail-fast: false
-    runs-on: ubuntu-latest
+    runs-on:
+      group: aws-general-8-plus
     container:
       image: huggingface/lerobot-cpu:latest
       options: --shm-size "16gb"
@@ -29,21 +28,18 @@ jobs:
         working-directory: /lerobot
     steps:
       - name: Tests
-        env:
-          DATA_DIR: tests/data
         run: pytest -v --cov=./lerobot --disable-warnings tests
 
       - name: Tests end-to-end
-        env:
-          DATA_DIR: tests/data
         run: make test-end-to-end
 
 
   run_all_tests_single_gpu:
-    name: "Test GPU"
+    name: GPU
     strategy:
       fail-fast: false
-    runs-on: [single-gpu, nvidia-gpu, t4, ci]
+    runs-on:
+      group: aws-g6-4xlarge-plus
     env:
       CUDA_VISIBLE_DEVICES: "0"
       TEST_TYPE: "single_gpu"
@@ -70,6 +66,8 @@ jobs:
       #     files: ./coverage.xml
       #     verbose: true
       - name: Tests end-to-end
+        env:
+          DEVICE: cuda
         run: make test-end-to-end
 
     #   - name: Generate Report

.github/workflows/quality.yml

@@ -0,0 +1,84 @@
+name: Quality
+
+on:
+  workflow_dispatch:
+  workflow_call:
+  pull_request:
+    branches:
+      - main
+  push:
+    branches:
+      - main
+
+env:
+  PYTHON_VERSION: "3.10"
+
+jobs:
+  style:
+    name: Style
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ env.PYTHON_VERSION }}
+
+      - name: Get Ruff Version from pre-commit-config.yaml
+        id: get-ruff-version
+        run: |
+          RUFF_VERSION=$(awk '/repo: https:\/\/github.com\/astral-sh\/ruff-pre-commit/{flag=1;next}/rev:/{if(flag){print $2;exit}}' .pre-commit-config.yaml)
+          echo "RUFF_VERSION=${RUFF_VERSION}" >> $GITHUB_ENV
+
+      - name: Install Ruff
+        run: python -m pip install "ruff==${{ env.RUFF_VERSION }}"
+
+      - name: Ruff check
+        run: ruff check
+
+      - name: Ruff format
+        run: ruff format --diff
+
+
+  poetry_check:
+    name: Poetry check
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Install poetry
+        run: pipx install poetry
+
+      - name: Poetry check
+        run: poetry check
+
+
+  poetry_relax:
+    name: Poetry relax
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Install poetry
+        run: pipx install poetry
+
+      - name: Install poetry-relax
+        run: poetry self add poetry-relax
+
+      - name: Poetry relax
+        id: poetry_relax
+        run: |
+          output=$(poetry relax --check 2>&1)
+          if echo "$output" | grep -q "Proposing updates"; then
+            echo "$output"
+            echo ""
+            echo "Some dependencies have caret '^' version requirement added by poetry by default."
+            echo "Please replace them with '>='. You can do this by hand or use poetry-relax to do this."
+            exit 1
+          else
+            echo "$output"
+          fi

.github/workflows/style.yml

@@ -1,38 +0,0 @@
-name: Style
-
-on:
-  workflow_dispatch:
-  workflow_call:
-  pull_request:
-    branches:
-      - main
-  push:
-    branches:
-      - main
-
-env:
-  PYTHON_VERSION: "3.10"
-
-jobs:
-  ruff_check:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout Repository
-        uses: actions/checkout@v3
-
-      - name: Set up Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: ${{ env.PYTHON_VERSION }}
-
-      - name: Get Ruff Version from pre-commit-config.yaml
-        id: get-ruff-version
-        run: |
-          RUFF_VERSION=$(awk '/repo: https:\/\/github.com\/astral-sh\/ruff-pre-commit/{flag=1;next}/rev:/{if(flag){print $2;exit}}' .pre-commit-config.yaml)
-          echo "RUFF_VERSION=${RUFF_VERSION}" >> $GITHUB_ENV
-
-      - name: Install Ruff
-        run: python -m pip install "ruff==${{ env.RUFF_VERSION }}"
-
-      - name: Run Ruff
-        run: ruff check .

.github/workflows/test-docker-build.yml

@@ -1,6 +1,6 @@
 # Inspired by
 # https://github.com/huggingface/peft/blob/main/.github/workflows/test-docker-build.yml
-name: Test Docker builds (PR)
+name: Test Dockerfiles
 
 on:
   pull_request:
@@ -15,7 +15,7 @@ env:
 
 jobs:
   get_changed_files:
-    name: "Get all modified Dockerfiles"
+    name: Detect modified Dockerfiles
     runs-on: ubuntu-latest
     outputs:
       matrix: ${{ steps.set-matrix.outputs.matrix }}
@@ -40,66 +40,22 @@ jobs:
 
 
   build_modified_dockerfiles:
-    name: "Build all modified Docker images"
+    name: Build modified Docker images
     needs: get_changed_files
-    runs-on: ubuntu-latest
+    runs-on:
+      group: aws-general-8-plus
     if: ${{ needs.get_changed_files.outputs.matrix }} != ''
     strategy:
       fail-fast: false
       matrix:
         docker-file: ${{ fromJson(needs.get_changed_files.outputs.matrix) }}
     steps:
-      - name: Cleanup disk
-        run: |
-          sudo df -h
-          # sudo ls -l /usr/local/lib/
-          # sudo ls -l /usr/share/
-          sudo du -sh /usr/local/lib/
-          sudo du -sh /usr/share/
-          sudo rm -rf /usr/local/lib/android
-          sudo rm -rf /usr/share/dotnet
-          sudo du -sh /usr/local/lib/
-          sudo du -sh /usr/share/
-          sudo df -h
-
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
 
       - name: Check out code
         uses: actions/checkout@v4
 
-      # HACK(aliberts): to be removed for release
-      # -----------------------------------------
-      - name: Checkout gym-aloha
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-aloha
-          path: envs/gym-aloha
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-xarm
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-xarm
-          path: envs/gym-xarm
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-pusht
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-pusht
-          path: envs/gym-pusht
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-
-      - name: Change envs dependencies as local path
-        run: python .github/scripts/dep_build.py
-      # -----------------------------------------
-
       - name: Build Docker image
         uses: docker/build-push-action@v5
         with:

.github/workflows/test.yml

@@ -10,6 +10,8 @@ on:
       - "examples/**"
       - ".github/**"
       - "poetry.lock"
+      - "Makefile"
+      - ".cache/**"
   push:
     branches:
       - main
@@ -19,29 +21,32 @@ on:
       - "examples/**"
       - ".github/**"
       - "poetry.lock"
+      - "Makefile"
+      - ".cache/**"
 
 jobs:
-  tests:
+  pytest:
+    name: Pytest
     runs-on: ubuntu-latest
     env:
-      DATA_DIR: tests/data
       MUJOCO_GL: egl
     steps:
-      - name: Add SSH key for installing envs
-        uses: webfactory/ssh-agent@v0.9.0
-        with:
-          ssh-private-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
       - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
 
-      - name: Install EGL
-        run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
+      - name: Install apt dependencies
+      # portaudio19-dev is needed to install pyaudio
+        run: |
+          sudo apt-get update && \
+          sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
 
       - name: Install poetry
         run: |
           pipx install poetry && poetry config virtualenvs.in-project true
           echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
 
+      # TODO(rcadene, aliberts): python 3.12 seems to be used in the tests, not python 3.10
       - name: Set up Python 3.10
         uses: actions/setup-python@v5
         with:
@@ -60,7 +65,75 @@ jobs:
             -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
             && rm -rf tests/outputs outputs
 
-      - name: Test end-to-end
+  pytest-minimal:
+    name: Pytest (minimal install)
+    runs-on: ubuntu-latest
+    env:
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          lfs: true  # Ensure LFS files are pulled
+
+      - name: Install apt dependencies
+        run: sudo apt-get update && sudo apt-get install -y ffmpeg
+
+      - name: Install poetry
         run: |
-          make test-end-to-end \
-            && rm -rf outputs
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+
+      # TODO(rcadene, aliberts): python 3.12 seems to be used in the tests, not python 3.10
+      - name: Set up Python 3.10
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+
+      - name: Install poetry dependencies
+        run: |
+          poetry install --extras "test"
+
+      - name: Test with pytest
+        run: |
+          pytest tests -v --cov=./lerobot --durations=0 \
+            -W ignore::DeprecationWarning:imageio_ffmpeg._utils:7 \
+            -W ignore::UserWarning:torch.utils.data.dataloader:558 \
+            -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
+            && rm -rf tests/outputs outputs
+
+  # TODO(aliberts, rcadene): redesign after v2 migration / removing hydra
+  # end-to-end:
+  #   name: End-to-end
+  #   runs-on: ubuntu-latest
+  #   env:
+  #     MUJOCO_GL: egl
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #       with:
+  #         lfs: true  # Ensure LFS files are pulled
+
+  #     - name: Install apt dependencies
+  #     # portaudio19-dev is needed to install pyaudio
+  #       run: |
+  #         sudo apt-get update && \
+  #         sudo apt-get install -y libegl1-mesa-dev portaudio19-dev
+
+  #     - name: Install poetry
+  #       run: |
+  #         pipx install poetry && poetry config virtualenvs.in-project true
+  #         echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+
+  #     - name: Set up Python 3.10
+  #       uses: actions/setup-python@v5
+  #       with:
+  #         python-version: "3.10"
+  #         cache: "poetry"
+
+  #     - name: Install poetry dependencies
+  #       run: |
+  #         poetry install --all-extras
+
+  #     - name: Test end-to-end
+  #       run: |
+  #         make test-end-to-end \
+  #           && rm -rf outputs

.github/workflows/trufflehog.yml

@@ -0,0 +1,20 @@
+on:
+  push:
+
+name: Secret Leaks
+
+permissions:
+  contents: read
+
+jobs:
+  trufflehog:
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v4
+      with:
+        fetch-depth: 0
+    - name: Secret Scanning
+      uses: trufflesecurity/trufflehog@main
+      with:
+        extra_args: --only-verified

.gitignore

@@ -2,10 +2,16 @@
 logs
 tmp
 wandb
+
+# Data
 data
 outputs
+
+# Apple
+.DS_Store
+
+# VS Code
 .vscode
-rl
 
 # HPC
 nautilus/*.yaml
@@ -60,7 +66,6 @@ htmlcov/
 .nox/
 .coverage
 .coverage.*
-.cache
 nosetests.xml
 coverage.xml
 *.cover
@@ -68,6 +73,11 @@ coverage.xml
 .hypothesis/
 .pytest_cache/
 
+# Ignore .cache except calibration
+.cache/*
+!.cache/calibration/
+!.cache/calibration/**
+
 # Translations
 *.mo
 *.pot
@@ -89,6 +99,7 @@ instance/
 docs/_build/
 
 # PyBuilder
+.pybuilder/
 target/
 
 # Jupyter Notebook
@@ -101,13 +112,6 @@ ipython_config.py
 # pyenv
 .python-version
 
-# pipenv
-#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
-#   However, in case of collaboration, if having platform-specific dependencies or dependencies
-#   having no cross-platform support, pipenv may install dependencies that don't work, or not
-#   install all needed dependencies.
-#Pipfile.lock
-
 # PEP 582; used by e.g. github.com/David-OConnor/pyflow
 __pypackages__/
 
@@ -118,6 +122,14 @@ celerybeat.pid
 # SageMath parsed files
 *.sage.py
 
+# Environments
+.env
+.venv
+env/
+venv/
+env.bak/
+venv.bak/
+
 # Spyder project settings
 .spyderproject
 .spyproject
@@ -135,3 +147,9 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/

.pre-commit-config.yaml

@@ -3,7 +3,7 @@ default_language_version:
     python: python3.10
 repos:
   - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.6.0
+    rev: v5.0.0
     hooks:
       - id: check-added-large-files
       - id: debug-statements
@@ -14,11 +14,11 @@ repos:
       - id: end-of-file-fixer
       - id: trailing-whitespace
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.15.2
+    rev: v3.19.0
     hooks:
     -   id: pyupgrade
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.4.2
+    rev: v0.8.2
     hooks:
       - id: ruff
         args: [--fix]
@@ -31,3 +31,7 @@ repos:
         args:
           - "--check"
           - "--no-update"
+  - repo: https://github.com/gitleaks/gitleaks
+    rev: v8.21.2
+    hooks:
+      - id: gitleaks

CONTRIBUTING.md

@@ -20,7 +20,7 @@ Some of the ways you can contribute to 🤗 LeRobot:
 * Contributing to the examples or to the documentation.
 * Submitting issues related to bugs or desired new features.
 
-Following the guides below, feel free to open issues and PRs and to coordinate your efforts with the community on our [Discord Channel](https://discord.gg/VjFz58wn3R). For specific inquiries, reach out to [Remi Cadene](remi.cadene@huggingface.co).
+Following the guides below, feel free to open issues and PRs and to coordinate your efforts with the community on our [Discord Channel](https://discord.gg/VjFz58wn3R). For specific inquiries, reach out to [Remi Cadene](mailto:remi.cadene@huggingface.co).
 
 If you are not sure how to contribute or want to know the next features we working on, look on this project page: [LeRobot TODO](https://github.com/orgs/huggingface/projects/46)
 
@@ -195,6 +195,11 @@ Follow these steps to start contributing:
    git commit
    ```
 
+   Note, if you already commited some changes that have a wrong formatting, you can use:
+   ```bash
+   pre-commit run --all-files
+   ```
+
    Please write [good commit messages](https://chris.beams.io/posts/git-commit/).
 
    It is a good idea to sync your copy of the code with the original
@@ -262,7 +267,7 @@ We use `pytest` in order to run the tests. From the root of the
 repository, here's how to run tests with `pytest` for the library:
 
 ```bash
-DATA_DIR="tests/data" python -m pytest -sv ./tests
+python -m pytest -sv ./tests
 ```
 
 

Makefile

@@ -5,11 +5,12 @@ PYTHON_PATH := $(shell which python)
 # If Poetry is installed, redefine PYTHON_PATH to use the Poetry-managed Python
 POETRY_CHECK := $(shell command -v poetry)
 ifneq ($(POETRY_CHECK),)
-    PYTHON_PATH := $(shell poetry run which python)
+	PYTHON_PATH := $(shell poetry run which python)
 endif
 
 export PATH := $(dir $(PYTHON_PATH)):$(PATH)
 
+DEVICE ?= cpu
 
 build-cpu:
 	docker build -t lerobot:latest -f docker/lerobot-cpu/Dockerfile .
@@ -18,78 +19,172 @@ build-gpu:
 	docker build -t lerobot:latest -f docker/lerobot-gpu/Dockerfile .
 
 test-end-to-end:
-	${MAKE} test-act-ete-train
-	${MAKE} test-act-ete-eval
-	${MAKE} test-diffusion-ete-train
-	${MAKE} test-diffusion-ete-eval
-	${MAKE} test-tdmpc-ete-train
-	${MAKE} test-tdmpc-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train-amp
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-eval-amp
+	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train-with-online
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-default-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-pusht-tutorial
 
 test-act-ete-train:
 	python lerobot/scripts/train.py \
 		policy=act \
+		policy.dim_model=64 \
 		env=aloha \
 		wandb.enable=False \
-		offline_steps=2 \
-		online_steps=0 \
-		eval_episodes=1 \
-		device=cpu \
-		save_model=true \
-		save_freq=2 \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
 		policy.n_action_steps=20 \
 		policy.chunk_size=20 \
-		policy.batch_size=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
 		hydra.run.dir=tests/outputs/act/
 
 test-act-ete-eval:
 	python lerobot/scripts/eval.py \
-		--config tests/outputs/act/.hydra/config.yaml \
-		eval_episodes=1 \
+		-p tests/outputs/act/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
-		policy.pretrained_model_path=tests/outputs/act/models/2.pt
+		device=$(DEVICE) \
+
+test-act-ete-train-amp:
+	python lerobot/scripts/train.py \
+		policy=act \
+		policy.dim_model=64 \
+		env=aloha \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		policy.n_action_steps=20 \
+		policy.chunk_size=20 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/act_amp/ \
+		training.image_transforms.enable=true \
+		use_amp=true
+
+test-act-ete-eval-amp:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/act_amp/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+		use_amp=true
 
 test-diffusion-ete-train:
 	python lerobot/scripts/train.py \
 		policy=diffusion \
+		policy.down_dims=\[64,128,256\] \
+		policy.diffusion_step_embed_dim=32 \
+		policy.num_inference_steps=10 \
 		env=pusht \
 		wandb.enable=False \
-		offline_steps=2 \
-		online_steps=0 \
-		eval_episodes=1 \
-		device=cpu \
-		save_model=true \
-		save_freq=2 \
-		policy.batch_size=2 \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
 		hydra.run.dir=tests/outputs/diffusion/
 
 test-diffusion-ete-eval:
 	python lerobot/scripts/eval.py \
-		--config tests/outputs/diffusion/.hydra/config.yaml \
-		eval_episodes=1 \
+		-p tests/outputs/diffusion/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
-		policy.pretrained_model_path=tests/outputs/diffusion/models/2.pt
+		device=$(DEVICE) \
 
 test-tdmpc-ete-train:
 	python lerobot/scripts/train.py \
 		policy=tdmpc \
 		env=xarm \
+		env.task=XarmLift-v0 \
+		dataset_repo_id=lerobot/xarm_lift_medium \
 		wandb.enable=False \
-		offline_steps=1 \
-		online_steps=2 \
-		eval_episodes=1 \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		env.episode_length=2 \
-		device=cpu \
-		save_model=true \
-		save_freq=2 \
-		policy.batch_size=2 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
 		hydra.run.dir=tests/outputs/tdmpc/
 
+test-tdmpc-ete-train-with-online:
+	python lerobot/scripts/train.py \
+		env=pusht \
+		env.gym.obs_type=environment_state_agent_pos \
+		policy=tdmpc_pusht_keypoints \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=10 \
+		device=$(DEVICE) \
+		training.offline_steps=2 \
+		training.online_steps=20 \
+		training.save_checkpoint=false \
+		training.save_freq=10 \
+		training.batch_size=2 \
+		training.online_rollout_n_episodes=2 \
+		training.online_rollout_batch_size=2 \
+		training.online_steps_between_rollouts=10 \
+		training.online_buffer_capacity=15 \
+		eval.use_async_envs=true \
+		hydra.run.dir=tests/outputs/tdmpc_online/
+
+
 test-tdmpc-ete-eval:
 	python lerobot/scripts/eval.py \
-		--config tests/outputs/tdmpc/.hydra/config.yaml \
-		eval_episodes=1 \
+		-p tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
-		policy.pretrained_model_path=tests/outputs/tdmpc/models/2.pt
+		device=$(DEVICE) \
+
+test-default-ete-eval:
+	python lerobot/scripts/eval.py \
+		--config lerobot/configs/default.yaml \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+
+test-act-pusht-tutorial:
+	cp examples/advanced/1_train_act_pusht/act_pusht.yaml lerobot/configs/policy/created_by_Makefile.yaml
+	python lerobot/scripts/train.py \
+		policy=created_by_Makefile.yaml \
+		env=pusht \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=2 \
+		device=$(DEVICE) \
+		training.save_model=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		training.image_transforms.enable=true \
+		hydra.run.dir=tests/outputs/act_pusht/
+	rm lerobot/configs/policy/created_by_Makefile.yaml

README.md

@@ -22,28 +22,42 @@
 
 </div>
 
+<h2 align="center">
+    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">New robot in town: SO-100</a></p>
+</h2>
+
+<div align="center">
+    <img src="media/so100/leader_follower.webp?raw=true" alt="SO-100 leader and follower arms" title="SO-100 leader and follower arms" width="50%">
+    <p>We just added a new tutorial on how to build a more affordable robot, at the price of $110 per arm!</p>
+    <p>Teach it new skills by showing it a few moves with just a laptop.</p>
+    <p>Then watch your homemade robot act autonomously 🤯</p>
+    <p>Follow the link to the <a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">full tutorial for SO-100</a>.</p>
+</div>
+
+<br/>
+
 <h3 align="center">
-    <p>State-of-the-art Machine Learning for real-world robotics</p>
+    <p>LeRobot: State-of-the-art AI for real-world robotics</p>
 </h3>
 
 ---
 
 
-🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier for entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
+🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
 
 🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
 
-🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulated environments so that everyone can get started. In the coming weeks, the plan is to add more and more support for real-world robotics on the most affordable and capable robots out there.
+🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulation environments to get started without assembling a robot. In the coming weeks, the plan is to add more and more support for real-world robotics on the most affordable and capable robots out there.
 
-🤗 LeRobot hosts pretrained models and datasets on this HuggingFace community page: [huggingface.co/lerobot](https://huggingface.co/lerobot)
+🤗 LeRobot hosts pretrained models and datasets on this Hugging Face community page: [huggingface.co/lerobot](https://huggingface.co/lerobot)
 
-#### Examples of pretrained models and environments
+#### Examples of pretrained models on simulation environments
 
 <table>
   <tr>
-    <td><img src="http://remicadene.com/assets/gif/aloha_act.gif" width="100%" alt="ACT policy on ALOHA env"/></td>
-    <td><img src="http://remicadene.com/assets/gif/simxarm_tdmpc.gif" width="100%" alt="TDMPC policy on SimXArm env"/></td>
-    <td><img src="http://remicadene.com/assets/gif/pusht_diffusion.gif" width="100%" alt="Diffusion policy on PushT env"/></td>
+    <td><img src="media/gym/aloha_act.gif" width="100%" alt="ACT policy on ALOHA env"/></td>
+    <td><img src="media/gym/simxarm_tdmpc.gif" width="100%" alt="TDMPC policy on SimXArm env"/></td>
+    <td><img src="media/gym/pusht_diffusion.gif" width="100%" alt="Diffusion policy on PushT env"/></td>
   </tr>
   <tr>
     <td align="center">ACT policy on ALOHA env</td>
@@ -54,28 +68,36 @@
 
 ### Acknowledgment
 
-- ACT policy and ALOHA environment are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha/)
-- Diffusion policy and Pusht environment are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu/)
-- TDMPC policy and Simxarm environment are adapted from [FOWM](https://www.yunhaifeng.com/FOWM/)
-- Abstractions and utilities for Reinforcement Learning come from [TorchRL](https://github.com/pytorch/rl)
+- Thanks to Tony Zaho, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
+- Thanks to Cheng Chi, Zhenjia Xu and colleagues for open sourcing Diffusion policy, Pusht environment and datasets, as well as UMI datasets. Ours are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu) and [UMI Gripper](https://umi-gripper.github.io).
+- Thanks to Nicklas Hansen, Yunhai Feng and colleagues for open sourcing TDMPC policy, Simxarm environments and datasets. Ours are adapted from [TDMPC](https://github.com/nicklashansen/tdmpc) and [FOWM](https://www.yunhaifeng.com/FOWM).
+- Thanks to Antonio Loquercio and Ashish Kumar for their early support.
+- Thanks to [Seungjae (Jay) Lee](https://sjlee.cc/), [Mahi Shafiullah](https://mahis.life/) and colleagues for open sourcing [VQ-BeT](https://sjlee.cc/vq-bet/) policy and helping us adapt the codebase to our repository. The policy is adapted from [VQ-BeT repo](https://github.com/jayLEE0301/vq_bet_official).
+
 
 ## Installation
 
 Download our source code:
 ```bash
-git clone https://github.com/huggingface/lerobot.git && cd lerobot
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
 ```
 
 Create a virtual environment with Python 3.10 and activate it, e.g. with [`miniconda`](https://docs.anaconda.com/free/miniconda/index.html):
 ```bash
-conda create -y -n lerobot python=3.10 && conda activate lerobot
+conda create -y -n lerobot python=3.10
+conda activate lerobot
 ```
 
 Install 🤗 LeRobot:
 ```bash
-pip install .
+pip install -e .
 ```
 
+> **NOTE:** Depending on your platform, If you encounter any build errors during this step
+you may need to install `cmake` and `build-essential` for building some of our dependencies.
+On linux: `sudo apt-get install cmake build-essential`
+
 For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
 - [aloha](https://github.com/huggingface/gym-aloha)
 - [xarm](https://github.com/huggingface/gym-xarm)
@@ -83,18 +105,22 @@ For simulations, 🤗 LeRobot comes with gymnasium environments that can be inst
 
 For instance, to install 🤗 LeRobot with aloha and pusht, use:
 ```bash
-pip install ".[aloha, pusht]"
+pip install -e ".[aloha, pusht]"
 ```
 
-To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiments tracking, log in with
+To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
 ```bash
 wandb login
 ```
 
+(note: you will also need to enable WandB in the configuration. See below.)
+
 ## Walkthrough
 
 ```
 .
+├── examples             # contains demonstration examples, start here to learn about LeRobot
+|   └── advanced         # contains even more examples for those who have mastered the basics
 ├── lerobot
 |   ├── configs          # contains hydra yaml files with all options that you can override in the command line
 |   |   ├── default.yaml   # selected by default, it loads pusht environment and diffusion policy
@@ -103,69 +129,174 @@ wandb login
 |   ├── common           # contains classes and utilities
 |   |   ├── datasets       # various datasets of human demonstrations: aloha, pusht, xarm
 |   |   ├── envs           # various sim environments: aloha, pusht, xarm
-|   |   └── policies       # various policies: act, diffusion, tdmpc
-|   └── scripts                  # contains functions to execute via command line
-|       ├── visualize_dataset.py  # load a dataset and render its demonstrations
-|       ├── eval.py               # load policy and evaluate it on an environment
-|       └── train.py              # train a policy via imitation learning and/or reinforcement learning
+|   |   ├── policies       # various policies: act, diffusion, tdmpc
+|   |   ├── robot_devices  # various real devices: dynamixel motors, opencv cameras, koch robots
+|   |   └── utils          # various utilities
+|   └── scripts          # contains functions to execute via command line
+|       ├── eval.py                 # load policy and evaluate it on an environment
+|       ├── train.py                # train a policy via imitation learning and/or reinforcement learning
+|       ├── control_robot.py        # teleoperate a real robot, record data, run a policy
+|       ├── push_dataset_to_hub.py  # convert your dataset into LeRobot dataset format and upload it to the Hugging Face hub
+|       └── visualize_dataset.py    # load a dataset and render its demonstrations
 ├── outputs               # contains results of scripts execution: logs, videos, model checkpoints
-├── .github
-|   └── workflows
-|       └── test.yml      # defines install settings for continuous integration and specifies end-to-end tests
 └── tests                 # contains pytest utilities for continuous integration
-
 ```
 
 ### Visualize datasets
 
-Check out [examples](./examples) to see how you can import our dataset class, download the data from the HuggingFace hub and use our rendering utilities.
+Check out [example 1](./examples/1_load_lerobot_dataset.py) that illustrates how to use our dataset class which automatically downloads data from the Hugging Face hub.
 
-Or you can achieve the same result by executing our script from the command line:
+You can also locally visualize episodes from a dataset on the hub by executing our script from the command line:
 ```bash
 python lerobot/scripts/visualize_dataset.py \
-env=pusht \
-hydra.run.dir=outputs/visualize_dataset/example
-# >>> ['outputs/visualize_dataset/example/episode_0.mp4']
+    --repo-id lerobot/pusht \
+    --episode-index 0
 ```
 
+or from a dataset in a local folder with the `root` option and the `--local-files-only` (in the following case the dataset will be searched for in `./my_local_data_dir/lerobot/pusht`)
+```bash
+python lerobot/scripts/visualize_dataset.py \
+    --repo-id lerobot/pusht \
+    --root ./my_local_data_dir \
+    --local-files-only 1 \
+    --episode-index 0
+```
+
+
+It will open `rerun.io` and display the camera streams, robot states and actions, like this:
+
+https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-fd46b787-b532-47e2-bb6f-fd536a55a7ed.mov?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20240505%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240505T172924Z&X-Amz-Expires=300&X-Amz-Signature=d680b26c532eeaf80740f08af3320d22ad0b8a4e4da1bcc4f33142c15b509eda&X-Amz-SignedHeaders=host&actor_id=24889239&key_id=0&repo_id=748713144
+
+
+Our script can also visualize datasets stored on a distant server. See `python lerobot/scripts/visualize_dataset.py --help` for more instructions.
+
+### The `LeRobotDataset` format
+
+A dataset in `LeRobotDataset` format is very simple to use. It can be loaded from a repository on the Hugging Face hub or a local folder simply with e.g. `dataset = LeRobotDataset("lerobot/aloha_static_coffee")` and can be indexed into like any Hugging Face and PyTorch dataset. For instance `dataset[0]` will retrieve a single temporal frame from the dataset containing observation(s) and an action as PyTorch tensors ready to be fed to a model.
+
+A specificity of `LeRobotDataset` is that, rather than retrieving a single frame by its index, we can retrieve several frames based on their temporal relationship with the indexed frame, by setting `delta_timestamps` to a list of relative times with respect to the indexed frame. For example, with `delta_timestamps = {"observation.image": [-1, -0.5, -0.2, 0]}`  one can retrieve, for a given index, 4 frames: 3 "previous" frames 1 second, 0.5 seconds, and 0.2 seconds before the indexed frame, and the indexed frame itself (corresponding to the 0 entry). See example [1_load_lerobot_dataset.py](examples/1_load_lerobot_dataset.py) for more details on `delta_timestamps`.
+
+Under the hood, the `LeRobotDataset` format makes use of several ways to serialize data which can be useful to understand if you plan to work more closely with this format. We tried to make a flexible yet simple dataset format that would cover most type of features and specificities present in reinforcement learning and robotics, in simulation and in real-world, with a focus on cameras and robot states but easily extended to other types of sensory inputs as long as they can be represented by a tensor.
+
+Here are the important details and internal structure organization of a typical `LeRobotDataset` instantiated with `dataset = LeRobotDataset("lerobot/aloha_static_coffee")`. The exact features will change from dataset to dataset but not the main aspects:
+
+```
+dataset attributes:
+  ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
+  │  ├ observation.images.cam_high (VideoFrame):
+  │  │   VideoFrame = {'path': path to a mp4 video, 'timestamp' (float32): timestamp in the video}
+  │  ├ observation.state (list of float32): position of an arm joints (for instance)
+  │  ... (more observations)
+  │  ├ action (list of float32): goal position of an arm joints (for instance)
+  │  ├ episode_index (int64): index of the episode for this sample
+  │  ├ frame_index (int64): index of the frame for this sample in the episode ; starts at 0 for each episode
+  │  ├ timestamp (float32): timestamp in the episode
+  │  ├ next.done (bool): indicates the end of en episode ; True for the last frame in each episode
+  │  └ index (int64): general index in the whole dataset
+  ├ episode_data_index: contains 2 tensors with the start and end indices of each episode
+  │  ├ from (1D int64 tensor): first frame index for each episode — shape (num episodes,) starts with 0
+  │  └ to: (1D int64 tensor): last frame index for each episode — shape (num episodes,)
+  ├ stats: a dictionary of statistics (max, mean, min, std) for each feature in the dataset, for instance
+  │  ├ observation.images.cam_high: {'max': tensor with same number of dimensions (e.g. `(c, 1, 1)` for images, `(c,)` for states), etc.}
+  │  ...
+  ├ info: a dictionary of metadata on the dataset
+  │  ├ codebase_version (str): this is to keep track of the codebase version the dataset was created with
+  │  ├ fps (float): frame per second the dataset is recorded/synchronized to
+  │  ├ video (bool): indicates if frames are encoded in mp4 video files to save space or stored as png files
+  │  └ encoding (dict): if video, this documents the main options that were used with ffmpeg to encode the videos
+  ├ videos_dir (Path): where the mp4 videos or png images are stored/accessed
+  └ camera_keys (list of string): the keys to access camera features in the item returned by the dataset (e.g. `["observation.images.cam_high", ...]`)
+```
+
+A `LeRobotDataset` is serialised using several widespread file formats for each of its parts, namely:
+- hf_dataset stored using Hugging Face datasets library serialization to parquet
+- videos are stored in mp4 format to save space
+- metadata are stored in plain json/jsonl files
+
+Dataset can be uploaded/downloaded from the HuggingFace hub seamlessly. To work on a local dataset, you can use the `local_files_only` argument and specify its location with the `root` argument if it's not in the default `~/.cache/huggingface/lerobot` location.
+
 ### Evaluate a pretrained policy
 
-Check out [examples](./examples) to see how you can load a pretrained policy from HuggingFace hub, load up the corresponding environment and model, and run an evaluation.
+Check out [example 2](./examples/2_evaluate_pretrained_policy.py) that illustrates how to download a pretrained policy from Hugging Face hub, and run an evaluation on its corresponding environment.
 
-Or you can achieve the same result by executing our script from the command line:
+We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):
 ```bash
 python lerobot/scripts/eval.py \
---hub-id lerobot/diffusion_policy_pusht_image \
-eval_episodes=10 \
-hydra.run.dir=outputs/eval/example_hub
+    -p lerobot/diffusion_pusht \
+    eval.n_episodes=10 \
+    eval.batch_size=10
 ```
 
-After training your own policy, you can also re-evaluate the checkpoints with:
+Note: After training your own policy, you can re-evaluate the checkpoints with:
+
 ```bash
-python lerobot/scripts/eval.py \
---config PATH/TO/FOLDER/config.yaml \
-policy.pretrained_model_path=PATH/TO/FOLDER/weights.pth \
-eval_episodes=10 \
-hydra.run.dir=outputs/eval/example_dir
+python lerobot/scripts/eval.py -p {OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```
 
 See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-Check out [examples](./examples) to see how you can start training a model on a dataset, which will be automatically downloaded if needed.
+Check out [example 3](./examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
+
+In general, you can use our training script to easily train any policy. Here is an example of training the ACT policy on trajectories collected by humans on the Aloha simulation environment for the insertion task:
 
-In general, you can use our training script to easily train any policy on any environment:
 ```bash
 python lerobot/scripts/train.py \
-env=aloha \
-task=sim_insertion \
-repo_id=lerobot/aloha_sim_insertion_scripted \
-policy=act \
-hydra.run.dir=outputs/train/aloha_act
+    policy=act \
+    env=aloha \
+    env.task=AlohaInsertion-v0 \
+    dataset_repo_id=lerobot/aloha_sim_insertion_human \
 ```
 
-After training, you may want to revisit model evaluation to change the evaluation settings. In fact, during training every checkpoint is already evaluated but on a low number of episodes for efficiency. Check out [example](./examples) to evaluate any model checkpoint on more episodes to increase statistical significance.
+The experiment directory is automatically generated and will show up in yellow in your terminal. It looks like `outputs/train/2024-05-05/20-21-12_aloha_act_default`. You can manually specify an experiment directory by adding this argument to the `train.py` python command:
+```bash
+    hydra.run.dir=your/new/experiment/dir
+```
+
+In the experiment directory there will be a folder called `checkpoints` which will have the following structure:
+
+```bash
+checkpoints
+├── 000250  # checkpoint_dir for training step 250
+│   ├── pretrained_model  # Hugging Face pretrained model dir
+│   │   ├── config.json  # Hugging Face pretrained model config
+│   │   ├── config.yaml  # consolidated Hydra config
+│   │   ├── model.safetensors  # model weights
+│   │   └── README.md  # Hugging Face model card
+│   └── training_state.pth  # optimizer/scheduler/rng state and training step
+```
+
+To resume training from a checkpoint, you can add these to the `train.py` python command:
+```bash
+    hydra.run.dir=your/original/experiment/dir resume=true
+```
+
+It will load the pretrained model, optimizer and scheduler states for training. For more information please see our tutorial on training resumption [here](https://github.com/huggingface/lerobot/blob/main/examples/5_resume_training.md).
+
+To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding:
+
+```bash
+    wandb.enable=true
+```
+
+A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of what they look like in your browser. Please also check [here](https://github.com/huggingface/lerobot/blob/main/examples/4_train_policy_with_script.md#typical-logs-and-metrics) for the explanation of some commonly used metrics in logs.
+
+![](media/wandb.png)
+
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
+
+#### Reproduce state-of-the-art (SOTA)
+
+We have organized our configuration files (found under [`lerobot/configs`](./lerobot/configs)) such that they reproduce SOTA results from a given model variant in their respective original works. Simply running:
+
+```bash
+python lerobot/scripts/train.py policy=diffusion env=pusht
+```
+
+reproduces SOTA results for Diffusion Policy on the PushT task.
+
+Pretrained policies, along with reproduction details, can be found under the "Models" section of https://huggingface.co/lerobot.
 
 ## Contribute
 
@@ -173,105 +304,40 @@ If you would like to contribute to 🤗 LeRobot, please check out our [contribut
 
 ### Add a new dataset
 
-```python
-# TODO(rcadene, AdilZouitine): rewrite this section
-```
-
-To add a dataset to the hub, first login and use a token generated from [huggingface settings](https://huggingface.co/settings/tokens) with write access:
+To add a dataset to the hub, you need to login using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
 ```bash
 huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
-Then you can upload it to the hub with:
+Then point to your raw dataset folder (e.g. `data/aloha_static_pingpong_test_raw`), and push your dataset to the hub with:
 ```bash
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload $HF_USER/$DATASET data/$DATASET \
---repo-type dataset  \
---revision v1.0
+python lerobot/scripts/push_dataset_to_hub.py \
+--raw-dir data/aloha_static_pingpong_test_raw \
+--out-dir data \
+--repo-id lerobot/aloha_static_pingpong_test \
+--raw-format aloha_hdf5
 ```
 
-You will need to set the corresponding version as a default argument in your dataset class:
-```python
-  version: str | None = "v1.1",
-```
-See: [`lerobot/common/datasets/pusht.py`](https://github.com/Cadene/lerobot/blob/main/lerobot/common/datasets/pusht.py)
+See `python lerobot/scripts/push_dataset_to_hub.py --help` for more instructions.
 
-For instance, for [lerobot/pusht](https://huggingface.co/datasets/lerobot/pusht), we used:
-```bash
-HF_USER=lerobot
-DATASET=pusht
-```
+If your dataset format is not supported, implement your own in `lerobot/common/datasets/push_dataset_to_hub/${raw_format}_format.py` by copying examples like [pusht_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py), [umi_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py), [aloha_hdf5](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py), or [xarm_pkl](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py).
 
-If you want to improve an existing dataset, you can download it locally with:
-```bash
-mkdir -p data/$DATASET
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download ${HF_USER}/$DATASET \
---repo-type dataset \
---local-dir data/$DATASET \
---local-dir-use-symlinks=False \
---revision v1.0
-```
-
-Iterate on your code and dataset with:
-```bash
-DATA_DIR=data python train.py
-```
-
-Upload a new version (v2.0 or v1.1 if the changes are respectively more or less significant):
-```bash
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload $HF_USER/$DATASET data/$DATASET \
---repo-type dataset \
---revision v1.1 \
---delete "*"
-```
-
-Then you will need to set the corresponding version as a default argument in your dataset class:
-```python
-  version: str | None = "v1.1",
-```
-See: [`lerobot/common/datasets/pusht.py`](https://github.com/Cadene/lerobot/blob/main/lerobot/common/datasets/pusht.py)
-
-
-Finally, you might want to mock the dataset if you need to update the unit tests as well:
-```bash
-python tests/scripts/mock_dataset.py --in-data-dir data/$DATASET --out-data-dir tests/data/$DATASET
-```
 
 ### Add a pretrained policy
 
-```python
-# TODO(rcadene, alexander-soare): rewrite this section
-```
+Once you have trained a policy you may upload it to the Hugging Face hub using a hub id that looks like `${hf_user}/${repo_name}` (e.g. [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)).
 
-Once you have trained a policy you may upload it to the HuggingFace hub.
-
-Firstly, make sure you have a model repository set up on the hub. The hub ID looks like HF_USER/REPO_NAME.
-
-Secondly, assuming you have trained a policy, you need:
-
-- `config.yaml` which you can get from the `.hydra` directory of your training output folder.
-- `model.pt` which should be one of the saved models in the `models` directory of your training output folder (they won't be named `model.pt` but you will need to choose one).
-
-To upload these to the hub, prepare a folder with the following structure (you can use symlinks rather than copying):
-
-```
-to_upload
-    ├── config.yaml
-    └── model.pt
-```
-
-With the folder prepared, run the following with a desired revision ID.
+You first need to find the checkpoint folder located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). Within that there is a `pretrained_model` directory which should contain:
+- `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
+- `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
+- `config.yaml`: A consolidated Hydra training configuration containing the policy, environment, and dataset configs. The policy configuration should match `config.json` exactly. The environment config is useful for anyone who wants to evaluate your policy. The dataset config just serves as a paper trail for reproducibility.
 
+To upload these to the hub, run the following:
 ```bash
-huggingface-cli upload $HUB_ID to_upload --revision $REVISION_ID
+huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
 ```
 
-If you want this to be the default revision also run the following (don't worry, it won't upload the files again; it will just adjust the file pointers):
-
-```bash
-huggingface-cli upload $HUB_ID to_upload
-```
-
-See `eval.py` for an example of how a user may use your policy.
+See [eval.py](https://github.com/huggingface/lerobot/blob/main/lerobot/scripts/eval.py) for an example of how other people may use your policy.
 
 
 ### Improve your code with profiling
@@ -298,9 +364,56 @@ with profile(
             # insert code to profile, potentially whole body of eval_policy function
 ```
 
-```bash
-python lerobot/scripts/eval.py \
---config outputs/pusht/.hydra/config.yaml \
-pretrained_model_path=outputs/pusht/model.pt \
-eval_episodes=7
+## Citation
+
+If you want, you can cite this work with:
+```bibtex
+@misc{cadene2024lerobot,
+    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Wolf, Thomas},
+    title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
+    howpublished = "\url{https://github.com/huggingface/lerobot}",
+    year = {2024}
+}
+```
+
+Additionally, if you are using any of the particular policy architecture, pretrained models, or datasets, it is recommended to cite the original authors of the work as they appear below:
+
+- [Diffusion Policy](https://diffusion-policy.cs.columbia.edu)
+```bibtex
+@article{chi2024diffusionpolicy,
+	author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
+	title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+	journal = {The International Journal of Robotics Research},
+	year = {2024},
+}
+```
+- [ACT or ALOHA](https://tonyzhaozh.github.io/aloha)
+```bibtex
+@article{zhao2023learning,
+  title={Learning fine-grained bimanual manipulation with low-cost hardware},
+  author={Zhao, Tony Z and Kumar, Vikash and Levine, Sergey and Finn, Chelsea},
+  journal={arXiv preprint arXiv:2304.13705},
+  year={2023}
+}
+```
+
+- [TDMPC](https://www.nicklashansen.com/td-mpc/)
+
+```bibtex
+@inproceedings{Hansen2022tdmpc,
+	title={Temporal Difference Learning for Model Predictive Control},
+	author={Nicklas Hansen and Xiaolong Wang and Hao Su},
+	booktitle={ICML},
+	year={2022}
+}
+```
+
+- [VQ-BeT](https://sjlee.cc/vq-bet/)
+```bibtex
+@article{lee2024behavior,
+  title={Behavior generation with latent actions},
+  author={Lee, Seungjae and Wang, Yibin and Etukuru, Haritheja and Kim, H Jin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},
+  journal={arXiv preprint arXiv:2403.03181},
+  year={2024}
+}
 ```

benchmarks/video/README.md

@@ -0,0 +1,271 @@
+# Video benchmark
+
+
+## Questions
+What is the optimal trade-off between:
+- maximizing loading time with random access,
+- minimizing memory space on disk,
+- maximizing success rate of policies,
+- compatibility across devices/platforms for decoding videos (e.g. video players, web browsers).
+
+How to encode videos?
+- Which video codec (`-vcodec`) to use? h264, h265, AV1?
+- What pixel format to use (`-pix_fmt`)? `yuv444p` or `yuv420p`?
+- How much compression (`-crf`)? No compression with `0`, intermediate compression with `25` or extreme with `50+`?
+- Which frequency to chose for key frames (`-g`)? A key frame every `10` frames?
+
+How to decode videos?
+- Which `decoder`? `torchvision`, `torchaudio`, `ffmpegio`, `decord`, or `nvc`?
+- What scenarios to use for the requesting timestamps during benchmark? (`timestamps_mode`)
+
+
+## Variables
+**Image content & size**
+We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an appartment, or in a factory, or outdoor, or with lots of moving objects in the scene, etc. Similarly, loading times might not vary linearly with the image size (resolution).
+For these reasons, we run this benchmark on four representative datasets:
+- `lerobot/pusht_image`: (96 x 96 pixels) simulation with simple geometric shapes, fixed camera.
+- `aliberts/aloha_mobile_shrimp_image`: (480 x 640 pixels) real-world indoor, moving camera.
+- `aliberts/paris_street`: (720 x 1280 pixels) real-world outdoor, moving camera.
+- `aliberts/kitchen`: (1080 x 1920 pixels) real-world indoor, fixed camera.
+
+Note: The datasets used for this benchmark need to be image datasets, not video datasets.
+
+**Data augmentations**
+We might revisit this benchmark and find better settings if we train our policies with various data augmentations to make them more robust (e.g. robust to color changes, compression, etc.).
+
+### Encoding parameters
+| parameter   | values                                                       |
+|-------------|--------------------------------------------------------------|
+| **vcodec**  | `libx264`, `libx265`, `libsvtav1`                            |
+| **pix_fmt** | `yuv444p`, `yuv420p`                                         |
+| **g**       | `1`, `2`, `3`, `4`, `5`, `6`, `10`, `15`, `20`, `40`, `None` |
+| **crf**     | `0`, `5`, `10`, `15`, `20`, `25`, `30`, `40`, `50`, `None`   |
+
+Note that `crf` value might be interpreted differently by various video codecs. In other words, the same value used with one codec doesn't necessarily translate into the same compression level with another codec. In fact, the default value (`None`) isn't the same amongst the different video codecs. Importantly, it is also the case for many other ffmpeg arguments like `g` which specifies the frequency of the key frames.
+
+For a comprehensive list and documentation of these parameters, see the ffmpeg documentation depending on the video codec used:
+- h264: https://trac.ffmpeg.org/wiki/Encode/H.264
+- h265: https://trac.ffmpeg.org/wiki/Encode/H.265
+- AV1: https://trac.ffmpeg.org/wiki/Encode/AV1
+
+### Decoding parameters
+**Decoder**
+We tested two video decoding backends from torchvision:
+- `pyav` (default)
+- `video_reader` (requires to build torchvision from source)
+
+**Requested timestamps**
+Given the way video decoding works, once a keyframe has been loaded, the decoding of subsequent frames is fast.
+This of course is affected by the `-g` parameter during encoding, which specifies the frequency of the keyframes. Given our typical use cases in robotics policies which might request a few timestamps in different random places, we want to replicate these use cases with the following scenarios:
+- `1_frame`: 1 frame,
+- `2_frames`: 2 consecutive frames (e.g. `[t, t + 1 / fps]`),
+- `6_frames`: 6 consecutive frames (e.g. `[t + i / fps for i in range(6)]`)
+
+Note that this differs significantly from a typical use case like watching a movie, in which every frame is loaded sequentially from the beginning to the end and it's acceptable to have big values for `-g`.
+
+Additionally, because some policies might request single timestamps that are a few frames appart, we also have the following scenario:
+- `2_frames_4_space`: 2 frames with 4 consecutive frames of spacing in between (e.g `[t, t + 5 / fps]`),
+
+However, due to how video decoding is implemented with `pyav`, we don't have access to an accurate seek so in practice this scenario is essentially the same as `6_frames` since all 6 frames between `t` and `t + 5 / fps` will be decoded.
+
+
+## Metrics
+**Data compression ratio (lower is better)**
+`video_images_size_ratio` is the ratio of the memory space on disk taken by the encoded video over the memory space taken by the original images. For instance, `video_images_size_ratio=25%` means that the video takes 4 times less memory space on disk compared to the original images.
+
+**Loading time ratio (lower is better)**
+`video_images_load_time_ratio` is the ratio of the time it takes to decode frames from the video at a given timestamps over the time it takes to load the exact same original images. Lower is better. For instance, `video_images_load_time_ratio=200%` means that decoding from video is 2 times slower than loading the original images.
+
+**Average Mean Square Error (lower is better)**
+`avg_mse` is the average mean square error between each decoded frame and its corresponding original image over all requested timestamps, and also divided by the number of pixels in the image to be comparable when switching to different image sizes.
+
+**Average Peak Signal to Noise Ratio (higher is better)**
+`avg_psnr` measures the ratio between the maximum possible power of a signal and the power of corrupting noise that affects the fidelity of its representation. Higher PSNR indicates better quality.
+
+**Average Structural Similarity Index Measure (higher is better)**
+`avg_ssim` evaluates the perceived quality of images by comparing luminance, contrast, and structure. SSIM values range from -1 to 1, where 1 indicates perfect similarity.
+
+One aspect that can't be measured here with those metrics is the compatibility of the encoding accross platforms, in particular on web browser, for visualization purposes.
+h264, h265 and AV1 are all commonly used codecs and should not be pose an issue. However, the chroma subsampling (`pix_fmt`) format might affect compatibility:
+- `yuv420p` is more widely supported across various platforms, including web browsers.
+- `yuv444p` offers higher color fidelity but might not be supported as broadly.
+
+
+<!-- **Loss of a pretrained policy (higher is better)** (not available)
+`loss_pretrained` is the result of evaluating with the selected encoding/decoding settings a policy pretrained on original images. It is easier to understand than `avg_l2_error`.
+
+**Success rate after retraining (higher is better)** (not available)
+`success_rate` is the result of training and evaluating a policy with the selected encoding/decoding settings. It is the most difficult metric to get but also the very best. -->
+
+
+## How the benchmark works
+The benchmark evaluates both encoding and decoding of video frames on the first episode of each dataset.
+
+**Encoding:** for each `vcodec` and `pix_fmt` pair, we use a default value for `g` and `crf` upon which we change a single value (either `g` or `crf`) to one of the specified values (we don't test every combination of those as this would be computationally too heavy).
+This gives a unique set of encoding parameters which is used to encode the episode.
+
+**Decoding:** Then, for each of those unique encodings, we iterate through every combination of the decoding parameters `backend` and `timestamps_mode`. For each of them, we record the metrics of a number of samples (given by `--num-samples`). This is parallelized for efficiency and the number of processes can be controlled with `--num-workers`. Ideally, it's best to have a `--num-samples` that is divisible by `--num-workers`.
+
+Intermediate results saved for each `vcodec` and `pix_fmt` combination in csv tables.
+These are then all concatenated to a single table ready for analysis.
+
+## Caveats
+We tried to measure the most impactful parameters for both encoding and decoding. However, for computational reasons we can't test out every combination.
+
+Additional encoding parameters exist that are not included in this benchmark. In particular:
+- `-preset` which allows for selecting encoding presets. This represents a collection of options that will provide a certain encoding speed to compression ratio. By leaving this parameter unspecified, it is considered to be `medium` for libx264 and libx265 and `8` for libsvtav1.
+- `-tune` which allows to optimize the encoding for certains aspects (e.g. film quality, fast decoding, etc.).
+
+See the documentation mentioned above for more detailled info on these settings and for a more comprehensive list of other parameters.
+
+Similarly on the decoding side, other decoders exist but are not implemented in our current benchmark. To name a few:
+- `torchaudio`
+- `ffmpegio`
+- `decord`
+- `nvc`
+
+Note as well that since we are mostly interested in the performance at decoding time (also because encoding is done only once before uploading a dataset), we did not measure encoding times nor have any metrics regarding encoding.
+However, besides the necessity to build ffmpeg from source, encoding did not pose any issue and it didn't take a significant amount of time during this benchmark.
+
+
+## Install
+Building ffmpeg from source is required to include libx265 and libaom/libsvtav1 (av1) video codecs ([compilation guide](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu)).
+
+**Note:** While you still need to build torchvision with a conda-installed `ffmpeg<4.3` to use the `video_reader` decoder (as described in [#220](https://github.com/huggingface/lerobot/pull/220)), you also need another version which is custom-built with all the video codecs for encoding. For the script to then use that version, you can prepend the command above with `PATH="$HOME/bin:$PATH"`, which is where ffmpeg should be built.
+
+
+## Adding a video decoder
+Right now, we're only benchmarking the two video decoder available with torchvision: `pyav` and `video_reader`.
+You can easily add a new decoder to benchmark by adding it to this function in the script:
+```diff
+def decode_video_frames(
+    video_path: str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str,
+) -> torch.Tensor:
+    if backend in ["pyav", "video_reader"]:
+        return decode_video_frames_torchvision(
+            video_path, timestamps, tolerance_s, backend
+        )
++    elif backend == ["your_decoder"]:
++        return your_decoder_function(
++            video_path, timestamps, tolerance_s, backend
++        )
+    else:
+        raise NotImplementedError(backend)
+```
+
+
+## Example
+For a quick run, you can try these parameters:
+```bash
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        aliberts/aloha_mobile_shrimp_image \
+    --vcodec libx264 libx265 \
+    --pix-fmt yuv444p yuv420p \
+    --g 2 20 None \
+    --crf 10 40 None \
+    --timestamps-modes 1_frame 2_frames \
+    --backends pyav video_reader \
+    --num-samples 5 \
+    --num-workers 5 \
+    --save-frames 0
+```
+
+
+## Results
+
+### Reproduce
+We ran the benchmark with the following parameters:
+```bash
+# h264 and h265 encodings
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/kitchen \
+    --vcodec libx264 libx265 \
+    --pix-fmt yuv444p yuv420p \
+    --g 1 2 3 4 5 6 10 15 20 40 None \
+    --crf 0 5 10 15 20 25 30 40 50 None \
+    --timestamps-modes 1_frame 2_frames 6_frames \
+    --backends pyav video_reader \
+    --num-samples 50 \
+    --num-workers 5 \
+    --save-frames 1
+
+# av1 encoding (only compatible with yuv420p and pyav decoder)
+python benchmark/video/run_video_benchmark.py \
+    --output-dir outputs/video_benchmark \
+    --repo-ids \
+        lerobot/pusht_image \
+        aliberts/aloha_mobile_shrimp_image \
+        aliberts/paris_street \
+        aliberts/kitchen \
+    --vcodec libsvtav1 \
+    --pix-fmt yuv420p \
+    --g 1 2 3 4 5 6 10 15 20 40 None \
+    --crf 0 5 10 15 20 25 30 40 50 None \
+    --timestamps-modes 1_frame 2_frames 6_frames \
+    --backends pyav \
+    --num-samples 50 \
+    --num-workers 5 \
+    --save-frames 1
+```
+
+The full results are available [here](https://docs.google.com/spreadsheets/d/1OYJB43Qu8fC26k_OyoMFgGBBKfQRCi4BIuYitQnq3sw/edit?usp=sharing)
+
+
+### Parameters selected for LeRobotDataset
+Considering these results, we chose what we think is the best set of encoding parameter:
+- vcodec: `libsvtav1`
+- pix-fmt: `yuv420p`
+- g: `2`
+- crf: `30`
+
+Since we're using av1 encoding, we're choosing the `pyav` decoder as `video_reader` does not support it (and `pyav` doesn't require a custom build of `torchvision`).
+
+### Summary
+
+These tables show the results for `g=2` and `crf=30`, using `timestamps-modes=6_frames` and `backend=pyav`
+
+| video_images_size_ratio            | vcodec     | pix_fmt |           |           |           |
+|------------------------------------|------------|---------|-----------|-----------|-----------|
+|                                    | libx264    |         | libx265   |           | libsvtav1 |
+| repo_id                            | yuv420p    | yuv444p | yuv420p   | yuv444p   | yuv420p   |
+| lerobot/pusht_image                | **16.97%** | 17.58%  | 18.57%    | 18.86%    | 22.06%    |
+| aliberts/aloha_mobile_shrimp_image | 2.14%      | 2.11%   | 1.38%     | **1.37%** | 5.59%     |
+| aliberts/paris_street              | 2.12%      | 2.13%   | **1.54%** | **1.54%** | 4.43%     |
+| aliberts/kitchen                   | 1.40%      | 1.39%   | **1.00%** | **1.00%** | 2.52%     |
+
+| video_images_load_time_ratio       | vcodec  | pix_fmt |          |         |           |
+|------------------------------------|---------|---------|----------|---------|-----------|
+|                                    | libx264 |         | libx265  |         | libsvtav1 |
+| repo_id                            | yuv420p | yuv444p | yuv420p  | yuv444p | yuv420p   |
+| lerobot/pusht_image                | 6.45    | 5.19    | **1.90** | 2.12    | 2.47      |
+| aliberts/aloha_mobile_shrimp_image | 11.80   | 7.92    | 0.71     | 0.85    | **0.48**  |
+| aliberts/paris_street              | 2.21    | 2.05    | 0.36     | 0.49    | **0.30**  |
+| aliberts/kitchen                   | 1.46    | 1.46    | 0.28     | 0.51    | **0.26**  |
+
+|                                    |          | vcodec   | pix_fmt      |          |           |              |
+|------------------------------------|----------|----------|--------------|----------|-----------|--------------|
+|                                    |          | libx264  |              | libx265  |           | libsvtav1    |
+| repo_id                            | metric   | yuv420p  | yuv444p      | yuv420p  | yuv444p   | yuv420p      |
+| lerobot/pusht_image                | avg_mse  | 2.90E-04 | **2.03E-04** | 3.13E-04 | 2.29E-04  | 2.19E-04     |
+|                                    | avg_psnr | 35.44    | 37.07        | 35.49    | **37.30** | 37.20        |
+|                                    | avg_ssim | 98.28%   | **98.85%**   | 98.31%   | 98.84%    | 98.72%       |
+| aliberts/aloha_mobile_shrimp_image | avg_mse  | 2.76E-04 | 2.59E-04     | 3.17E-04 | 3.06E-04  | **1.30E-04** |
+|                                    | avg_psnr | 35.91    | 36.21        | 35.88    | 36.09     | **40.17**    |
+|                                    | avg_ssim | 95.19%   | 95.18%       | 95.00%   | 95.05%    | **97.73%**   |
+| aliberts/paris_street              | avg_mse  | 6.89E-04 | 6.70E-04     | 4.03E-03 | 4.02E-03  | **3.09E-04** |
+|                                    | avg_psnr | 33.48    | 33.68        | 32.05    | 32.15     | **35.40**    |
+|                                    | avg_ssim | 93.76%   | 93.75%       | 89.46%   | 89.46%    | **95.46%**   |
+| aliberts/kitchen                   | avg_mse  | 2.50E-04 | 2.24E-04     | 4.28E-04 | 4.18E-04  | **1.53E-04** |
+|                                    | avg_psnr | 36.73    | 37.33        | 36.56    | 36.75     | **39.12**    |
+|                                    | avg_ssim | 95.47%   | 95.58%       | 95.52%   | 95.53%    | **96.82%**   |

benchmarks/video/capture_camera_feed.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Capture video feed from a camera as raw images."""
+
+import argparse
+import datetime as dt
+from pathlib import Path
+
+import cv2
+
+
+def display_and_save_video_stream(output_dir: Path, fps: int, width: int, height: int):
+    now = dt.datetime.now()
+    capture_dir = output_dir / f"{now:%Y-%m-%d}" / f"{now:%H-%M-%S}"
+    if not capture_dir.exists():
+        capture_dir.mkdir(parents=True, exist_ok=True)
+
+    # Opens the default webcam
+    cap = cv2.VideoCapture(0)
+    if not cap.isOpened():
+        print("Error: Could not open video stream.")
+        return
+
+    cap.set(cv2.CAP_PROP_FPS, fps)
+    cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
+    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
+
+    frame_index = 0
+    while True:
+        ret, frame = cap.read()
+
+        if not ret:
+            print("Error: Could not read frame.")
+            break
+
+        cv2.imshow("Video Stream", frame)
+        cv2.imwrite(str(capture_dir / f"frame_{frame_index:06d}.png"), frame)
+        frame_index += 1
+
+        # Break the loop on 'q' key press
+        if cv2.waitKey(1) & 0xFF == ord("q"):
+            break
+
+    # Release the capture and destroy all windows
+    cap.release()
+    cv2.destroyAllWindows()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path("outputs/cam_capture/"),
+        help="Directory where the capture images are written. A subfolder named with the current date & time will be created inside it for each capture.",
+    )
+    parser.add_argument(
+        "--fps",
+        type=int,
+        default=30,
+        help="Frames Per Second of the capture.",
+    )
+    parser.add_argument(
+        "--width",
+        type=int,
+        default=1280,
+        help="Width of the captured images.",
+    )
+    parser.add_argument(
+        "--height",
+        type=int,
+        default=720,
+        help="Height of the captured images.",
+    )
+    args = parser.parse_args()
+    display_and_save_video_stream(**vars(args))

benchmarks/video/run_video_benchmark.py

@@ -0,0 +1,490 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Assess the performance of video decoding in various configurations.
+
+This script will benchmark different video encoding and decoding parameters.
+See the provided README.md or run `python benchmark/video/run_video_benchmark.py --help` for usage info.
+"""
+
+import argparse
+import datetime as dt
+import random
+import shutil
+from collections import OrderedDict
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+
+import einops
+import numpy as np
+import pandas as pd
+import PIL
+import torch
+from skimage.metrics import mean_squared_error, peak_signal_noise_ratio, structural_similarity
+from tqdm import tqdm
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.video_utils import (
+    decode_video_frames_torchvision,
+    encode_video_frames,
+)
+from lerobot.common.utils.benchmark import TimeBenchmark
+
+BASE_ENCODING = OrderedDict(
+    [
+        ("vcodec", "libx264"),
+        ("pix_fmt", "yuv444p"),
+        ("g", 2),
+        ("crf", None),
+        # TODO(aliberts): Add fastdecode
+        # ("fastdecode", 0),
+    ]
+)
+
+
+# TODO(rcadene, aliberts): move to `utils.py` folder when we want to refactor
+def parse_int_or_none(value) -> int | None:
+    if value.lower() == "none":
+        return None
+    try:
+        return int(value)
+    except ValueError as e:
+        raise argparse.ArgumentTypeError(f"Invalid int or None: {value}") from e
+
+
+def check_datasets_formats(repo_ids: list) -> None:
+    for repo_id in repo_ids:
+        dataset = LeRobotDataset(repo_id)
+        if dataset.video:
+            raise ValueError(
+                f"Use only image dataset for running this benchmark. Video dataset provided: {repo_id}"
+            )
+
+
+def get_directory_size(directory: Path) -> int:
+    total_size = 0
+    for item in directory.rglob("*"):
+        if item.is_file():
+            total_size += item.stat().st_size
+    return total_size
+
+
+def load_original_frames(imgs_dir: Path, timestamps: list[float], fps: int) -> torch.Tensor:
+    frames = []
+    for ts in timestamps:
+        idx = int(ts * fps)
+        frame = PIL.Image.open(imgs_dir / f"frame_{idx:06d}.png")
+        frame = torch.from_numpy(np.array(frame))
+        frame = frame.type(torch.float32) / 255
+        frame = einops.rearrange(frame, "h w c -> c h w")
+        frames.append(frame)
+    return torch.stack(frames)
+
+
+def save_decoded_frames(
+    imgs_dir: Path, save_dir: Path, frames: torch.Tensor, timestamps: list[float], fps: int
+) -> None:
+    if save_dir.exists() and len(list(save_dir.glob("frame_*.png"))) == len(timestamps):
+        return
+
+    save_dir.mkdir(parents=True, exist_ok=True)
+    for i, ts in enumerate(timestamps):
+        idx = int(ts * fps)
+        frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
+        PIL.Image.fromarray(frame_hwc).save(save_dir / f"frame_{idx:06d}_decoded.png")
+        shutil.copyfile(imgs_dir / f"frame_{idx:06d}.png", save_dir / f"frame_{idx:06d}_original.png")
+
+
+def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
+    ep_num_images = dataset.episode_data_index["to"][0].item()
+    if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
+        return
+
+    imgs_dir.mkdir(parents=True, exist_ok=True)
+    hf_dataset = dataset.hf_dataset.with_format(None)
+
+    # We only save images from the first camera
+    img_keys = [key for key in hf_dataset.features if key.startswith("observation.image")]
+    imgs_dataset = hf_dataset.select_columns(img_keys[0])
+
+    for i, item in enumerate(
+        tqdm(imgs_dataset, desc=f"saving {dataset.repo_id} first episode images", leave=False)
+    ):
+        img = item[img_keys[0]]
+        img.save(str(imgs_dir / f"frame_{i:06d}.png"), quality=100)
+
+        if i >= ep_num_images - 1:
+            break
+
+
+def sample_timestamps(timestamps_mode: str, ep_num_images: int, fps: int) -> list[float]:
+    # Start at 5 to allow for 2_frames_4_space and 6_frames
+    idx = random.randint(5, ep_num_images - 1)
+    match timestamps_mode:
+        case "1_frame":
+            frame_indexes = [idx]
+        case "2_frames":
+            frame_indexes = [idx - 1, idx]
+        case "2_frames_4_space":
+            frame_indexes = [idx - 5, idx]
+        case "6_frames":
+            frame_indexes = [idx - i for i in range(6)][::-1]
+        case _:
+            raise ValueError(timestamps_mode)
+
+    return [idx / fps for idx in frame_indexes]
+
+
+def decode_video_frames(
+    video_path: str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str,
+) -> torch.Tensor:
+    if backend in ["pyav", "video_reader"]:
+        return decode_video_frames_torchvision(video_path, timestamps, tolerance_s, backend)
+    else:
+        raise NotImplementedError(backend)
+
+
+def benchmark_decoding(
+    imgs_dir: Path,
+    video_path: Path,
+    timestamps_mode: str,
+    backend: str,
+    ep_num_images: int,
+    fps: int,
+    num_samples: int = 50,
+    num_workers: int = 4,
+    save_frames: bool = False,
+) -> dict:
+    def process_sample(sample: int):
+        time_benchmark = TimeBenchmark()
+        timestamps = sample_timestamps(timestamps_mode, ep_num_images, fps)
+        num_frames = len(timestamps)
+        result = {
+            "psnr_values": [],
+            "ssim_values": [],
+            "mse_values": [],
+        }
+
+        with time_benchmark:
+            frames = decode_video_frames(video_path, timestamps=timestamps, tolerance_s=5e-1, backend=backend)
+        result["load_time_video_ms"] = time_benchmark.result_ms / num_frames
+
+        with time_benchmark:
+            original_frames = load_original_frames(imgs_dir, timestamps, fps)
+        result["load_time_images_ms"] = time_benchmark.result_ms / num_frames
+
+        frames_np, original_frames_np = frames.numpy(), original_frames.numpy()
+        for i in range(num_frames):
+            result["mse_values"].append(mean_squared_error(original_frames_np[i], frames_np[i]))
+            result["psnr_values"].append(
+                peak_signal_noise_ratio(original_frames_np[i], frames_np[i], data_range=1.0)
+            )
+            result["ssim_values"].append(
+                structural_similarity(original_frames_np[i], frames_np[i], data_range=1.0, channel_axis=0)
+            )
+
+        if save_frames and sample == 0:
+            save_dir = video_path.with_suffix("") / f"{timestamps_mode}_{backend}"
+            save_decoded_frames(imgs_dir, save_dir, frames, timestamps, fps)
+
+        return result
+
+    load_times_video_ms = []
+    load_times_images_ms = []
+    mse_values = []
+    psnr_values = []
+    ssim_values = []
+
+    # A sample is a single set of decoded frames specified by timestamps_mode (e.g. a single frame, 2 frames, etc.).
+    # For each sample, we record metrics (loading time and quality metrics) which are then averaged over all samples.
+    # As these samples are independent, we run them in parallel threads to speed up the benchmark.
+    with ThreadPoolExecutor(max_workers=num_workers) as executor:
+        futures = [executor.submit(process_sample, i) for i in range(num_samples)]
+        for future in tqdm(as_completed(futures), total=num_samples, desc="samples", leave=False):
+            result = future.result()
+            load_times_video_ms.append(result["load_time_video_ms"])
+            load_times_images_ms.append(result["load_time_images_ms"])
+            psnr_values.extend(result["psnr_values"])
+            ssim_values.extend(result["ssim_values"])
+            mse_values.extend(result["mse_values"])
+
+    avg_load_time_video_ms = float(np.array(load_times_video_ms).mean())
+    avg_load_time_images_ms = float(np.array(load_times_images_ms).mean())
+    video_images_load_time_ratio = avg_load_time_video_ms / avg_load_time_images_ms
+
+    return {
+        "avg_load_time_video_ms": avg_load_time_video_ms,
+        "avg_load_time_images_ms": avg_load_time_images_ms,
+        "video_images_load_time_ratio": video_images_load_time_ratio,
+        "avg_mse": float(np.mean(mse_values)),
+        "avg_psnr": float(np.mean(psnr_values)),
+        "avg_ssim": float(np.mean(ssim_values)),
+    }
+
+
+def benchmark_encoding_decoding(
+    dataset: LeRobotDataset,
+    video_path: Path,
+    imgs_dir: Path,
+    encoding_cfg: dict,
+    decoding_cfg: dict,
+    num_samples: int,
+    num_workers: int,
+    save_frames: bool,
+    overwrite: bool = False,
+    seed: int = 1337,
+) -> list[dict]:
+    fps = dataset.fps
+
+    if overwrite or not video_path.is_file():
+        tqdm.write(f"encoding {video_path}")
+        encode_video_frames(
+            imgs_dir=imgs_dir,
+            video_path=video_path,
+            fps=fps,
+            vcodec=encoding_cfg["vcodec"],
+            pix_fmt=encoding_cfg["pix_fmt"],
+            g=encoding_cfg.get("g"),
+            crf=encoding_cfg.get("crf"),
+            # fast_decode=encoding_cfg.get("fastdecode"),
+            overwrite=True,
+        )
+
+    ep_num_images = dataset.episode_data_index["to"][0].item()
+    width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
+    num_pixels = width * height
+    video_size_bytes = video_path.stat().st_size
+    images_size_bytes = get_directory_size(imgs_dir)
+    video_images_size_ratio = video_size_bytes / images_size_bytes
+
+    random.seed(seed)
+    benchmark_table = []
+    for timestamps_mode in tqdm(
+        decoding_cfg["timestamps_modes"], desc="decodings (timestamps_modes)", leave=False
+    ):
+        for backend in tqdm(decoding_cfg["backends"], desc="decodings (backends)", leave=False):
+            benchmark_row = benchmark_decoding(
+                imgs_dir,
+                video_path,
+                timestamps_mode,
+                backend,
+                ep_num_images,
+                fps,
+                num_samples,
+                num_workers,
+                save_frames,
+            )
+            benchmark_row.update(
+                **{
+                    "repo_id": dataset.repo_id,
+                    "resolution": f"{width} x {height}",
+                    "num_pixels": num_pixels,
+                    "video_size_bytes": video_size_bytes,
+                    "images_size_bytes": images_size_bytes,
+                    "video_images_size_ratio": video_images_size_ratio,
+                    "timestamps_mode": timestamps_mode,
+                    "backend": backend,
+                },
+                **encoding_cfg,
+            )
+            benchmark_table.append(benchmark_row)
+
+    return benchmark_table
+
+
+def main(
+    output_dir: Path,
+    repo_ids: list[str],
+    vcodec: list[str],
+    pix_fmt: list[str],
+    g: list[int],
+    crf: list[int],
+    # fastdecode: list[int],
+    timestamps_modes: list[str],
+    backends: list[str],
+    num_samples: int,
+    num_workers: int,
+    save_frames: bool,
+):
+    check_datasets_formats(repo_ids)
+    encoding_benchmarks = {
+        "g": g,
+        "crf": crf,
+        # "fastdecode": fastdecode,
+    }
+    decoding_benchmarks = {
+        "timestamps_modes": timestamps_modes,
+        "backends": backends,
+    }
+    headers = ["repo_id", "resolution", "num_pixels"]
+    headers += list(BASE_ENCODING.keys())
+    headers += [
+        "timestamps_mode",
+        "backend",
+        "video_size_bytes",
+        "images_size_bytes",
+        "video_images_size_ratio",
+        "avg_load_time_video_ms",
+        "avg_load_time_images_ms",
+        "video_images_load_time_ratio",
+        "avg_mse",
+        "avg_psnr",
+        "avg_ssim",
+    ]
+    file_paths = []
+    for video_codec in tqdm(vcodec, desc="encodings (vcodec)"):
+        for pixel_format in tqdm(pix_fmt, desc="encodings (pix_fmt)", leave=False):
+            benchmark_table = []
+            for repo_id in tqdm(repo_ids, desc="encodings (datasets)", leave=False):
+                dataset = LeRobotDataset(repo_id)
+                imgs_dir = output_dir / "images" / dataset.repo_id.replace("/", "_")
+                # We only use the first episode
+                save_first_episode(imgs_dir, dataset)
+                for key, values in tqdm(encoding_benchmarks.items(), desc="encodings (g, crf)", leave=False):
+                    for value in tqdm(values, desc=f"encodings ({key})", leave=False):
+                        encoding_cfg = BASE_ENCODING.copy()
+                        encoding_cfg["vcodec"] = video_codec
+                        encoding_cfg["pix_fmt"] = pixel_format
+                        encoding_cfg[key] = value
+                        args_path = Path("_".join(str(value) for value in encoding_cfg.values()))
+                        video_path = output_dir / "videos" / args_path / f"{repo_id.replace('/', '_')}.mp4"
+                        benchmark_table += benchmark_encoding_decoding(
+                            dataset,
+                            video_path,
+                            imgs_dir,
+                            encoding_cfg,
+                            decoding_benchmarks,
+                            num_samples,
+                            num_workers,
+                            save_frames,
+                        )
+
+            # Save intermediate results
+            benchmark_df = pd.DataFrame(benchmark_table, columns=headers)
+            now = dt.datetime.now()
+            csv_path = (
+                output_dir
+                / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_{video_codec}_{pixel_format}_{num_samples}-samples.csv"
+            )
+            benchmark_df.to_csv(csv_path, header=True, index=False)
+            file_paths.append(csv_path)
+            del benchmark_df
+
+    # Concatenate all results
+    df_list = [pd.read_csv(csv_path) for csv_path in file_paths]
+    concatenated_df = pd.concat(df_list, ignore_index=True)
+    concatenated_path = output_dir / f"{now:%Y-%m-%d}_{now:%H-%M-%S}_all_{num_samples}-samples.csv"
+    concatenated_df.to_csv(concatenated_path, header=True, index=False)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path("outputs/video_benchmark"),
+        help="Directory where the video benchmark outputs are written.",
+    )
+    parser.add_argument(
+        "--repo-ids",
+        type=str,
+        nargs="*",
+        default=[
+            "lerobot/pusht_image",
+            "aliberts/aloha_mobile_shrimp_image",
+            "aliberts/paris_street",
+            "aliberts/kitchen",
+        ],
+        help="Datasets repo-ids to test against. First episodes only are used. Must be images.",
+    )
+    parser.add_argument(
+        "--vcodec",
+        type=str,
+        nargs="*",
+        default=["libx264", "libx265", "libsvtav1"],
+        help="Video codecs to be tested",
+    )
+    parser.add_argument(
+        "--pix-fmt",
+        type=str,
+        nargs="*",
+        default=["yuv444p", "yuv420p"],
+        help="Pixel formats (chroma subsampling) to be tested",
+    )
+    parser.add_argument(
+        "--g",
+        type=parse_int_or_none,
+        nargs="*",
+        default=[1, 2, 3, 4, 5, 6, 10, 15, 20, 40, 100, None],
+        help="Group of pictures sizes to be tested.",
+    )
+    parser.add_argument(
+        "--crf",
+        type=parse_int_or_none,
+        nargs="*",
+        default=[0, 5, 10, 15, 20, 25, 30, 40, 50, None],
+        help="Constant rate factors to be tested.",
+    )
+    # parser.add_argument(
+    #     "--fastdecode",
+    #     type=int,
+    #     nargs="*",
+    #     default=[0, 1],
+    #     help="Use the fastdecode tuning option. 0 disables it. "
+    #         "For libx264 and libx265, only 1 is possible. "
+    #         "For libsvtav1, 1, 2 or 3 are possible values with a higher number meaning a faster decoding optimization",
+    # )
+    parser.add_argument(
+        "--timestamps-modes",
+        type=str,
+        nargs="*",
+        default=[
+            "1_frame",
+            "2_frames",
+            "2_frames_4_space",
+            "6_frames",
+        ],
+        help="Timestamps scenarios to be tested.",
+    )
+    parser.add_argument(
+        "--backends",
+        type=str,
+        nargs="*",
+        default=["pyav", "video_reader"],
+        help="Torchvision decoding backend to be tested.",
+    )
+    parser.add_argument(
+        "--num-samples",
+        type=int,
+        default=50,
+        help="Number of samples for each encoding x decoding config.",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=10,
+        help="Number of processes for parallelized sample processing.",
+    )
+    parser.add_argument(
+        "--save-frames",
+        type=int,
+        default=0,
+        help="Whether to save decoded frames or not. Enter a non-zero number for true.",
+    )
+    args = parser.parse_args()
+    main(**vars(args))

docker/lerobot-cpu/Dockerfile

@@ -8,7 +8,8 @@ ARG DEBIAN_FRONTEND=noninteractive
 # Install apt dependencies
 RUN apt-get update && apt-get install -y --no-install-recommends \
     build-essential cmake \
-    libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    speech-dispatcher \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
 # Create virtual environment
@@ -21,7 +22,7 @@ RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 COPY . /lerobot
 WORKDIR /lerobot
 RUN pip install --upgrade --no-cache-dir pip
-RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht]" \
+RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]" \
     --extra-index-url https://download.pytorch.org/whl/cpu
 
 # Set EGL as the rendering backend for MuJoCo

docker/lerobot-gpu-dev/Dockerfile

@@ -0,0 +1,68 @@
+FROM nvidia/cuda:12.2.2-devel-ubuntu22.04
+
+# Configure image
+ARG PYTHON_VERSION=3.10
+ARG DEBIAN_FRONTEND=noninteractive
+
+# Install apt dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake \
+    git git-lfs openssh-client \
+    nano vim less util-linux tree \
+    htop atop nvtop \
+    sed gawk grep curl wget zip unzip \
+    tcpdump sysstat screen tmux \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
+    speech-dispatcher \
+    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Install ffmpeg build dependencies. See:
+# https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu
+# TODO(aliberts): create image to build dependencies from source instead
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    autoconf automake yasm \
+    libass-dev \
+    libfreetype6-dev \
+    libgnutls28-dev \
+    libunistring-dev \
+    libmp3lame-dev \
+    libtool \
+    libvorbis-dev \
+    meson \
+    ninja-build \
+    pkg-config \
+    texinfo \
+    yasm \
+    zlib1g-dev \
+    nasm \
+    libx264-dev \
+    libx265-dev libnuma-dev \
+    libvpx-dev \
+    libfdk-aac-dev \
+    libopus-dev \
+    libsvtav1-dev libsvtav1enc-dev libsvtav1dec-dev \
+    libdav1d-dev
+
+# Install gh cli tool
+RUN (type -p wget >/dev/null || (apt update && apt-get install wget -y)) \
+    && mkdir -p -m 755 /etc/apt/keyrings \
+    && wget -qO- https://cli.github.com/packages/githubcli-archive-keyring.gpg | tee /etc/apt/keyrings/githubcli-archive-keyring.gpg > /dev/null \
+    && chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg \
+    && echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" | tee /etc/apt/sources.list.d/github-cli.list > /dev/null \
+    && apt update \
+    && apt install gh -y \
+    && apt clean && rm -rf /var/lib/apt/lists/*
+
+# Setup `python`
+RUN ln -s /usr/bin/python3 /usr/bin/python
+
+# Install poetry
+RUN curl -sSL https://install.python-poetry.org | python -
+ENV PATH="/root/.local/bin:$PATH"
+RUN echo 'if [ "$HOME" != "/root" ]; then ln -sf /root/.local/bin/poetry $HOME/.local/bin/poetry; fi' >> /root/.bashrc
+RUN poetry config virtualenvs.create false
+RUN poetry config virtualenvs.in-project true
+
+# Set EGL as the rendering backend for MuJoCo
+ENV MUJOCO_GL="egl"

docker/lerobot-gpu/Dockerfile

@@ -4,13 +4,16 @@ FROM nvidia/cuda:12.4.1-base-ubuntu22.04
 ARG PYTHON_VERSION=3.10
 ARG DEBIAN_FRONTEND=noninteractive
 
+
 # Install apt dependencies
 RUN apt-get update && apt-get install -y --no-install-recommends \
     build-essential cmake \
-    libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
-    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    speech-dispatcher \
+    python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
+
 # Create virtual environment
 RUN ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python
 RUN python -m venv /opt/venv
@@ -21,7 +24,7 @@ RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 COPY . /lerobot
 WORKDIR /lerobot
 RUN pip install --upgrade --no-cache-dir pip
-RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht]"
+RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht, dynamixel]"
 
 # Set EGL as the rendering backend for MuJoCo
 ENV MUJOCO_GL="egl"

examples/10_use_so100.md

@@ -0,0 +1,275 @@
+This tutorial explains how to use [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot.
+
+## Source the parts
+
+Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with link to source the parts, as well as the instructions to 3D print the parts, and advices if it's your first time printing or if you don't own a 3D printer already.
+
+**Important**: Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
+
+## Install LeRobot
+
+On your computer:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
+bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+rm ~/miniconda3/miniconda.sh
+~/miniconda3/bin/conda init bash
+```
+
+2. Restart shell or `source ~/.bashrc`
+
+3. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+5. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+For Linux only (not Mac), install extra dependencies for recording datasets:
+```bash
+conda install -y -c conda-forge ffmpeg
+pip uninstall -y opencv-python
+conda install -y -c conda-forge "opencv>=4.10.0"
+```
+
+## Configure the motors
+
+Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the use of our scripts below.
+
+**Find USB ports associated to your arms**
+To find the correct ports for each arm, run the utility script twice:
+```bash
+python lerobot/scripts/find_motors_bus_port.py
+```
+
+Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
+
+[...Disconnect leader arm and press Enter...]
+
+The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
+Reconnect the usb cable.
+```
+
+Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
+
+[...Disconnect follower arm and press Enter...]
+
+The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the usb cable.
+```
+
+Troubleshooting: On Linux, you might need to give access to the USB ports by running:
+```bash
+sudo chmod 666 /dev/ttyACM0
+sudo chmod 666 /dev/ttyACM1
+```
+
+**Configure your motors**
+Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 1
+```
+
+Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
+
+Then unplug your motor and plug the second motor and set its ID to 2.
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 2
+```
+
+Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
+
+**Remove the gears of the 6 leader motors**
+Follow step 2 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
+
+**Add motor horn to the motors**
+Follow step 3 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). For SO-100, you need to align the holes on the motor horn to the motor spline to be approximately 1:30, 4:30, 7:30 and 10:30.
+Try to avoid rotating the motor while doing so to keep position 2048 set during configuration. It is especially tricky for the leader motors as it is more sensible without the gears, but it's ok if it's a bit rotated.
+
+## Assemble the arms
+
+Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
+
+## Calibrate
+
+Next, you'll need to calibrate your SO-100 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one SO-100 robot to work on another.
+
+**Manual calibration of follower arm**
+/!\ Contrarily to step 6 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
+
+You will need to move the follower arm to these positions sequentially:
+
+| 1. Zero position | 2. Rotated position | 3. Rest position |
+|---|---|---|
+| <img src="../media/so100/follower_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/so100/follower_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/so100/follower_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
+
+Make sure both arms are connected and run this script to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --robot-overrides '~cameras' --arms main_follower
+```
+
+**Manual calibration of leader arm**
+Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
+
+| 1. Zero position | 2. Rotated position | 3. Rest position |
+|---|---|---|
+| <img src="../media/so100/leader_zero.webp?raw=true" alt="SO-100 leader arm zero position" title="SO-100 leader arm zero position" style="width:100%;"> | <img src="../media/so100/leader_rotated.webp?raw=true" alt="SO-100 leader arm rotated position" title="SO-100 leader arm rotated position" style="width:100%;"> | <img src="../media/so100/leader_rest.webp?raw=true" alt="SO-100 leader arm rest position" title="SO-100 leader arm rest position" style="width:100%;"> |
+
+Run this script to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --robot-overrides '~cameras' --arms main_leader
+```
+
+## Teleoperate
+
+**Simple teleop**
+Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --robot-overrides '~cameras' \
+    --display-cameras 0
+```
+
+
+**Teleop with displaying cameras**
+Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/so100.yaml
+```
+
+## Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset with SO-100.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py record \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --fps 30 \
+    --repo-id ${HF_USER}/so100_test \
+    --tags so100 tutorial \
+    --warmup-time-s 5 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 2 \
+    --push-to-hub 1
+```
+
+## Visualize a dataset
+
+If you uploaded your dataset to the hub with `--push-to-hub 1`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+```bash
+echo ${HF_USER}/so100_test
+```
+
+If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/so100_test
+```
+
+## Replay an episode
+
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py replay \
+    --robot-path lerobot/configs/robot/so100.yaml \
+    --fps 30 \
+    --repo-id ${HF_USER}/so100_test \
+    --episode 0
+```
+
+## Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+python lerobot/scripts/train.py \
+  dataset_repo_id=${HF_USER}/so100_test \
+  policy=act_so100_real \
+  env=so100_real \
+  hydra.run.dir=outputs/train/act_so100_test \
+  hydra.job.name=act_so100_test \
+  device=cuda \
+  wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/so100_test`.
+2. We provided the policy with `policy=act_so100_real`. This loads configurations from [`lerobot/configs/policy/act_so100_real.yaml`](../lerobot/configs/policy/act_so100_real.yaml). Importantly, this policy uses 2 cameras as input `laptop`, `phone`.
+3. We provided an environment as argument with `env=so100_real`. This loads configurations from [`lerobot/configs/env/so100_real.yaml`](../lerobot/configs/env/so100_real.yaml).
+4. We provided `device=cuda` since we are training on a Nvidia GPU, but you can also use `device=mps` if you are using a Mac with Apple silicon, or `device=cpu` otherwise.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+
+Training should take several hours. You will find checkpoints in `outputs/train/act_so100_test/checkpoints`.
+
+## Evaluate your policy
+
+You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
+```bash
+python lerobot/scripts/control_robot.py record \
+  --robot-path lerobot/configs/robot/so100.yaml \
+  --fps 30 \
+  --repo-id ${HF_USER}/eval_act_so100_test \
+  --tags so100 tutorial eval \
+  --warmup-time-s 5 \
+  --episode-time-s 40 \
+  --reset-time-s 10 \
+  --num-episodes 10 \
+  -p outputs/train/act_so100_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_so100_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_so100_test`).
+
+## More
+
+Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
+
+If you have any question or need help, please reach out on Discord in the channel [`#so100-arm`](https://discord.com/channels/1216765309076115607/1237741463832363039).

examples/11_use_moss.md

@@ -0,0 +1,275 @@
+This tutorial explains how to use [Moss v1](https://github.com/jess-moss/moss-robot-arms) with LeRobot.
+
+## Source the parts
+
+Follow this [README](https://github.com/jess-moss/moss-robot-arms). It contains the bill of materials, with link to source the parts, as well as the instructions to 3D print the parts, and advices if it's your first time printing or if you don't own a 3D printer already.
+
+**Important**: Before assembling, you will first need to configure your motors. To this end, we provide a nice script, so let's first install LeRobot. After configuration, we will also guide you through assembly.
+
+## Install LeRobot
+
+On your computer:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
+bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+rm ~/miniconda3/miniconda.sh
+~/miniconda3/bin/conda init bash
+```
+
+2. Restart shell or `source ~/.bashrc`
+
+3. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+5. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+For Linux only (not Mac), install extra dependencies for recording datasets:
+```bash
+conda install -y -c conda-forge ffmpeg
+pip uninstall -y opencv-python
+conda install -y -c conda-forge "opencv>=4.10.0"
+```
+
+## Configure the motors
+
+Follow steps 1 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the use of our scripts below.
+
+**Find USB ports associated to your arms**
+To find the correct ports for each arm, run the utility script twice:
+```bash
+python lerobot/scripts/find_motors_bus_port.py
+```
+
+Example output when identifying the leader arm's port (e.g., `/dev/tty.usbmodem575E0031751` on Mac, or possibly `/dev/ttyACM0` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
+
+[...Disconnect leader arm and press Enter...]
+
+The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0031751
+Reconnect the usb cable.
+```
+
+Example output when identifying the follower arm's port (e.g., `/dev/tty.usbmodem575E0032081`, or possibly `/dev/ttyACM1` on Linux):
+```
+Finding all available ports for the MotorBus.
+['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
+Remove the usb cable from your DynamixelMotorsBus and press Enter when done.
+
+[...Disconnect follower arm and press Enter...]
+
+The port of this DynamixelMotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the usb cable.
+```
+
+Troubleshooting: On Linux, you might need to give access to the USB ports by running:
+```bash
+sudo chmod 666 /dev/ttyACM0
+sudo chmod 666 /dev/ttyACM1
+```
+
+**Configure your motors**
+Plug your first motor and run this script to set its ID to 1. It will also set its present position to 2048, so expect your motor to rotate:
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 1
+```
+
+Note: These motors are currently limitated. They can take values between 0 and 4096 only, which corresponds to a full turn. They can't turn more than that. 2048 is at the middle of this range, so we can take -2048 steps (180 degrees anticlockwise) and reach the maximum range, or take +2048 steps (180 degrees clockwise) and reach the maximum range. The configuration step also sets the homing offset to 0, so that if you misassembled the arm, you can always update the homing offset to account for a shift up to ± 2048 steps (± 180 degrees).
+
+Then unplug your motor and plug the second motor and set its ID to 2.
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 2
+```
+
+Redo the process for all your motors until ID 6. Do the same for the 6 motors of the leader arm.
+
+**Remove the gears of the 6 leader motors**
+Follow step 2 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). You need to remove the gear for the motors of the leader arm. As a result, you will only use the position encoding of the motor and reduce friction to more easily operate the leader arm.
+
+**Add motor horn to the motors**
+Follow step 3 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). For Moss v1, you need to align the holes on the motor horn to the motor spline to be approximately 3, 6, 9 and 12 o'clock.
+Try to avoid rotating the motor while doing so to keep position 2048 set during configuration. It is especially tricky for the leader motors as it is more sensible without the gears, but it's ok if it's a bit rotated.
+
+## Assemble the arms
+
+Follow step 4 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic). The first arm should take a bit more than 1 hour to assemble, but once you get use to it, you can do it under 1 hour for the second arm.
+
+## Calibrate
+
+Next, you'll need to calibrate your Moss v1 robot to ensure that the leader and follower arms have the same position values when they are in the same physical position. This calibration is essential because it allows a neural network trained on one Moss v1 robot to work on another.
+
+**Manual calibration of follower arm**
+/!\ Contrarily to step 6 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the auto calibration, we will actually do manual calibration of follower for now.
+
+You will need to move the follower arm to these positions sequentially:
+
+| 1. Zero position | 2. Rotated position | 3. Rest position |
+|---|---|---|
+| <img src="../media/moss/follower_zero.webp?raw=true" alt="Moss v1 follower arm zero position" title="Moss v1 follower arm zero position" style="width:100%;"> | <img src="../media/moss/follower_rotated.webp?raw=true" alt="Moss v1 follower arm rotated position" title="Moss v1 follower arm rotated position" style="width:100%;"> | <img src="../media/moss/follower_rest.webp?raw=true" alt="Moss v1 follower arm rest position" title="Moss v1 follower arm rest position" style="width:100%;"> |
+
+Make sure both arms are connected and run this script to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --robot-overrides '~cameras' --arms main_follower
+```
+
+**Manual calibration of leader arm**
+Follow step 6 of the [assembly video](https://www.youtube.com/watch?v=DA91NJOtMic) which illustrates the manual calibration. You will need to move the leader arm to these positions sequentially:
+
+| 1. Zero position | 2. Rotated position | 3. Rest position |
+|---|---|---|
+| <img src="../media/moss/leader_zero.webp?raw=true" alt="Moss v1 leader arm zero position" title="Moss v1 leader arm zero position" style="width:100%;"> | <img src="../media/moss/leader_rotated.webp?raw=true" alt="Moss v1 leader arm rotated position" title="Moss v1 leader arm rotated position" style="width:100%;"> | <img src="../media/moss/leader_rest.webp?raw=true" alt="Moss v1 leader arm rest position" title="Moss v1 leader arm rest position" style="width:100%;"> |
+
+Run this script to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --robot-overrides '~cameras' --arms main_leader
+```
+
+## Teleoperate
+
+**Simple teleop**
+Then you are ready to teleoperate your robot! Run this simple script (it won't connect and display the cameras):
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --robot-overrides '~cameras' \
+    --display-cameras 0
+```
+
+
+**Teleop with displaying cameras**
+Follow [this guide to setup your cameras](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#c-add-your-cameras-with-opencvcamera). Then you will be able to display the cameras on your computer while you are teleoperating by running the following code. This is useful to prepare your setup before recording your first dataset.
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/moss.yaml
+```
+
+## Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset with Moss v1.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py record \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --fps 30 \
+    --repo-id ${HF_USER}/moss_test \
+    --tags moss tutorial \
+    --warmup-time-s 5 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 2 \
+    --push-to-hub 1
+```
+
+## Visualize a dataset
+
+If you uploaded your dataset to the hub with `--push-to-hub 1`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+```bash
+echo ${HF_USER}/moss_test
+```
+
+If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/moss_test
+```
+
+## Replay an episode
+
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py replay \
+    --robot-path lerobot/configs/robot/moss.yaml \
+    --fps 30 \
+    --repo-id ${HF_USER}/moss_test \
+    --episode 0
+```
+
+## Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+python lerobot/scripts/train.py \
+  dataset_repo_id=${HF_USER}/moss_test \
+  policy=act_moss_real \
+  env=moss_real \
+  hydra.run.dir=outputs/train/act_moss_test \
+  hydra.job.name=act_moss_test \
+  device=cuda \
+  wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/moss_test`.
+2. We provided the policy with `policy=act_moss_real`. This loads configurations from [`lerobot/configs/policy/act_moss_real.yaml`](../lerobot/configs/policy/act_moss_real.yaml). Importantly, this policy uses 2 cameras as input `laptop`, `phone`.
+3. We provided an environment as argument with `env=moss_real`. This loads configurations from [`lerobot/configs/env/moss_real.yaml`](../lerobot/configs/env/moss_real.yaml).
+4. We provided `device=cuda` since we are training on a Nvidia GPU, but you can also use `device=mps` if you are using a Mac with Apple silicon, or `device=cpu` otherwise.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+
+Training should take several hours. You will find checkpoints in `outputs/train/act_moss_test/checkpoints`.
+
+## Evaluate your policy
+
+You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
+```bash
+python lerobot/scripts/control_robot.py record \
+  --robot-path lerobot/configs/robot/moss.yaml \
+  --fps 30 \
+  --repo-id ${HF_USER}/eval_act_moss_test \
+  --tags moss tutorial eval \
+  --warmup-time-s 5 \
+  --episode-time-s 40 \
+  --reset-time-s 10 \
+  --num-episodes 10 \
+  -p outputs/train/act_moss_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_moss_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_moss_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_moss_test`).
+
+## More
+
+Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth tutorial on controlling real robots with LeRobot.
+
+If you have any question or need help, please reach out on Discord in the channel [`#moss-arm`](https://discord.com/channels/1216765309076115607/1275374638985252925).

examples/12_train_hilserl_classifier.md

@@ -0,0 +1,83 @@
+# Training a HIL-SERL Reward Classifier with LeRobot
+
+This tutorial provides step-by-step instructions for training a reward classifier using LeRobot.
+
+---
+
+## Training Script Overview
+
+LeRobot includes a ready-to-use training script located at [`lerobot/scripts/train_hilserl_classifier.py`](../../lerobot/scripts/train_hilserl_classifier.py). Here's an outline of its workflow:
+
+1. **Configuration Loading**
+   The script uses Hydra to load a configuration file for subsequent steps. (Details on Hydra follow below.)
+
+2. **Dataset Initialization**
+   It loads a `LeRobotDataset` containing images and rewards. To optimize performance, a weighted random sampler is used to balance class sampling.
+
+3. **Classifier Initialization**
+   A lightweight classification head is built on top of a frozen, pretrained image encoder from HuggingFace. The classifier outputs either:
+   - A single probability (binary classification), or
+   - Logits (multi-class classification).
+
+4. **Training Loop Execution**
+   The script performs:
+   - Forward and backward passes,
+   - Optimization steps,
+   - Periodic logging, evaluation, and checkpoint saving.
+
+---
+
+## Configuring with Hydra
+
+For detailed information about Hydra usage, refer to [`examples/4_train_policy_with_script.md`](../examples/4_train_policy_with_script.md). However, note that training the reward classifier differs slightly and requires a separate configuration file.
+
+### Config File Setup
+
+The default `default.yaml` cannot launch the reward classifier training directly. Instead, you need a configuration file like [`lerobot/configs/policy/hilserl_classifier.yaml`](../../lerobot/configs/policy/hilserl_classifier.yaml), with the following adjustment:
+
+Replace the `dataset_repo_id` field with the identifier for your dataset, which contains images and sparse rewards:
+
+```yaml
+# Example: lerobot/configs/policy/reward_classifier.yaml
+dataset_repo_id: "my_dataset_repo_id"
+## Typical logs and metrics
+```
+When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
+
+After that, you will see training log like this one:
+
+```
+[2024-11-29 18:26:36,999][root][INFO] -
+Epoch 5/5
+Training:  82%|██████████████████████████████████████████████████████████████████████████████▋                 | 91/111 [00:50<00:09,  2.04it/s, loss=0.2999, acc=69.99%]
+```
+
+or evaluation log like:
+
+```
+Validation: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:20<00:00,  1.37it/s]
+```
+
+### Metrics Tracking with Weights & Biases (WandB)
+
+If `wandb.enable` is set to `true`, the training and evaluation logs will also be saved in WandB. This allows you to track key metrics in real-time, including:
+
+- **Training Metrics**:
+  - `train/accuracy`
+  - `train/loss`
+  - `train/dataloading_s`
+- **Evaluation Metrics**:
+  - `eval/accuracy`
+  - `eval/loss`
+  - `eval/eval_s`
+
+#### Additional Features
+
+You can also log sample predictions during evaluation. Each logged sample will include:
+
+- The **input image**.
+- The **predicted label**.
+- The **true label**.
+- The **classifier's "confidence" (logits/probability)**.
+
+These logs can be useful for diagnosing and debugging performance issues.

examples/1_load_lerobot_dataset.py

@@ -3,89 +3,132 @@ This script demonstrates the use of `LeRobotDataset` class for handling and proc
 It illustrates how to load datasets, manipulate them, and apply transformations suitable for machine learning tasks in PyTorch.
 
 Features included in this script:
-- Loading a dataset and accessing its properties.
-- Filtering data by episode number.
-- Converting tensor data for visualization.
-- Saving video files from dataset frames.
+- Viewing a dataset's metadata and exploring its properties.
+- Loading an existing dataset from the hub or a subset of it.
+- Accessing frames by episode number.
 - Using advanced dataset features like timestamp-based frame selection.
 - Demonstrating compatibility with PyTorch DataLoader for batch processing.
 
 The script ends with examples of how to batch process data using PyTorch's DataLoader.
 """
 
-from pathlib import Path
+from pprint import pprint
 
-import imageio
 import torch
+from huggingface_hub import HfApi
 
 import lerobot
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
 
-print("List of available datasets", lerobot.available_datasets)
-# # >>> ['lerobot/aloha_sim_insertion_human', 'lerobot/aloha_sim_insertion_scripted',
-# #     'lerobot/aloha_sim_transfer_cube_human', 'lerobot/aloha_sim_transfer_cube_scripted',
-# #     'lerobot/pusht', 'lerobot/xarm_lift_medium']
+# We ported a number of existing datasets ourselves, use this to see the list:
+print("List of available datasets:")
+pprint(lerobot.available_datasets)
 
-repo_id = "lerobot/pusht"
+# You can also browse through the datasets created/ported by the community on the hub using the hub api:
+hub_api = HfApi()
+repo_ids = [info.id for info in hub_api.list_datasets(task_categories="robotics", tags=["LeRobot"])]
+pprint(repo_ids)
 
-# You can easily load a dataset from a Hugging Face repositery
+# Or simply explore them in your web browser directly at:
+# https://huggingface.co/datasets?other=LeRobot
+
+# Let's take this one for this example
+repo_id = "lerobot/aloha_mobile_cabinet"
+# We can have a look and fetch its metadata to know more about it:
+ds_meta = LeRobotDatasetMetadata(repo_id)
+
+# By instantiating just this class, you can quickly access useful information about the content and the
+# structure of the dataset without downloading the actual data yet (only metadata files — which are
+# lightweight).
+print(f"Total number of episodes: {ds_meta.total_episodes}")
+print(f"Average number of frames per episode: {ds_meta.total_frames / ds_meta.total_episodes:.3f}")
+print(f"Frames per second used during data collection: {ds_meta.fps}")
+print(f"Robot type: {ds_meta.robot_type}")
+print(f"keys to access images from cameras: {ds_meta.camera_keys=}\n")
+
+print("Tasks:")
+print(ds_meta.tasks)
+print("Features:")
+pprint(ds_meta.features)
+
+# You can also get a short summary by simply printing the object:
+print(ds_meta)
+
+# You can then load the actual dataset from the hub.
+# Either load any subset of episodes:
+dataset = LeRobotDataset(repo_id, episodes=[0, 10, 11, 23])
+
+# And see how many frames you have:
+print(f"Selected episodes: {dataset.episodes}")
+print(f"Number of episodes selected: {dataset.num_episodes}")
+print(f"Number of frames selected: {dataset.num_frames}")
+
+# Or simply load the entire dataset:
 dataset = LeRobotDataset(repo_id)
+print(f"Number of episodes selected: {dataset.num_episodes}")
+print(f"Number of frames selected: {dataset.num_frames}")
 
-# LeRobotDataset is actually a thin wrapper around an underlying Hugging Face dataset  (see https://huggingface.co/docs/datasets/index for more information).
-# TODO(rcadene): update to make the print pretty
-print(f"{dataset=}")
-print(f"{dataset.hf_dataset=}")
+# The previous metadata class is contained in the 'meta' attribute of the dataset:
+print(dataset.meta)
 
-# and provides additional utilities for robotics and compatibility with pytorch
-print(f"number of samples/frames: {dataset.num_samples=}")
-print(f"number of episodes: {dataset.num_episodes=}")
-print(f"average number of frames per episode: {dataset.num_samples / dataset.num_episodes:.3f}")
-print(f"frames per second used during data collection: {dataset.fps=}")
-print(f"keys to access images from cameras: {dataset.image_keys=}")
+# LeRobotDataset actually wraps an underlying Hugging Face dataset
+# (see https://huggingface.co/docs/datasets for more information).
+print(dataset.hf_dataset)
 
-# While the LeRobotDataset adds helpers for working within our library, we still expose the underling Hugging Face dataset.
-# It may be freely replaced or modified in place. Here we use the filtering to keep only frames from episode 5.
-# TODO(rcadene): remove this example of accessing hf_dataset
-dataset.hf_dataset = dataset.hf_dataset.filter(lambda frame: frame["episode_index"] == 5)
+# LeRobot datasets also subclasses PyTorch datasets so you can do everything you know and love from working
+# with the latter, like iterating through the dataset.
+# The __getitem__ iterates over the frames of the dataset. Since our datasets are also structured by
+# episodes, you can access the frame indices of any episode using the episode_data_index. Here, we access
+# frame indices associated to the first episode:
+episode_index = 0
+from_idx = dataset.episode_data_index["from"][episode_index].item()
+to_idx = dataset.episode_data_index["to"][episode_index].item()
 
-# LeRobot datsets actually subclass PyTorch datasets. So you can do everything you know and love from working with the latter, for example: iterating through the dataset. Here we grab all the image frames.
-frames = [sample["observation.image"] for sample in dataset]
+# Then we grab all the image frames from the first camera:
+camera_key = dataset.meta.camera_keys[0]
+frames = [dataset[idx][camera_key] for idx in range(from_idx, to_idx)]
 
-# but frames are now float32 range [0,1] channel first (c,h,w) to follow pytorch convention,
-# to view them, we convert to uint8 range [0,255]
-frames = [(frame * 255).type(torch.uint8) for frame in frames]
-# and to channel last (h,w,c)
-frames = [frame.permute((1, 2, 0)).numpy() for frame in frames]
+# The objects returned by the dataset are all torch.Tensors
+print(type(frames[0]))
+print(frames[0].shape)
 
-# and finally save them to a mp4 video
-Path("outputs/examples/1_load_lerobot_dataset").mkdir(parents=True, exist_ok=True)
-imageio.mimsave("outputs/examples/1_load_lerobot_dataset/episode_5.mp4", frames, fps=dataset.fps)
+# Since we're using pytorch, the shape is in pytorch, channel-first convention (c, h, w).
+# We can compare this shape with the information available for that feature
+pprint(dataset.features[camera_key])
+# In particular:
+print(dataset.features[camera_key]["shape"])
+# The shape is in (h, w, c) which is a more universal format.
 
-# For many machine learning applications we need to load histories of past observations, or trajectorys of future actions. Our datasets can load previous and future frames for each key/modality,
-# using timestamps differences with the current loaded frame. For instance:
+# For many machine learning applications we need to load the history of past observations or trajectories of
+# future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
+# differences with the current loaded frame. For instance:
 delta_timestamps = {
     # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
-    "observation.image": [-1, -0.5, -0.20, 0],
-    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 20 ms, 10 ms, and current frame
-    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, -0.02, -0.01, 0],
+    camera_key: [-1, -0.5, -0.20, 0],
+    # loads 8 state vectors: 1.5 seconds before, 1 second before, ... 200 ms, 100 ms, and current frame
+    "observation.state": [-1.5, -1, -0.5, -0.20, -0.10, 0],
     # loads 64 action vectors: current frame, 1 frame in the future, 2 frames, ... 63 frames in the future
     "action": [t / dataset.fps for t in range(64)],
 }
-dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
-print(f"{dataset[0]['observation.image'].shape=}")  # (4,c,h,w)
-print(f"{dataset[0]['observation.state'].shape=}")  # (8,c)
-print(f"{dataset[0]['action'].shape=}")  # (64,c)
+# Note that in any case, these delta_timestamps values need to be multiples of (1/fps) so that added to any
+# timestamp, you still get a valid timestamp.
 
-# Finally, our datasets are fully compatible with PyTorch dataloaders and samplers
-# because they are just PyTorch datasets.
+dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
+print(f"\n{dataset[0][camera_key].shape=}")  # (4, c, h, w)
+print(f"{dataset[0]['observation.state'].shape=}")  # (6, c)
+print(f"{dataset[0]['action'].shape=}\n")  # (64, c)
+
+# Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
+# PyTorch datasets.
 dataloader = torch.utils.data.DataLoader(
     dataset,
     num_workers=0,
     batch_size=32,
     shuffle=True,
 )
+
 for batch in dataloader:
-    print(f"{batch['observation.image'].shape=}")  # (32,4,c,h,w)
-    print(f"{batch['observation.state'].shape=}")  # (32,8,c)
-    print(f"{batch['action'].shape=}")  # (32,64,c)
+    print(f"{batch[camera_key].shape=}")  # (32, 4, c, h, w)
+    print(f"{batch['observation.state'].shape=}")  # (32, 5, c)
+    print(f"{batch['action'].shape=}")  # (32, 64, c)
     break

examples/2_evaluate_pretrained_policy.py

@@ -5,34 +5,117 @@ training outputs directory. In the latter case, you might want to run examples/3
 
 from pathlib import Path
 
+import gym_pusht  # noqa: F401
+import gymnasium as gym
+import imageio
+import numpy
+import torch
 from huggingface_hub import snapshot_download
 
-from lerobot.common.utils.utils import init_hydra_config
-from lerobot.scripts.eval import eval
+from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 
-# Get a pretrained policy from the hub.
-# TODO(alexander-soare): This no longer works until we upload a new model that uses the current configs.
-hub_id = "lerobot/diffusion_policy_pusht_image"
-folder = Path(snapshot_download(hub_id))
+# Create a directory to store the video of the evaluation
+output_directory = Path("outputs/eval/example_pusht_diffusion")
+output_directory.mkdir(parents=True, exist_ok=True)
+
+# Download the diffusion policy for pusht environment
+pretrained_policy_path = Path(snapshot_download("lerobot/diffusion_pusht"))
 # OR uncomment the following to evaluate a policy from the local outputs/train folder.
-# folder = Path("outputs/train/example_pusht_diffusion")
+# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
 
-config_path = folder / "config.yaml"
-weights_path = folder / "model.pt"
+policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
+policy.eval()
 
-# Override some config parameters to do with evaluation.
-overrides = [
-    f"policy.pretrained_model_path={weights_path}",
-    "eval_episodes=10",
-    "rollout_batch_size=10",
-    "device=cuda",
-]
+# Check if GPU is available
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+    print("GPU is available. Device set to:", device)
+else:
+    device = torch.device("cpu")
+    print(f"GPU is not available. Device set to: {device}. Inference will be slower than on GPU.")
+    # Decrease the number of reverse-diffusion steps (trades off a bit of quality for 10x speed)
+    policy.diffusion.num_inference_steps = 10
 
-# Create a Hydra config.
-cfg = init_hydra_config(config_path, overrides)
+policy.to(device)
 
-# Evaluate the policy and save the outputs including metrics and videos.
-eval(
-    cfg,
-    out_dir=f"outputs/eval/example_{cfg.env.name}_{cfg.policy.name}",
+# Initialize evaluation environment to render two observation types:
+# an image of the scene and state/position of the agent. The environment
+# also automatically stops running after 300 interactions/steps.
+env = gym.make(
+    "gym_pusht/PushT-v0",
+    obs_type="pixels_agent_pos",
+    max_episode_steps=300,
 )
+
+# Reset the policy and environmens to prepare for rollout
+policy.reset()
+numpy_observation, info = env.reset(seed=42)
+
+# Prepare to collect every rewards and all the frames of the episode,
+# from initial state to final state.
+rewards = []
+frames = []
+
+# Render frame of the initial state
+frames.append(env.render())
+
+step = 0
+done = False
+while not done:
+    # Prepare observation for the policy running in Pytorch
+    state = torch.from_numpy(numpy_observation["agent_pos"])
+    image = torch.from_numpy(numpy_observation["pixels"])
+
+    # Convert to float32 with image from channel first in [0,255]
+    # to channel last in [0,1]
+    state = state.to(torch.float32)
+    image = image.to(torch.float32) / 255
+    image = image.permute(2, 0, 1)
+
+    # Send data tensors from CPU to GPU
+    state = state.to(device, non_blocking=True)
+    image = image.to(device, non_blocking=True)
+
+    # Add extra (empty) batch dimension, required to forward the policy
+    state = state.unsqueeze(0)
+    image = image.unsqueeze(0)
+
+    # Create the policy input dictionary
+    observation = {
+        "observation.state": state,
+        "observation.image": image,
+    }
+
+    # Predict the next action with respect to the current observation
+    with torch.inference_mode():
+        action = policy.select_action(observation)
+
+    # Prepare the action for the environment
+    numpy_action = action.squeeze(0).to("cpu").numpy()
+
+    # Step through the environment and receive a new observation
+    numpy_observation, reward, terminated, truncated, info = env.step(numpy_action)
+    print(f"{step=} {reward=} {terminated=}")
+
+    # Keep track of all the rewards and frames
+    rewards.append(reward)
+    frames.append(env.render())
+
+    # The rollout is considered done when the success state is reach (i.e. terminated is True),
+    # or the maximum number of iterations is reached (i.e. truncated is True)
+    done = terminated | truncated | done
+    step += 1
+
+if terminated:
+    print("Success!")
+else:
+    print("Failure!")
+
+# Get the speed of environment (i.e. its number of frames per second).
+fps = env.metadata["render_fps"]
+
+# Encode all frames into a mp4 video.
+video_path = output_directory / "rollout.mp4"
+imageio.mimsave(str(video_path), numpy.stack(frames), fps=fps)
+
+print(f"Video of the evaluation is available in '{video_path}'.")

examples/3_train_policy.py

@@ -4,45 +4,53 @@ Once you have trained a model with this script, you can try to evaluate it on
 examples/2_evaluate_pretrained_policy.py
 """
 
-import os
 from pathlib import Path
 
 import torch
-from omegaconf import OmegaConf
 
-from lerobot.common.datasets.factory import make_dataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
-from lerobot.common.utils.utils import init_hydra_config
 
+# Create a directory to store the training checkpoint.
 output_directory = Path("outputs/train/example_pusht_diffusion")
-os.makedirs(output_directory, exist_ok=True)
+output_directory.mkdir(parents=True, exist_ok=True)
 
-# Number of offline training steps (we'll only do offline training for this example.
+# Number of offline training steps (we'll only do offline training for this example.)
 # Adjust as you prefer. 5000 steps are needed to get something worth evaluating.
 training_steps = 5000
 device = torch.device("cuda")
 log_freq = 250
 
 # Set up the dataset.
-hydra_cfg = init_hydra_config("lerobot/configs/default.yaml", overrides=["env=pusht"])
-dataset = make_dataset(hydra_cfg)
+delta_timestamps = {
+    # Load the previous image and state at -0.1 seconds before current frame,
+    # then load current image and state corresponding to 0.0 second.
+    "observation.image": [-0.1, 0.0],
+    "observation.state": [-0.1, 0.0],
+    # Load the previous action (-0.1), the next action to be executed (0.0),
+    # and 14 future actions with a 0.1 seconds spacing. All these actions will be
+    # used to supervise the policy.
+    "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
+}
+dataset = LeRobotDataset("lerobot/pusht", delta_timestamps=delta_timestamps)
 
 # Set up the the policy.
 # Policies are initialized with a configuration class, in this case `DiffusionConfig`.
 # For this example, no arguments need to be passed because the defaults are set up for PushT.
 # If you're doing something different, you will likely need to change at least some of the defaults.
 cfg = DiffusionConfig()
-# TODO(alexander-soare): Remove LR scheduler from the policy.
-policy = DiffusionPolicy(cfg, lr_scheduler_num_training_steps=training_steps, dataset_stats=dataset.stats)
+policy = DiffusionPolicy(cfg, dataset_stats=dataset.meta.stats)
 policy.train()
 policy.to(device)
 
+optimizer = torch.optim.Adam(policy.parameters(), lr=1e-4)
+
 # Create dataloader for offline training.
 dataloader = torch.utils.data.DataLoader(
     dataset,
     num_workers=4,
-    batch_size=cfg.batch_size,
+    batch_size=64,
     shuffle=True,
     pin_memory=device != torch.device("cpu"),
     drop_last=True,
@@ -54,14 +62,18 @@ done = False
 while not done:
     for batch in dataloader:
         batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
-        info = policy.update(batch)
+        output_dict = policy.forward(batch)
+        loss = output_dict["loss"]
+        loss.backward()
+        optimizer.step()
+        optimizer.zero_grad()
+
         if step % log_freq == 0:
-            print(f"step: {step} loss: {info['loss']:.3f} update_time: {info['update_s']:.3f} (seconds)")
+            print(f"step: {step} loss: {loss.item():.3f}")
         step += 1
         if step >= training_steps:
             done = True
             break
 
-# Save the policy and configuration for later use.
-policy.save(output_directory / "model.pt")
-OmegaConf.save(hydra_cfg, output_directory / "config.yaml")
+# Save a policy checkpoint.
+policy.save_pretrained(output_directory)

examples/4_train_policy_with_script.md

@@ -0,0 +1,213 @@
+This tutorial will explain the training script, how to use it, and particularly the use of Hydra to configure everything needed for the training run.
+
+## The training script
+
+LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
+
+- Loads a Hydra configuration file for the following steps (more on Hydra in a moment).
+- Makes a simulation environment.
+- Makes a dataset corresponding to that simulation environment.
+- Makes a policy.
+- Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
+
+## Basics of how we use Hydra
+
+Explaining the ins and outs of [Hydra](https://hydra.cc/docs/intro/) is beyond the scope of this document, but here we'll share the main points you need to know.
+
+First, `lerobot/configs` has a directory structure like this:
+
+```
+.
+├── default.yaml
+├── env
+│   ├── aloha.yaml
+│   ├── pusht.yaml
+│   └── xarm.yaml
+└── policy
+    ├── act.yaml
+    ├── diffusion.yaml
+    └── tdmpc.yaml
+```
+
+**_For brevity, in the rest of this document we'll drop the leading `lerobot/configs` path. So `default.yaml` really refers to `lerobot/configs/default.yaml`._**
+
+When you run the training script with
+
+```python
+python lerobot/scripts/train.py
+```
+
+Hydra is set up to read `default.yaml` (via the `@hydra.main` decorator). If you take a look at the `@hydra.main`'s arguments you will see `config_path="../configs", config_name="default"`. At the top of `default.yaml`, is a `defaults` section which looks likes this:
+
+```yaml
+defaults:
+  - _self_
+  - env: pusht
+  - policy: diffusion
+```
+
+This logic tells Hydra to incorporate configuration parameters from `env/pusht.yaml` and `policy/diffusion.yaml`. _Note: Be aware of the order as any configuration parameters with the same name will be overidden. Thus, `default.yaml` is overridden by `env/pusht.yaml`  which is overidden by `policy/diffusion.yaml`_.
+
+Then, `default.yaml` also contains common configuration parameters such as `device: cuda` or `use_amp: false` (for enabling fp16 training). Some other parameters are set to `???` which indicates that they are expected to be set in additional yaml files. For instance, `training.offline_steps: ???` in `default.yaml` is set to `200000` in `diffusion.yaml`.
+
+Thanks to this `defaults` section in `default.yaml`, if you want to train Diffusion Policy with PushT, you really only need to run:
+
+```bash
+python lerobot/scripts/train.py
+```
+
+However, you can be more explicit and launch the exact same Diffusion Policy training on PushT with:
+
+```bash
+python lerobot/scripts/train.py policy=diffusion env=pusht
+```
+
+This way of overriding defaults via the CLI is especially useful when you want to change the policy and/or environment. For instance, you can train ACT on the default Aloha environment with:
+
+```bash
+python lerobot/scripts/train.py policy=act env=aloha
+```
+
+There are two things to note here:
+- Config overrides are passed as `param_name=param_value`.
+- Here we have overridden the defaults section. `policy=act` tells Hydra to use `policy/act.yaml`, and `env=aloha` tells Hydra to use `env/aloha.yaml`.
+
+_As an aside: we've set up all of our configurations so that they reproduce state-of-the-art results from papers in the literature._
+
+## Overriding configuration parameters in the CLI
+
+Now let's say that we want to train on a different task in the Aloha environment. If you look in `env/aloha.yaml` you will see something like:
+
+```yaml
+# lerobot/configs/env/aloha.yaml
+env:
+  task: AlohaInsertion-v0
+```
+
+And if you look in `policy/act.yaml` you will see something like:
+
+```yaml
+# lerobot/configs/policy/act.yaml
+dataset_repo_id: lerobot/aloha_sim_insertion_human
+```
+
+But our Aloha environment actually supports a cube transfer task as well. To train for this task, you could manually modify the two yaml configuration files respectively.
+
+First, we'd need to switch to using the cube transfer task for the ALOHA environment.
+
+```diff
+# lerobot/configs/env/aloha.yaml
+env:
+-  task: AlohaInsertion-v0
++  task: AlohaTransferCube-v0
+```
+
+Then, we'd also need to switch to using the cube transfer dataset.
+
+```diff
+# lerobot/configs/policy/act.yaml
+-dataset_repo_id: lerobot/aloha_sim_insertion_human
++dataset_repo_id: lerobot/aloha_sim_transfer_cube_human
+```
+
+Then, you'd be able to run:
+
+```bash
+python lerobot/scripts/train.py policy=act env=aloha
+```
+
+and you'd be training and evaluating on the cube transfer task.
+
+An alternative approach to editing the yaml configuration files, would be to override the defaults via the command line:
+
+```bash
+python lerobot/scripts/train.py \
+    policy=act \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    env=aloha \
+    env.task=AlohaTransferCube-v0
+```
+
+There's something new here. Notice the `.` delimiter used to traverse the configuration hierarchy. _But be aware that the `defaults` section is an exception. As you saw above, we didn't need to write `defaults.policy=act` in the CLI. `policy=act` was enough._
+
+Putting all that knowledge together, here's the command that was used to train https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human.
+
+```bash
+python lerobot/scripts/train.py \
+    hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human \
+    device=cuda
+    env=aloha \
+    env.task=AlohaTransferCube-v0 \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    policy=act \
+    training.eval_freq=10000 \
+    training.log_freq=250 \
+    training.offline_steps=100000 \
+    training.save_model=true \
+    training.save_freq=25000 \
+    eval.n_episodes=50 \
+    eval.batch_size=50 \
+    wandb.enable=false \
+```
+
+There's one new thing here: `hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human`, which specifies where to save the training output.
+
+## Using a configuration file not in `lerobot/configs`
+
+Above we discusses the our training script is set up such that Hydra looks for `default.yaml` in `lerobot/configs`. But, if you have a configuration file elsewhere in your filesystem you may use:
+
+```bash
+python lerobot/scripts/train.py --config-dir PARENT/PATH --config-name FILE_NAME_WITHOUT_EXTENSION
+```
+
+Note: here we use regular syntax for providing CLI arguments to a Python script, not Hydra's `param_name=param_value` syntax.
+
+As a concrete example, this becomes particularly handy when you have a folder with training outputs, and would like to re-run the training. For example, say you previously ran the training script with one of the earlier commands and have `outputs/train/my_experiment/checkpoints/pretrained_model/config.yaml`. This `config.yaml` file will have the full set of configuration parameters within it. To run the training with the same configuration again, do:
+
+```bash
+python lerobot/scripts/train.py --config-dir outputs/train/my_experiment/checkpoints/last/pretrained_model --config-name config
+```
+
+Note that you may still use the regular syntax for config parameter overrides (eg: by adding `training.offline_steps=200000`).
+
+## Typical logs and metrics
+
+When you start the training process, you will first see your full configuration being printed in the terminal. You can check it to make sure that you config it correctly and your config is not overrided by other files. The final configuration will also be saved with the checkpoint.
+
+After that, you will see training log like this one:
+
+```
+INFO 2024-08-14 13:35:12 ts/train.py:192 step:0 smpl:64 ep:1 epch:0.00 loss:1.112 grdn:15.387 lr:2.0e-07 updt_s:1.738 data_s:4.774
+```
+
+or evaluation log like:
+
+```
+INFO 2024-08-14 13:38:45 ts/train.py:226 step:100 smpl:6K ep:52 epch:0.25 ∑rwrd:20.693 success:0.0% eval_s:120.266
+```
+
+These logs will also be saved in wandb if `wandb.enable` is set to `true`. Here are the meaning of some abbreviations:
+
+- `smpl`: number of samples seen during training.
+- `ep`: number of episodes seen during training. An episode contains multiple samples in a complete manipulation task.
+- `epch`: number of time all unique samples are seen (epoch).
+- `grdn`: gradient norm.
+- `∑rwrd`: compute the sum of rewards in every evaluation episode and then take an average of them.
+- `success`: average success rate of eval episodes. Reward and success are usually different except for the sparsing reward setting, where reward=1 only when the task is completed successfully.
+- `eval_s`: time to evaluate the policy in the environment, in second.
+- `updt_s`: time to update the network parameters, in second.
+- `data_s`: time to load a batch of data, in second.
+
+Some metrics are useful for initial performance profiling. For example, if you find the current GPU utilization is low via the `nvidia-smi` command and `data_s` sometimes is too high, you may need to modify batch size or number of dataloading workers to accelerate dataloading. We also recommend [pytorch profiler](https://github.com/huggingface/lerobot?tab=readme-ov-file#improve-your-code-with-profiling) for detailed performance probing.
+
+---
+
+So far we've seen how to train Diffusion Policy for PushT and ACT for ALOHA. Now, what if we want to train ACT for PushT? Well, there are aspects of the ACT configuration that are specific to the ALOHA environments, and these happen to be incompatible with PushT. Therefore, trying to run the following will almost certainly raise an exception of sorts (eg: feature dimension mismatch):
+
+```bash
+python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
+```
+
+Please, head on over to our [advanced tutorial on adapting policy configuration to various environments](./advanced/train_act_pusht/train_act_pusht.md) to learn more.
+
+Or in the meantime, happy coding! 🤗

examples/5_resume_training.md

@@ -0,0 +1,37 @@
+This tutorial explains how to resume a training run that you've started with the training script. If you don't know how our training script and configuration system works, please read [4_train_policy_with_script.md](./4_train_policy_with_script.md) first.
+
+## Basic training resumption
+
+Let's consider the example of training ACT for one of the ALOHA tasks. Here's a command that can achieve that:
+
+```bash
+python lerobot/scripts/train.py \
+    hydra.run.dir=outputs/train/run_resumption \
+    policy=act \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    env=aloha \
+    env.task=AlohaTransferCube-v0 \
+    training.log_freq=25 \
+    training.save_checkpoint=true \
+    training.save_freq=100
+```
+
+Here we're using the default dataset and environment for ACT, and we've taken care to set up the log frequency and checkpointing frequency to low numbers so we can test resumption. You should be able to see some logging and have a first checkpoint within 1 minute. Please interrupt the training after the first checkpoint.
+
+To resume, all that we have to do is run the training script, providing the run directory, and the resume option:
+
+```bash
+python lerobot/scripts/train.py \
+    hydra.run.dir=outputs/train/run_resumption \
+    resume=true
+```
+
+You should see from the logging that your training picks up from where it left off.
+
+Note that with `resume=true`, the configuration file from the last checkpoint in the training output directory is loaded. So it doesn't matter that we haven't provided all the other configuration parameters from our previous command (although there may be warnings to notify you that your command has a different configuration than than the checkpoint).
+
+---
+
+Now you should know how to resume your training run in case it gets interrupted or you want to extend a finished training run.
+
+Happy coding! 🤗

examples/6_add_image_transforms.py

@@ -0,0 +1,53 @@
+"""
+This script demonstrates how to use torchvision's image transformation with LeRobotDataset for data
+augmentation purposes. The transformations are passed to the dataset as an argument upon creation, and
+transforms are applied to the observation images before they are returned in the dataset's __getitem__.
+"""
+
+from pathlib import Path
+
+from torchvision.transforms import ToPILImage, v2
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+dataset_repo_id = "lerobot/aloha_static_screw_driver"
+
+# Create a LeRobotDataset with no transformations
+dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
+# This is equivalent to `dataset = LeRobotDataset(dataset_repo_id, image_transforms=None)`
+
+# Get the index of the first observation in the first episode
+first_idx = dataset.episode_data_index["from"][0].item()
+
+# Get the frame corresponding to the first camera
+frame = dataset[first_idx][dataset.meta.camera_keys[0]]
+
+
+# Define the transformations
+transforms = v2.Compose(
+    [
+        v2.ColorJitter(brightness=(0.5, 1.5)),
+        v2.ColorJitter(contrast=(0.5, 1.5)),
+        v2.ColorJitter(hue=(-0.1, 0.1)),
+        v2.RandomAdjustSharpness(sharpness_factor=2, p=1),
+    ]
+)
+
+# Create another LeRobotDataset with the defined transformations
+transformed_dataset = LeRobotDataset(dataset_repo_id, episodes=[0], image_transforms=transforms)
+
+# Get a frame from the transformed dataset
+transformed_frame = transformed_dataset[first_idx][transformed_dataset.meta.camera_keys[0]]
+
+# Create a directory to store output images
+output_dir = Path("outputs/image_transforms")
+output_dir.mkdir(parents=True, exist_ok=True)
+
+# Save the original frame
+to_pil = ToPILImage()
+to_pil(frame).save(output_dir / "original_frame.png", quality=100)
+print(f"Original frame saved to {output_dir / 'original_frame.png'}.")
+
+# Save the transformed frame
+to_pil(transformed_frame).save(output_dir / "transformed_frame.png", quality=100)
+print(f"Transformed frame saved to {output_dir / 'transformed_frame.png'}.")

examples/8_use_stretch.md

@@ -0,0 +1,156 @@
+This tutorial explains how to use [Stretch 3](https://hello-robot.com/stretch-3-product) with LeRobot.
+
+## Setup
+
+Familiarize yourself with Stretch by following its [tutorials](https://docs.hello-robot.com/0.3/getting_started/hello_robot/) (recommended).
+
+To use LeRobot on Stretch, 3 options are available:
+- [tethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#tethered-setup)
+- [untethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#untethered-setup)
+- ssh directly into Stretch (you will first need to install and configure openssh-server on stretch using one of the two above setups)
+
+
+## Install LeRobot
+
+On Stretch's CLI, follow these steps:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
+bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+rm ~/miniconda3/miniconda.sh
+~/miniconda3/bin/conda init bash
+```
+
+2. Comment out these lines in `~/.profile` (this can mess up paths used by conda and ~/.local/bin should already be in your PATH)
+```
+# set PATH so it includes user's private bin if it exists
+if [ -d "$HOME/.local/bin" ] ; then
+    PATH="$HOME/.local/bin:$PATH"
+fi
+```
+
+3. Restart shell or `source ~/.bashrc`
+
+4. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+5. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+6. Install LeRobot with stretch dependencies:
+```bash
+cd ~/lerobot && pip install -e ".[stretch]"
+```
+
+> **Note:** If you get this message, you can ignore it: `ERROR: pip's dependency resolver does not currently take into account all the packages that are installed.`
+
+For Linux only (not Mac), install extra dependencies for recording datasets:
+```bash
+conda install -y -c conda-forge ffmpeg
+pip uninstall -y opencv-python
+conda install -y -c conda-forge "opencv>=4.10.0"
+```
+
+7. Run a [system check](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#system-check) to make sure your robot is ready:
+```bash
+stretch_system_check.py
+```
+
+> **Note:** You may need to free the "robot process" after booting Stretch by running `stretch_free_robot_process.py`. For more info this Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#turning-off-gamepad-teleoperation).
+
+You should get something like this:
+```bash
+For use with S T R E T C H (R) from Hello Robot Inc.
+---------------------------------------------------------------------
+
+Model = Stretch 3
+Tool = DexWrist 3 w/ Gripper
+Serial Number = stretch-se3-3054
+
+---- Checking Hardware ----
+[Pass] Comms are ready
+[Pass] Actuators are ready
+[Warn] Sensors not ready (IMU AZ = -10.19 out of range -10.1 to -9.5)
+[Pass] Battery voltage is 13.6 V
+
+---- Checking Software ----
+[Pass] Ubuntu 22.04 is ready
+[Pass] All APT pkgs are setup correctly
+[Pass] Firmware is up-to-date
+[Pass] Python pkgs are up-to-date
+[Pass] ROS2 Humble is ready
+```
+
+## Teleoperate, record a dataset and run a policy
+
+**Calibrate (Optional)**
+Before operating Stretch, you need to [home](https://docs.hello-robot.com/0.3/getting_started/stretch_hardware_overview/#homing) it first. Be mindful about giving Stretch some space as this procedure will move the robot's arm and gripper. Now run this command:
+```bash
+python lerobot/scripts/control_robot.py calibrate \
+    --robot-path lerobot/configs/robot/stretch.yaml
+```
+This is equivalent to running `stretch_robot_home.py`
+
+> **Note:** If you run any of the LeRobot scripts below and Stretch is not poperly homed, it will automatically home/calibrate first.
+
+**Teleoperate**
+Before trying teleoperation, you need activate the gamepad controller by pressing the middle button. For more info, see Stretch's [doc](https://docs.hello-robot.com/0.3/getting_started/hello_robot/#gamepad-teleoperation).
+
+Now try out teleoperation (see above documentation to learn about the gamepad controls):
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/stretch.yaml
+```
+This is essentially the same as running `stretch_gamepad_teleop.py`
+
+**Record a dataset**
+Once you're familiar with the gamepad controls and after a bit of practice, you can try to record your first dataset with Stretch.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record one episode:
+```bash
+python lerobot/scripts/control_robot.py record \
+    --robot-path lerobot/configs/robot/stretch.yaml \
+    --fps 20 \
+    --repo-id ${HF_USER}/stretch_test \
+    --tags stretch tutorial \
+    --warmup-time-s 3 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 1 \
+    --push-to-hub 0
+```
+
+> **Note:** If you're using ssh to connect to Stretch and run this script, you won't be able to visualize its cameras feed (though they will still be recording). To see the cameras stream, use [tethered](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#tethered-setup) or [untethered setup](https://docs.hello-robot.com/0.3/getting_started/connecting_to_stretch/#untethered-setup).
+
+**Replay an episode**
+Now try to replay this episode (make sure the robot's initial position is the same):
+```bash
+python lerobot/scripts/control_robot.py replay \
+    --robot-path lerobot/configs/robot/stretch.yaml \
+    --fps 20 \
+    --repo-id ${HF_USER}/stretch_test \
+    --episode 0
+```
+
+Follow [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) to train a policy on your data and run inference on your robot. You will need to adapt the code for Stretch.
+
+> TODO(rcadene, aliberts): Add already setup environment and policy yaml configuration files
+
+If you need help, please reach out on Discord in the channel `#stretch3-mobile-arm`.

examples/9_use_aloha.md

@@ -0,0 +1,174 @@
+This tutorial explains how to use [Aloha and Aloha 2 stationary](https://www.trossenrobotics.com/aloha-stationary) with LeRobot.
+
+## Setup
+
+Follow the [documentation from Trossen Robotics](https://docs.trossenrobotics.com/aloha_docs/getting_started/stationary/hardware_setup.html) for setting up the hardware and plugging the 4 arms and 4 cameras to your computer.
+
+
+## Install LeRobot
+
+On your computer:
+
+1. [Install Miniconda](https://docs.anaconda.com/miniconda/#quick-command-line-install):
+```bash
+mkdir -p ~/miniconda3
+wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
+bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
+rm ~/miniconda3/miniconda.sh
+~/miniconda3/bin/conda init bash
+```
+
+2. Restart shell or `source ~/.bashrc`
+
+3. Create and activate a fresh conda environment for lerobot
+```bash
+conda create -y -n lerobot python=3.10 && conda activate lerobot
+```
+
+4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+5. Install LeRobot with dependencies for the Aloha motors (dynamixel) and cameras (intelrealsense):
+```bash
+cd ~/lerobot && pip install -e ".[dynamixel, intelrealsense]"
+```
+
+For Linux only (not Mac), install extra dependencies for recording datasets:
+```bash
+conda install -y -c conda-forge ffmpeg
+pip uninstall -y opencv-python
+conda install -y -c conda-forge "opencv>=4.10.0"
+```
+
+## Teleoperate
+
+**/!\ FOR SAFETY, READ THIS /!\**
+Teleoperation consists in manually operating the leader arms to move the follower arms. Importantly:
+1. Make sure your leader arms are in the same position as the follower arms, so that the follower arms don't move too fast to match the leader arms,
+2. Our code assumes that your robot has been assembled following Trossen Robotics instructions. This allows us to skip calibration, as we use the pre-defined calibration files in `.cache/calibration/aloha_default`. If you replace a motor, make sure you follow the exact instructions from Trossen Robotics.
+
+By running the following code, you can start your first **SAFE** teleoperation:
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=5
+```
+
+By adding `--robot-overrides max_relative_target=5`, we override the default value for `max_relative_target` defined in `lerobot/configs/robot/aloha.yaml`. It is expected to be `5` to limit the magnitude of the movement for more safety, but the teleoperation won't be smooth. When you feel confident, you can disable this limit by adding `--robot-overrides max_relative_target=null` to the command line:
+```bash
+python lerobot/scripts/control_robot.py teleoperate \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=null
+```
+
+## Record a dataset
+
+Once you're familiar with teleoperation, you can record your first dataset with Aloha.
+
+If you want to use the Hugging Face hub features for uploading your dataset and you haven't previously done it, make sure you've logged in using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
+```bash
+huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
+```
+
+Store your Hugging Face repository name in a variable to run these commands:
+```bash
+HF_USER=$(huggingface-cli whoami | head -n 1)
+echo $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py record \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=null \
+    --fps 30 \
+    --repo-id ${HF_USER}/aloha_test \
+    --tags aloha tutorial \
+    --warmup-time-s 5 \
+    --episode-time-s 40 \
+    --reset-time-s 10 \
+    --num-episodes 2 \
+    --push-to-hub 1
+```
+
+## Visualize a dataset
+
+If you uploaded your dataset to the hub with `--push-to-hub 1`, you can [visualize your dataset online](https://huggingface.co/spaces/lerobot/visualize_dataset) by copy pasting your repo id given by:
+```bash
+echo ${HF_USER}/aloha_test
+```
+
+If you didn't upload with `--push-to-hub 0`, you can also visualize it locally with:
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/aloha_test
+```
+
+## Replay an episode
+
+**/!\ FOR SAFETY, READ THIS /!\**
+Replay consists in automatically replaying the sequence of actions (i.e. goal positions for your motors) recorded in a given dataset episode. Make sure the current initial position of your robot is similar to the one in your episode, so that your follower arms don't move too fast to go to the first goal positions. For safety, you might want to add `--robot-overrides max_relative_target=5` to your command line as explained above.
+
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py replay \
+    --robot-path lerobot/configs/robot/aloha.yaml \
+    --robot-overrides max_relative_target=null \
+    --fps 30 \
+    --repo-id ${HF_USER}/aloha_test \
+    --episode 0
+```
+
+## Train a policy
+
+To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
+```bash
+python lerobot/scripts/train.py \
+  dataset_repo_id=${HF_USER}/aloha_test \
+  policy=act_aloha_real \
+  env=aloha_real \
+  hydra.run.dir=outputs/train/act_aloha_test \
+  hydra.job.name=act_aloha_test \
+  device=cuda \
+  wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `dataset_repo_id=${HF_USER}/aloha_test`.
+2. We provided the policy with `policy=act_aloha_real`. This loads configurations from [`lerobot/configs/policy/act_aloha_real.yaml`](../lerobot/configs/policy/act_aloha_real.yaml). Importantly, this policy uses 4 cameras as input `cam_right_wrist`, `cam_left_wrist`, `cam_high`, and `cam_low`.
+3. We provided an environment as argument with `env=aloha_real`. This loads configurations from [`lerobot/configs/env/aloha_real.yaml`](../lerobot/configs/env/aloha_real.yaml). Note: this yaml defines 18 dimensions for the `state_dim` and `action_dim`, corresponding to 18 motors, not 14 motors as used in previous Aloha work. This is because, we include the `shoulder_shadow` and `elbow_shadow` motors for simplicity.
+4. We provided `device=cuda` since we are training on a Nvidia GPU.
+5. We provided `wandb.enable=true` to use [Weights and Biases](https://docs.wandb.ai/quickstart) for visualizing training plots. This is optional but if you use it, make sure you are logged in by running `wandb login`.
+
+Training should take several hours. You will find checkpoints in `outputs/train/act_aloha_test/checkpoints`.
+
+## Evaluate your policy
+
+You can use the `record` function from [`lerobot/scripts/control_robot.py`](../lerobot/scripts/control_robot.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
+```bash
+python lerobot/scripts/control_robot.py record \
+  --robot-path lerobot/configs/robot/aloha.yaml \
+  --robot-overrides max_relative_target=null \
+  --fps 30 \
+  --repo-id ${HF_USER}/eval_act_aloha_test \
+  --tags aloha tutorial eval \
+  --warmup-time-s 5 \
+  --episode-time-s 40 \
+  --reset-time-s 10 \
+  --num-episodes 10 \
+  --num-image-writer-processes 1 \
+  -p outputs/train/act_aloha_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `-p` argument which indicates the path to your policy checkpoint with  (e.g. `-p outputs/train/eval_aloha_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `-p ${HF_USER}/act_aloha_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `--repo-id ${HF_USER}/eval_act_aloha_test`).
+3. We use `--num-image-writer-processes 1` instead of the default value (`0`). On our computer, using a dedicated process to write images from the 4 cameras on disk allows to reach constent 30 fps during inference. Feel free to explore different values for `--num-image-writer-processes`.
+
+## More
+
+Follow this [previous tutorial](https://github.com/huggingface/lerobot/blob/main/examples/7_get_started_with_real_robot.md#4-train-a-policy-on-your-data) for a more in-depth explaination.
+
+If you have any question or need help, please reach out on Discord in the channel `#aloha-arm`.

examples/advanced/1_train_act_pusht/act_pusht.yaml

@@ -0,0 +1,87 @@
+# @package _global_
+
+# Change the seed to match what PushT eval uses
+# (to avoid evaluating on seeds used for generating the training data).
+seed: 100000
+# Change the dataset repository to the PushT one.
+dataset_repo_id: lerobot/pusht
+
+override_dataset_stats:
+  observation.image:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+training:
+  offline_steps: 80000
+  online_steps: 0
+  eval_freq: 10000
+  save_freq: 100000
+  log_freq: 250
+  save_model: true
+
+  batch_size: 8
+  lr: 1e-5
+  lr_backbone: 1e-5
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
+
+  delta_timestamps:
+    action: "[i / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    observation.image: [3, 96, 96]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.image: mean_std
+    # Use min_max normalization just because it's more standard.
+    observation.state: min_max
+  output_normalization_modes:
+    # Use min_max normalization just because it's more standard.
+    action: min_max
+
+  # Architecture.
+  # Vision backbone.
+  vision_backbone: resnet18
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
+  replace_final_stride_with_dilation: false
+  # Transformer layers.
+  pre_norm: false
+  dim_model: 512
+  n_heads: 8
+  dim_feedforward: 3200
+  feedforward_activation: relu
+  n_encoder_layers: 4
+    # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
+  # that means only the first layer is used. Here we match the original implementation by setting this to 1.
+  # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
+  n_decoder_layers: 1
+  # VAE.
+  use_vae: true
+  latent_dim: 32
+  n_vae_encoder_layers: 4
+
+  # Inference.
+  temporal_ensemble_coeff: null
+
+  # Training and loss computation.
+  dropout: 0.1
+  kl_weight: 10.0

examples/advanced/1_train_act_pusht/train_act_pusht.md

@@ -0,0 +1,70 @@
+In this tutorial we will learn how to adapt a policy configuration to be compatible with a new environment and dataset. As a concrete example, we will adapt the default configuration for ACT to be compatible with the PushT environment and dataset.
+
+If you haven't already read our tutorial on the [training script and configuration tooling](../4_train_policy_with_script.md) please do so prior to tackling this tutorial.
+
+Let's get started!
+
+Suppose we want to train ACT for PushT. Well, there are aspects of the ACT configuration that are specific to the ALOHA environments, and these happen to be incompatible with PushT. Therefore, trying to run the following will almost certainly raise an exception of sorts (eg: feature dimension mismatch):
+
+```bash
+python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
+```
+
+We need to adapt the parameters of the ACT policy configuration to the PushT environment. The most important ones are the image keys.
+
+ALOHA's datasets and environments typically use a variable number of cameras. In `lerobot/configs/policy/act.yaml` you may notice two relevant sections. Here we show you the minimal diff needed to adjust to PushT:
+
+```diff
+override_dataset_stats:
+-  observation.images.top:
++  observation.image:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+policy:
+  input_shapes:
+-    observation.images.top: [3, 480, 640]
++    observation.image: [3, 96, 96]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  input_normalization_modes:
+-    observation.images.top: mean_std
++    observation.image: mean_std
+     observation.state: min_max
+  output_normalization_modes:
+    action: min_max
+```
+
+Here we've accounted for the following:
+- PushT uses "observation.image" for its image key.
+- PushT provides smaller images.
+
+_Side note: technically we could override these via the CLI, but with many changes it gets a bit messy, and we also have a bit of a challenge in that we're using `.` in our observation keys which is treated by Hydra as a hierarchical separator_.
+
+For your convenience, we provide [`act_pusht.yaml`](./act_pusht.yaml) in this directory. It contains the diff above, plus some other (optional) ones that are explained within. Please copy it into `lerobot/configs/policy` with:
+
+```bash
+cp examples/advanced/1_train_act_pusht/act_pusht.yaml lerobot/configs/policy/act_pusht.yaml
+```
+
+(remember from a [previous tutorial](../4_train_policy_with_script.md) that Hydra will look in the `lerobot/configs` directory). Now try running the following.
+
+<!-- Note to contributor: are you changing this command? Note that it's tested in `Makefile`, so change it there too! -->
+```bash
+python lerobot/scripts/train.py policy=act_pusht env=pusht
+```
+
+Notice that this is much the same as the command that failed at the start of the tutorial, only:
+- Now we are using `policy=act_pusht` to point to our new configuration file.
+- We can drop `dataset_repo_id=lerobot/pusht` as the change is incorporated in our new configuration file.
+
+Hurrah! You're now training ACT for the PushT environment.
+
+---
+
+The bottom line of this tutorial is that when training policies for different environments and datasets you will need to understand what parts of the policy configuration are specific to those and make changes accordingly.
+
+Happy coding! 🤗

examples/advanced/2_calculate_validation_loss.py

@@ -0,0 +1,84 @@
+"""This script demonstrates how to slice a dataset and calculate the loss on a subset of the data.
+
+This technique can be useful for debugging and testing purposes, as well as identifying whether a policy
+is learning effectively.
+
+Furthermore, relying on validation loss to evaluate performance is generally not considered a good practice,
+especially in the context of imitation learning. The most reliable approach is to evaluate the policy directly
+on the target environment, whether that be in simulation or the real world.
+"""
+
+import math
+from pathlib import Path
+
+import torch
+from huggingface_hub import snapshot_download
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
+
+device = torch.device("cuda")
+
+# Download the diffusion policy for pusht environment
+pretrained_policy_path = Path(snapshot_download("lerobot/diffusion_pusht"))
+# OR uncomment the following to evaluate a policy from the local outputs/train folder.
+# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
+
+policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
+policy.eval()
+policy.to(device)
+
+# Set up the dataset.
+delta_timestamps = {
+    # Load the previous image and state at -0.1 seconds before current frame,
+    # then load current image and state corresponding to 0.0 second.
+    "observation.image": [-0.1, 0.0],
+    "observation.state": [-0.1, 0.0],
+    # Load the previous action (-0.1), the next action to be executed (0.0),
+    # and 14 future actions with a 0.1 seconds spacing. All these actions will be
+    # used to calculate the loss.
+    "action": [-0.1, 0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4],
+}
+
+# Load the last 10% of episodes of the dataset as a validation set.
+# - Load dataset metadata
+dataset_metadata = LeRobotDatasetMetadata("lerobot/pusht")
+# - Calculate train and val episodes
+total_episodes = dataset_metadata.total_episodes
+episodes = list(range(dataset_metadata.total_episodes))
+num_train_episodes = math.floor(total_episodes * 90 / 100)
+train_episodes = episodes[:num_train_episodes]
+val_episodes = episodes[num_train_episodes:]
+print(f"Number of episodes in full dataset: {total_episodes}")
+print(f"Number of episodes in training dataset (90% subset): {len(train_episodes)}")
+print(f"Number of episodes in validation dataset (10% subset): {len(val_episodes)}")
+# - Load train an val datasets
+train_dataset = LeRobotDataset("lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps)
+val_dataset = LeRobotDataset("lerobot/pusht", episodes=val_episodes, delta_timestamps=delta_timestamps)
+print(f"Number of frames in training dataset (90% subset): {len(train_dataset)}")
+print(f"Number of frames in validation dataset (10% subset): {len(val_dataset)}")
+
+# Create dataloader for evaluation.
+val_dataloader = torch.utils.data.DataLoader(
+    val_dataset,
+    num_workers=4,
+    batch_size=64,
+    shuffle=False,
+    pin_memory=device != torch.device("cpu"),
+    drop_last=False,
+)
+
+# Run validation loop.
+loss_cumsum = 0
+n_examples_evaluated = 0
+for batch in val_dataloader:
+    batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+    output_dict = policy.forward(batch)
+
+    loss_cumsum += output_dict["loss"].item()
+    n_examples_evaluated += batch["index"].shape[0]
+
+# Calculate the average loss over the validation set.
+average_loss = loss_cumsum / n_examples_evaluated
+
+print(f"Average loss on validation set: {average_loss:.4f}")

examples/port_datasets/pusht_zarr.py

@@ -0,0 +1,222 @@
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LEROBOT_HOME, LeRobotDataset
+from lerobot.common.datasets.push_dataset_to_hub._download_raw import download_raw
+
+PUSHT_TASK = "Push the T-shaped blue block onto the T-shaped green target surface."
+PUSHT_FEATURES = {
+    "observation.state": {
+        "dtype": "float32",
+        "shape": (2,),
+        "names": {
+            "axes": ["x", "y"],
+        },
+    },
+    "action": {
+        "dtype": "float32",
+        "shape": (2,),
+        "names": {
+            "axes": ["x", "y"],
+        },
+    },
+    "next.reward": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": None,
+    },
+    "next.success": {
+        "dtype": "bool",
+        "shape": (1,),
+        "names": None,
+    },
+    "observation.environment_state": {
+        "dtype": "float32",
+        "shape": (16,),
+        "names": [
+            "keypoints",
+        ],
+    },
+    "observation.image": {
+        "dtype": None,
+        "shape": (3, 96, 96),
+        "names": [
+            "channel",
+            "height",
+            "width",
+        ],
+    },
+}
+
+
+def build_features(mode: str) -> dict:
+    features = PUSHT_FEATURES
+    if mode == "keypoints":
+        features.pop("observation.image")
+    else:
+        features.pop("observation.environment_state")
+        features["observation.image"]["dtype"] = mode
+
+    return features
+
+
+def load_raw_dataset(zarr_path: Path):
+    try:
+        from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
+            ReplayBuffer as DiffusionPolicyReplayBuffer,
+        )
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+
+    zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
+    return zarr_data
+
+
+def calculate_coverage(zarr_data):
+    try:
+        import pymunk
+        from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+
+    block_pos = zarr_data["state"][:, 2:4]
+    block_angle = zarr_data["state"][:, 4]
+
+    num_frames = len(block_pos)
+
+    coverage = np.zeros((num_frames,))
+    # 8 keypoints with 2 coords each
+    keypoints = np.zeros((num_frames, 16))
+
+    # Set x, y, theta (in radians)
+    goal_pos_angle = np.array([256, 256, np.pi / 4])
+    goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
+
+    for i in range(num_frames):
+        space = pymunk.Space()
+        space.gravity = 0, 0
+        space.damping = 0
+
+        # Add walls.
+        walls = [
+            PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
+            PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
+            PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
+            PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
+        ]
+        space.add(*walls)
+
+        block_body, block_shapes = PushTEnv.add_tee(space, block_pos[i].tolist(), block_angle[i].item())
+        goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
+        block_geom = pymunk_to_shapely(block_body, block_body.shapes)
+        intersection_area = goal_geom.intersection(block_geom).area
+        goal_area = goal_geom.area
+        coverage[i] = intersection_area / goal_area
+        keypoints[i] = torch.from_numpy(PushTEnv.get_keypoints(block_shapes).flatten())
+
+    return coverage, keypoints
+
+
+def calculate_success(coverage: float, success_threshold: float):
+    return coverage > success_threshold
+
+
+def calculate_reward(coverage: float, success_threshold: float):
+    return np.clip(coverage / success_threshold, 0, 1)
+
+
+def main(raw_dir: Path, repo_id: str, mode: str = "video", push_to_hub: bool = True):
+    if mode not in ["video", "image", "keypoints"]:
+        raise ValueError(mode)
+
+    if (LEROBOT_HOME / repo_id).exists():
+        shutil.rmtree(LEROBOT_HOME / repo_id)
+
+    if not raw_dir.exists():
+        download_raw(raw_dir, repo_id="lerobot-raw/pusht_raw")
+
+    zarr_data = load_raw_dataset(zarr_path=raw_dir / "pusht_cchi_v7_replay.zarr")
+
+    env_state = zarr_data["state"][:]
+    agent_pos = env_state[:, :2]
+
+    action = zarr_data["action"][:]
+    image = zarr_data["img"]  # (b, h, w, c)
+
+    episode_data_index = {
+        "from": np.concatenate(([0], zarr_data.meta["episode_ends"][:-1])),
+        "to": zarr_data.meta["episode_ends"],
+    }
+
+    # Calculate success and reward based on the overlapping area
+    # of the T-object and the T-area.
+    coverage, keypoints = calculate_coverage(zarr_data)
+    success = calculate_success(coverage, success_threshold=0.95)
+    reward = calculate_reward(coverage, success_threshold=0.95)
+
+    features = build_features(mode)
+    dataset = LeRobotDataset.create(
+        repo_id=repo_id,
+        fps=10,
+        robot_type="2d pointer",
+        features=features,
+        image_writer_threads=4,
+    )
+    episodes = range(len(episode_data_index["from"]))
+    for ep_idx in episodes:
+        from_idx = episode_data_index["from"][ep_idx]
+        to_idx = episode_data_index["to"][ep_idx]
+        num_frames = to_idx - from_idx
+
+        for frame_idx in range(num_frames):
+            i = from_idx + frame_idx
+            frame = {
+                "action": torch.from_numpy(action[i]),
+                # Shift reward and success by +1 until the last item of the episode
+                "next.reward": reward[i + (frame_idx < num_frames - 1)],
+                "next.success": success[i + (frame_idx < num_frames - 1)],
+            }
+
+            frame["observation.state"] = torch.from_numpy(agent_pos[i])
+
+            if mode == "keypoints":
+                frame["observation.environment_state"] = torch.from_numpy(keypoints[i])
+            else:
+                frame["observation.image"] = torch.from_numpy(image[i])
+
+            dataset.add_frame(frame)
+
+        dataset.save_episode(task=PUSHT_TASK)
+
+    dataset.consolidate()
+
+    if push_to_hub:
+        dataset.push_to_hub()
+
+
+if __name__ == "__main__":
+    # To try this script, modify the repo id with your own HuggingFace user (e.g cadene/pusht)
+    repo_id = "lerobot/pusht"
+
+    modes = ["video", "image", "keypoints"]
+    # Uncomment if you want to try with a specific mode
+    # modes = ["video"]
+    # modes = ["image"]
+    # modes = ["keypoints"]
+
+    raw_dir = Path("data/lerobot-raw/pusht_raw")
+    for mode in modes:
+        if mode in ["image", "keypoints"]:
+            repo_id += f"_{mode}"
+
+        # download and load raw dataset, create LeRobotDataset, populate it, push to hub
+        main(raw_dir, repo_id=repo_id, mode=mode)
+
+        # Uncomment if you want to load the local dataset and explore it
+        # dataset = LeRobotDataset(repo_id=repo_id, local_files_only=True)
+        # breakpoint()

lerobot/__init__.py

@@ -1,3 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 """
 This file contains lists of available environments, dataset and policies to reflect the current state of LeRobot library.
 We do not want to import all the dependencies, but instead we keep it lightweight to ensure fast access to these variables.
@@ -9,8 +24,12 @@ Example:
         print(lerobot.available_tasks_per_env)
         print(lerobot.available_datasets)
         print(lerobot.available_datasets_per_env)
+        print(lerobot.available_real_world_datasets)
         print(lerobot.available_policies)
         print(lerobot.available_policies_per_env)
+        print(lerobot.available_robots)
+        print(lerobot.available_cameras)
+        print(lerobot.available_motors)
     ```
 
 When implementing a new dataset loadable with LeRobotDataset follow these steps:
@@ -29,6 +48,9 @@ import itertools
 
 from lerobot.__version__ import __version__  # noqa: F401
 
+# TODO(rcadene): Improve policies and envs. As of now, an item in `available_policies`
+# refers to a yaml file AND a modeling name. Same for `available_envs` which refers to
+# a yaml file AND a environment name. The difference should be more obvious.
 available_tasks_per_env = {
     "aloha": [
         "AlohaInsertion-v0",
@@ -36,6 +58,7 @@ available_tasks_per_env = {
     ],
     "pusht": ["PushT-v0"],
     "xarm": ["XarmLift-v0"],
+    "dora_aloha_real": ["DoraAloha-v0", "DoraKoch-v0", "DoraReachy2-v0"],
 }
 available_envs = list(available_tasks_per_env.keys())
 
@@ -45,39 +68,160 @@ available_datasets_per_env = {
         "lerobot/aloha_sim_insertion_scripted",
         "lerobot/aloha_sim_transfer_cube_human",
         "lerobot/aloha_sim_transfer_cube_scripted",
+        "lerobot/aloha_sim_insertion_human_image",
+        "lerobot/aloha_sim_insertion_scripted_image",
+        "lerobot/aloha_sim_transfer_cube_human_image",
+        "lerobot/aloha_sim_transfer_cube_scripted_image",
     ],
-    "pusht": ["lerobot/pusht"],
+    # TODO(alexander-soare): Add "lerobot/pusht_keypoints". Right now we can't because this is too tightly
+    # coupled with tests.
+    "pusht": ["lerobot/pusht", "lerobot/pusht_image"],
     "xarm": [
         "lerobot/xarm_lift_medium",
         "lerobot/xarm_lift_medium_replay",
         "lerobot/xarm_push_medium",
         "lerobot/xarm_push_medium_replay",
+        "lerobot/xarm_lift_medium_image",
+        "lerobot/xarm_lift_medium_replay_image",
+        "lerobot/xarm_push_medium_image",
+        "lerobot/xarm_push_medium_replay_image",
+    ],
+    "dora_aloha_real": [
+        "lerobot/aloha_static_battery",
+        "lerobot/aloha_static_candy",
+        "lerobot/aloha_static_coffee",
+        "lerobot/aloha_static_coffee_new",
+        "lerobot/aloha_static_cups_open",
+        "lerobot/aloha_static_fork_pick_up",
+        "lerobot/aloha_static_pingpong_test",
+        "lerobot/aloha_static_pro_pencil",
+        "lerobot/aloha_static_screw_driver",
+        "lerobot/aloha_static_tape",
+        "lerobot/aloha_static_thread_velcro",
+        "lerobot/aloha_static_towel",
+        "lerobot/aloha_static_vinh_cup",
+        "lerobot/aloha_static_vinh_cup_left",
+        "lerobot/aloha_static_ziploc_slide",
     ],
 }
 
-available_datasets_without_env = ["lerobot/umi_cup_in_the_wild"]
+available_real_world_datasets = [
+    "lerobot/aloha_mobile_cabinet",
+    "lerobot/aloha_mobile_chair",
+    "lerobot/aloha_mobile_elevator",
+    "lerobot/aloha_mobile_shrimp",
+    "lerobot/aloha_mobile_wash_pan",
+    "lerobot/aloha_mobile_wipe_wine",
+    "lerobot/aloha_static_battery",
+    "lerobot/aloha_static_candy",
+    "lerobot/aloha_static_coffee",
+    "lerobot/aloha_static_coffee_new",
+    "lerobot/aloha_static_cups_open",
+    "lerobot/aloha_static_fork_pick_up",
+    "lerobot/aloha_static_pingpong_test",
+    "lerobot/aloha_static_pro_pencil",
+    "lerobot/aloha_static_screw_driver",
+    "lerobot/aloha_static_tape",
+    "lerobot/aloha_static_thread_velcro",
+    "lerobot/aloha_static_towel",
+    "lerobot/aloha_static_vinh_cup",
+    "lerobot/aloha_static_vinh_cup_left",
+    "lerobot/aloha_static_ziploc_slide",
+    "lerobot/umi_cup_in_the_wild",
+    "lerobot/unitreeh1_fold_clothes",
+    "lerobot/unitreeh1_rearrange_objects",
+    "lerobot/unitreeh1_two_robot_greeting",
+    "lerobot/unitreeh1_warehouse",
+    "lerobot/nyu_rot_dataset",
+    "lerobot/utokyo_saytap",
+    "lerobot/imperialcollege_sawyer_wrist_cam",
+    "lerobot/utokyo_xarm_bimanual",
+    "lerobot/tokyo_u_lsmo",
+    "lerobot/utokyo_pr2_opening_fridge",
+    "lerobot/cmu_franka_exploration_dataset",
+    "lerobot/cmu_stretch",
+    "lerobot/asu_table_top",
+    "lerobot/utokyo_pr2_tabletop_manipulation",
+    "lerobot/utokyo_xarm_pick_and_place",
+    "lerobot/ucsd_kitchen_dataset",
+    "lerobot/austin_buds_dataset",
+    "lerobot/dlr_sara_grid_clamp",
+    "lerobot/conq_hose_manipulation",
+    "lerobot/columbia_cairlab_pusht_real",
+    "lerobot/dlr_sara_pour",
+    "lerobot/dlr_edan_shared_control",
+    "lerobot/ucsd_pick_and_place_dataset",
+    "lerobot/berkeley_cable_routing",
+    "lerobot/nyu_franka_play_dataset",
+    "lerobot/austin_sirius_dataset",
+    "lerobot/cmu_play_fusion",
+    "lerobot/berkeley_gnm_sac_son",
+    "lerobot/nyu_door_opening_surprising_effectiveness",
+    "lerobot/berkeley_fanuc_manipulation",
+    "lerobot/jaco_play",
+    "lerobot/viola",
+    "lerobot/kaist_nonprehensile",
+    "lerobot/berkeley_mvp",
+    "lerobot/uiuc_d3field",
+    "lerobot/berkeley_gnm_recon",
+    "lerobot/austin_sailor_dataset",
+    "lerobot/utaustin_mutex",
+    "lerobot/roboturk",
+    "lerobot/stanford_hydra_dataset",
+    "lerobot/berkeley_autolab_ur5",
+    "lerobot/stanford_robocook",
+    "lerobot/toto",
+    "lerobot/fmb",
+    "lerobot/droid_100",
+    "lerobot/berkeley_rpt",
+    "lerobot/stanford_kuka_multimodal_dataset",
+    "lerobot/iamlab_cmu_pickup_insert",
+    "lerobot/taco_play",
+    "lerobot/berkeley_gnm_cory_hall",
+    "lerobot/usc_cloth_sim",
+]
 
-available_datasets = list(
-    itertools.chain(*available_datasets_per_env.values(), available_datasets_without_env)
-)
-
-# TODO(rcadene, aliberts, alexander-soare): Add real-world env with a gym API
-available_datasets_without_env = ["lerobot/umi_cup_in_the_wild"]
-
-available_datasets = list(
-    itertools.chain(*available_datasets_per_env.values(), available_datasets_without_env)
+available_datasets = sorted(
+    set(itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets))
 )
 
+# lists all available policies from `lerobot/common/policies`
 available_policies = [
     "act",
     "diffusion",
     "tdmpc",
+    "vqbet",
 ]
 
+# lists all available robots from `lerobot/common/robot_devices/robots`
+available_robots = [
+    "koch",
+    "koch_bimanual",
+    "aloha",
+    "so100",
+    "moss",
+]
+
+# lists all available cameras from `lerobot/common/robot_devices/cameras`
+available_cameras = [
+    "opencv",
+    "intelrealsense",
+]
+
+# lists all available motors from `lerobot/common/robot_devices/motors`
+available_motors = [
+    "dynamixel",
+    "feetech",
+]
+
+# keys and values refer to yaml files
 available_policies_per_env = {
     "aloha": ["act"],
-    "pusht": ["diffusion"],
+    "pusht": ["diffusion", "vqbet"],
     "xarm": ["tdmpc"],
+    "koch_real": ["act_koch_real"],
+    "aloha_real": ["act_aloha_real"],
+    "dora_aloha_real": ["act_aloha_real"],
 }
 
 env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]

lerobot/__version__.py

@@ -1,3 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 """To enable `lerobot.__version__`"""
 
 from importlib.metadata import PackageNotFoundError, version

lerobot/common/datasets/card_template.md

@@ -0,0 +1,27 @@
+---
+# For reference on dataset card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/datasetcard.md?plain=1
+# Doc / guide: https://huggingface.co/docs/hub/datasets-cards
+{{ card_data }}
+---
+
+This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
+
+## Dataset Description
+
+{{ dataset_description | default("", true) }}
+
+- **Homepage:** {{ url | default("[More Information Needed]", true)}}
+- **Paper:** {{ paper | default("[More Information Needed]", true)}}
+- **License:** {{ license | default("[More Information Needed]", true)}}
+
+## Dataset Structure
+
+{{ dataset_structure | default("[More Information Needed]", true)}}
+
+## Citation
+
+**BibTeX:**
+
+```bibtex
+{{ citation_bibtex | default("[More Information Needed]", true)}}
+```

lerobot/common/datasets/compute_stats.py

@@ -0,0 +1,214 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from copy import deepcopy
+from math import ceil
+
+import einops
+import torch
+import tqdm
+
+
+def get_stats_einops_patterns(dataset, num_workers=0):
+    """These einops patterns will be used to aggregate batches and compute statistics.
+
+    Note: We assume the images are in channel first format
+    """
+
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=num_workers,
+        batch_size=2,
+        shuffle=False,
+    )
+    batch = next(iter(dataloader))
+
+    stats_patterns = {}
+
+    for key in dataset.features:
+        # sanity check that tensors are not float64
+        assert batch[key].dtype != torch.float64
+
+        # if isinstance(feats_type, (VideoFrame, Image)):
+        if key in dataset.meta.camera_keys:
+            # sanity check that images are channel first
+            _, c, h, w = batch[key].shape
+            assert c < h and c < w, f"expect channel first images, but instead {batch[key].shape}"
+
+            # sanity check that images are float32 in range [0,1]
+            assert batch[key].dtype == torch.float32, f"expect torch.float32, but instead {batch[key].dtype=}"
+            assert batch[key].max() <= 1, f"expect pixels lower than 1, but instead {batch[key].max()=}"
+            assert batch[key].min() >= 0, f"expect pixels greater than 1, but instead {batch[key].min()=}"
+
+            stats_patterns[key] = "b c h w -> c 1 1"
+        elif batch[key].ndim == 2:
+            stats_patterns[key] = "b c -> c "
+        elif batch[key].ndim == 1:
+            stats_patterns[key] = "b -> 1"
+        else:
+            raise ValueError(f"{key}, {batch[key].shape}")
+
+    return stats_patterns
+
+
+def compute_stats(dataset, batch_size=8, num_workers=8, max_num_samples=None):
+    """Compute mean/std and min/max statistics of all data keys in a LeRobotDataset."""
+    if max_num_samples is None:
+        max_num_samples = len(dataset)
+
+    # for more info on why we need to set the same number of workers, see `load_from_videos`
+    stats_patterns = get_stats_einops_patterns(dataset, num_workers)
+
+    # mean and std will be computed incrementally while max and min will track the running value.
+    mean, std, max, min = {}, {}, {}, {}
+    for key in stats_patterns:
+        mean[key] = torch.tensor(0.0).float()
+        std[key] = torch.tensor(0.0).float()
+        max[key] = torch.tensor(-float("inf")).float()
+        min[key] = torch.tensor(float("inf")).float()
+
+    def create_seeded_dataloader(dataset, batch_size, seed):
+        generator = torch.Generator()
+        generator.manual_seed(seed)
+        dataloader = torch.utils.data.DataLoader(
+            dataset,
+            num_workers=num_workers,
+            batch_size=batch_size,
+            shuffle=True,
+            drop_last=False,
+            generator=generator,
+        )
+        return dataloader
+
+    # Note: Due to be refactored soon. The point of storing `first_batch` is to make sure we don't get
+    # surprises when rerunning the sampler.
+    first_batch = None
+    running_item_count = 0  # for online mean computation
+    dataloader = create_seeded_dataloader(dataset, batch_size, seed=1337)
+    for i, batch in enumerate(
+        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute mean, min, max")
+    ):
+        this_batch_size = len(batch["index"])
+        running_item_count += this_batch_size
+        if first_batch is None:
+            first_batch = deepcopy(batch)
+        for key, pattern in stats_patterns.items():
+            batch[key] = batch[key].float()
+            # Numerically stable update step for mean computation.
+            batch_mean = einops.reduce(batch[key], pattern, "mean")
+            # Hint: to update the mean we need x̄ₙ = (Nₙ₋₁x̄ₙ₋₁ + Bₙxₙ) / Nₙ, where the subscript represents
+            # the update step, N is the running item count, B is this batch size, x̄ is the running mean,
+            # and x is the current batch mean. Some rearrangement is then required to avoid risking
+            # numerical overflow. Another hint: Nₙ₋₁ = Nₙ - Bₙ. Rearrangement yields
+            # x̄ₙ = x̄ₙ₋₁ + Bₙ * (xₙ - x̄ₙ₋₁) / Nₙ
+            mean[key] = mean[key] + this_batch_size * (batch_mean - mean[key]) / running_item_count
+            max[key] = torch.maximum(max[key], einops.reduce(batch[key], pattern, "max"))
+            min[key] = torch.minimum(min[key], einops.reduce(batch[key], pattern, "min"))
+
+        if i == ceil(max_num_samples / batch_size) - 1:
+            break
+
+    first_batch_ = None
+    running_item_count = 0  # for online std computation
+    dataloader = create_seeded_dataloader(dataset, batch_size, seed=1337)
+    for i, batch in enumerate(
+        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute std")
+    ):
+        this_batch_size = len(batch["index"])
+        running_item_count += this_batch_size
+        # Sanity check to make sure the batches are still in the same order as before.
+        if first_batch_ is None:
+            first_batch_ = deepcopy(batch)
+            for key in stats_patterns:
+                assert torch.equal(first_batch_[key], first_batch[key])
+        for key, pattern in stats_patterns.items():
+            batch[key] = batch[key].float()
+            # Numerically stable update step for mean computation (where the mean is over squared
+            # residuals).See notes in the mean computation loop above.
+            batch_std = einops.reduce((batch[key] - mean[key]) ** 2, pattern, "mean")
+            std[key] = std[key] + this_batch_size * (batch_std - std[key]) / running_item_count
+
+        if i == ceil(max_num_samples / batch_size) - 1:
+            break
+
+    for key in stats_patterns:
+        std[key] = torch.sqrt(std[key])
+
+    stats = {}
+    for key in stats_patterns:
+        stats[key] = {
+            "mean": mean[key],
+            "std": std[key],
+            "max": max[key],
+            "min": min[key],
+        }
+    return stats
+
+
+def aggregate_stats(ls_datasets) -> dict[str, torch.Tensor]:
+    """Aggregate stats of multiple LeRobot datasets into one set of stats without recomputing from scratch.
+
+    The final stats will have the union of all data keys from each of the datasets.
+
+    The final stats will have the union of all data keys from each of the datasets. For instance:
+    - new_max = max(max_dataset_0, max_dataset_1, ...)
+    - new_min = min(min_dataset_0, min_dataset_1, ...)
+    - new_mean = (mean of all data)
+    - new_std = (std of all data)
+    """
+    data_keys = set()
+    for dataset in ls_datasets:
+        data_keys.update(dataset.meta.stats.keys())
+    stats = {k: {} for k in data_keys}
+    for data_key in data_keys:
+        for stat_key in ["min", "max"]:
+            # compute `max(dataset_0["max"], dataset_1["max"], ...)`
+            stats[data_key][stat_key] = einops.reduce(
+                torch.stack(
+                    [ds.meta.stats[data_key][stat_key] for ds in ls_datasets if data_key in ds.meta.stats],
+                    dim=0,
+                ),
+                "n ... -> ...",
+                stat_key,
+            )
+        total_samples = sum(d.num_frames for d in ls_datasets if data_key in d.meta.stats)
+        # Compute the "sum" statistic by multiplying each mean by the number of samples in the respective
+        # dataset, then divide by total_samples to get the overall "mean".
+        # NOTE: the brackets around (d.num_frames / total_samples) are needed tor minimize the risk of
+        # numerical overflow!
+        stats[data_key]["mean"] = sum(
+            d.meta.stats[data_key]["mean"] * (d.num_frames / total_samples)
+            for d in ls_datasets
+            if data_key in d.meta.stats
+        )
+        # The derivation for standard deviation is a little more involved but is much in the same spirit as
+        # the computation of the mean.
+        # Given two sets of data where the statistics are known:
+        # σ_combined = sqrt[ (n1 * (σ1^2 + d1^2) + n2 * (σ2^2 + d2^2)) / (n1 + n2) ]
+        # where d1 = μ1 - μ_combined, d2 = μ2 - μ_combined
+        # NOTE: the brackets around (d.num_frames / total_samples) are needed tor minimize the risk of
+        # numerical overflow!
+        stats[data_key]["std"] = torch.sqrt(
+            sum(
+                (
+                    d.meta.stats[data_key]["std"] ** 2
+                    + (d.meta.stats[data_key]["mean"] - stats[data_key]["mean"]) ** 2
+                )
+                * (d.num_frames / total_samples)
+                for d in ls_datasets
+                if data_key in d.meta.stats
+            )
+        )
+    return stats

lerobot/common/datasets/factory.py

@@ -1,44 +1,116 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 import logging
-import os
-from pathlib import Path
 
 import torch
-from omegaconf import OmegaConf
+from omegaconf import ListConfig, OmegaConf
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-
-DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
+from lerobot.common.datasets.transforms import get_image_transforms
 
 
-def make_dataset(
-    cfg,
-    split="train",
-):
-    if cfg.env.name not in cfg.dataset.repo_id:
-        logging.warning(
-            f"There might be a mismatch between your training dataset ({cfg.dataset.repo_id=}) and your environment ({cfg.env.name=})."
-        )
+def resolve_delta_timestamps(cfg):
+    """Resolves delta_timestamps config key (in-place) by using `eval`.
 
-    delta_timestamps = cfg.policy.get("delta_timestamps")
+    Doesn't do anything if delta_timestamps is not specified or has already been resolve (as evidenced by
+    the data type of its values).
+    """
+    delta_timestamps = cfg.training.get("delta_timestamps")
     if delta_timestamps is not None:
         for key in delta_timestamps:
             if isinstance(delta_timestamps[key], str):
-                delta_timestamps[key] = eval(delta_timestamps[key])
+                # TODO(rcadene, alexander-soare): remove `eval` to avoid exploit
+                cfg.training.delta_timestamps[key] = eval(delta_timestamps[key])
 
-    # TODO(rcadene): add data augmentations
 
-    dataset = LeRobotDataset(
-        cfg.dataset.repo_id,
-        split=split,
-        root=DATA_DIR,
-        delta_timestamps=delta_timestamps,
-    )
+def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotDataset:
+    """
+    Args:
+        cfg: A Hydra config as per the LeRobot config scheme.
+        split: Select the data subset used to create an instance of LeRobotDataset.
+            All datasets hosted on [lerobot](https://huggingface.co/lerobot) contain only one subset: "train".
+            Thus, by default, `split="train"` selects all the available data. `split` aims to work like the
+            slicer in the hugging face datasets:
+            https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
+            As of now, it only supports `split="train[:n]"` to load the first n frames of the dataset or
+            `split="train[n:]"` to load the last n frames. For instance `split="train[:1000]"`.
+    Returns:
+        The LeRobotDataset.
+    """
+    if not isinstance(cfg.dataset_repo_id, (str, ListConfig)):
+        raise ValueError(
+            "Expected cfg.dataset_repo_id to be either a single string to load one dataset or a list of "
+            "strings to load multiple datasets."
+        )
+
+    # A soft check to warn if the environment matches the dataset. Don't check if we are using a real world env (dora).
+    if cfg.env.name != "dora":
+        if isinstance(cfg.dataset_repo_id, str):
+            dataset_repo_ids = [cfg.dataset_repo_id]  # single dataset
+        else:
+            dataset_repo_ids = cfg.dataset_repo_id  # multiple datasets
+
+        for dataset_repo_id in dataset_repo_ids:
+            if cfg.env.name not in dataset_repo_id:
+                logging.warning(
+                    f"There might be a mismatch between your training dataset ({dataset_repo_id=}) and your "
+                    f"environment ({cfg.env.name=})."
+                )
+
+    resolve_delta_timestamps(cfg)
+
+    image_transforms = None
+    if cfg.training.image_transforms.enable:
+        cfg_tf = cfg.training.image_transforms
+        image_transforms = get_image_transforms(
+            brightness_weight=cfg_tf.brightness.weight,
+            brightness_min_max=cfg_tf.brightness.min_max,
+            contrast_weight=cfg_tf.contrast.weight,
+            contrast_min_max=cfg_tf.contrast.min_max,
+            saturation_weight=cfg_tf.saturation.weight,
+            saturation_min_max=cfg_tf.saturation.min_max,
+            hue_weight=cfg_tf.hue.weight,
+            hue_min_max=cfg_tf.hue.min_max,
+            sharpness_weight=cfg_tf.sharpness.weight,
+            sharpness_min_max=cfg_tf.sharpness.min_max,
+            max_num_transforms=cfg_tf.max_num_transforms,
+            random_order=cfg_tf.random_order,
+        )
+
+    if isinstance(cfg.dataset_repo_id, str):
+        # TODO (aliberts): add 'episodes' arg from config after removing hydra
+        dataset = LeRobotDataset(
+            cfg.dataset_repo_id,
+            delta_timestamps=cfg.training.get("delta_timestamps"),
+            image_transforms=image_transforms,
+            video_backend=cfg.video_backend,
+        )
+    else:
+        dataset = MultiLeRobotDataset(
+            cfg.dataset_repo_id,
+            delta_timestamps=cfg.training.get("delta_timestamps"),
+            image_transforms=image_transforms,
+            video_backend=cfg.video_backend,
+        )
 
     if cfg.get("override_dataset_stats"):
         for key, stats_dict in cfg.override_dataset_stats.items():
             for stats_type, listconfig in stats_dict.items():
                 # example of stats_type: min, max, mean, std
                 stats = OmegaConf.to_container(listconfig, resolve=True)
-                dataset.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
+                dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
     return dataset

lerobot/common/datasets/image_writer.py

@@ -0,0 +1,160 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import multiprocessing
+import queue
+import threading
+from pathlib import Path
+
+import numpy as np
+import PIL.Image
+import torch
+
+
+def safe_stop_image_writer(func):
+    def wrapper(*args, **kwargs):
+        try:
+            return func(*args, **kwargs)
+        except Exception as e:
+            dataset = kwargs.get("dataset")
+            image_writer = getattr(dataset, "image_writer", None) if dataset else None
+            if image_writer is not None:
+                print("Waiting for image writer to terminate...")
+                image_writer.stop()
+            raise e
+
+    return wrapper
+
+
+def image_array_to_image(image_array: np.ndarray) -> PIL.Image.Image:
+    # TODO(aliberts): handle 1 channel and 4 for depth images
+    if image_array.ndim == 3 and image_array.shape[0] in [1, 3]:
+        # Transpose from pytorch convention (C, H, W) to (H, W, C)
+        image_array = image_array.transpose(1, 2, 0)
+    if image_array.dtype != np.uint8:
+        # Assume the image is in [0, 1] range for floating-point data
+        image_array = np.clip(image_array, 0, 1)
+        image_array = (image_array * 255).astype(np.uint8)
+    return PIL.Image.fromarray(image_array)
+
+
+def write_image(image: np.ndarray | PIL.Image.Image, fpath: Path):
+    try:
+        if isinstance(image, np.ndarray):
+            img = image_array_to_image(image)
+        elif isinstance(image, PIL.Image.Image):
+            img = image
+        else:
+            raise TypeError(f"Unsupported image type: {type(image)}")
+        img.save(fpath)
+    except Exception as e:
+        print(f"Error writing image {fpath}: {e}")
+
+
+def worker_thread_loop(queue: queue.Queue):
+    while True:
+        item = queue.get()
+        if item is None:
+            queue.task_done()
+            break
+        image_array, fpath = item
+        write_image(image_array, fpath)
+        queue.task_done()
+
+
+def worker_process(queue: queue.Queue, num_threads: int):
+    threads = []
+    for _ in range(num_threads):
+        t = threading.Thread(target=worker_thread_loop, args=(queue,))
+        t.daemon = True
+        t.start()
+        threads.append(t)
+    for t in threads:
+        t.join()
+
+
+class AsyncImageWriter:
+    """
+    This class abstract away the initialisation of processes or/and threads to
+    save images on disk asynchrounously, which is critical to control a robot and record data
+    at a high frame rate.
+
+    When `num_processes=0`, it creates a threads pool of size `num_threads`.
+    When `num_processes>0`, it creates processes pool of size `num_processes`, where each subprocess starts
+    their own threads pool of size `num_threads`.
+
+    The optimal number of processes and threads depends on your computer capabilities.
+    We advise to use 4 threads per camera with 0 processes. If the fps is not stable, try to increase or lower
+    the number of threads. If it is still not stable, try to use 1 subprocess, or more.
+    """
+
+    def __init__(self, num_processes: int = 0, num_threads: int = 1):
+        self.num_processes = num_processes
+        self.num_threads = num_threads
+        self.queue = None
+        self.threads = []
+        self.processes = []
+        self._stopped = False
+
+        if num_threads <= 0 and num_processes <= 0:
+            raise ValueError("Number of threads and processes must be greater than zero.")
+
+        if self.num_processes == 0:
+            # Use threading
+            self.queue = queue.Queue()
+            for _ in range(self.num_threads):
+                t = threading.Thread(target=worker_thread_loop, args=(self.queue,))
+                t.daemon = True
+                t.start()
+                self.threads.append(t)
+        else:
+            # Use multiprocessing
+            self.queue = multiprocessing.JoinableQueue()
+            for _ in range(self.num_processes):
+                p = multiprocessing.Process(target=worker_process, args=(self.queue, self.num_threads))
+                p.daemon = True
+                p.start()
+                self.processes.append(p)
+
+    def save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path):
+        if isinstance(image, torch.Tensor):
+            # Convert tensor to numpy array to minimize main process time
+            image = image.cpu().numpy()
+        self.queue.put((image, fpath))
+
+    def wait_until_done(self):
+        self.queue.join()
+
+    def stop(self):
+        if self._stopped:
+            return
+
+        if self.num_processes == 0:
+            for _ in self.threads:
+                self.queue.put(None)
+            for t in self.threads:
+                t.join()
+        else:
+            num_nones = self.num_processes * self.num_threads
+            for _ in range(num_nones):
+                self.queue.put(None)
+            for p in self.processes:
+                p.join()
+                if p.is_alive():
+                    p.terminate()
+            self.queue.close()
+            self.queue.join_thread()
+
+        self._stopped = True

lerobot/common/datasets/online_buffer.py

@@ -0,0 +1,384 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""An online buffer for the online training loop in train.py
+
+Note to maintainers: This duplicates some logic from LeRobotDataset and EpisodeAwareSampler. We should
+consider converging to one approach. Here we have opted to use numpy.memmap to back the data buffer. It's much
+faster than using HuggingFace Datasets as there's no conversion to an intermediate non-python object. Also it
+supports in-place slicing and mutation which is very handy for a dynamic buffer.
+"""
+
+import os
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+
+def _make_memmap_safe(**kwargs) -> np.memmap:
+    """Make a numpy memmap with checks on available disk space first.
+
+    Expected kwargs are: "filename", "dtype" (must by np.dtype), "mode" and "shape"
+
+    For information on dtypes:
+    https://numpy.org/doc/stable/reference/arrays.dtypes.html#arrays-dtypes-constructing
+    """
+    if kwargs["mode"].startswith("w"):
+        required_space = kwargs["dtype"].itemsize * np.prod(kwargs["shape"])  # bytes
+        stats = os.statvfs(Path(kwargs["filename"]).parent)
+        available_space = stats.f_bavail * stats.f_frsize  # bytes
+        if required_space >= available_space * 0.8:
+            raise RuntimeError(
+                f"You're about to take up {required_space} of {available_space} bytes available."
+            )
+    return np.memmap(**kwargs)
+
+
+class OnlineBuffer(torch.utils.data.Dataset):
+    """FIFO data buffer for the online training loop in train.py.
+
+    Follows the protocol of LeRobotDataset as much as is required to have it be used by the online training
+    loop in the same way that a LeRobotDataset would be used.
+
+    The underlying data structure will have data inserted in a circular fashion. Always insert after the
+    last index, and when you reach the end, wrap around to the start.
+
+    The data is stored in a numpy memmap.
+    """
+
+    NEXT_INDEX_KEY = "_next_index"
+    OCCUPANCY_MASK_KEY = "_occupancy_mask"
+    INDEX_KEY = "index"
+    FRAME_INDEX_KEY = "frame_index"
+    EPISODE_INDEX_KEY = "episode_index"
+    TIMESTAMP_KEY = "timestamp"
+    IS_PAD_POSTFIX = "_is_pad"
+
+    def __init__(
+        self,
+        write_dir: str | Path,
+        data_spec: dict[str, Any] | None,
+        buffer_capacity: int | None,
+        fps: float | None = None,
+        delta_timestamps: dict[str, list[float]] | dict[str, np.ndarray] | None = None,
+    ):
+        """
+        The online buffer can be provided from scratch or you can load an existing online buffer by passing
+        a `write_dir` associated with an existing buffer.
+
+        Args:
+            write_dir: Where to keep the numpy memmap files. One memmap file will be stored for each data key.
+                Note that if the files already exist, they are opened in read-write mode (used for training
+                resumption.)
+            data_spec: A mapping from data key to data specification, like {data_key: {"shape": tuple[int],
+                "dtype": np.dtype}}. This should include all the data that you wish to record into the buffer,
+                but note that "index", "frame_index" and "episode_index" are already accounted for by this
+                class, so you don't need to include them.
+            buffer_capacity: How many frames should be stored in the buffer as a maximum. Be aware of your
+                system's available disk space when choosing this.
+            fps: Same as the fps concept in LeRobot dataset. Here it needs to be provided for the
+                 delta_timestamps logic. You can pass None if you are not using delta_timestamps.
+            delta_timestamps: Same as the delta_timestamps concept in LeRobotDataset. This is internally
+                converted to dict[str, np.ndarray] for optimization purposes.
+
+        """
+        self.set_delta_timestamps(delta_timestamps)
+        self._fps = fps
+        # Tolerance in seconds used to discard loaded frames when their timestamps are not close enough from
+        # the requested frames. It is only used when `delta_timestamps` is provided.
+        # minus 1e-4 to account for possible numerical error
+        self.tolerance_s = 1 / self.fps - 1e-4 if fps is not None else None
+        self._buffer_capacity = buffer_capacity
+        data_spec = self._make_data_spec(data_spec, buffer_capacity)
+        Path(write_dir).mkdir(parents=True, exist_ok=True)
+        self._data = {}
+        for k, v in data_spec.items():
+            self._data[k] = _make_memmap_safe(
+                filename=Path(write_dir) / k,
+                dtype=v["dtype"] if v is not None else None,
+                mode="r+" if (Path(write_dir) / k).exists() else "w+",
+                shape=tuple(v["shape"]) if v is not None else None,
+            )
+
+    @property
+    def delta_timestamps(self) -> dict[str, np.ndarray] | None:
+        return self._delta_timestamps
+
+    def set_delta_timestamps(self, value: dict[str, list[float]] | None):
+        """Set delta_timestamps converting the values to numpy arrays.
+
+        The conversion is for an optimization in the __getitem__. The loop is much slower if the arrays
+        need to be converted into numpy arrays.
+        """
+        if value is not None:
+            self._delta_timestamps = {k: np.array(v) for k, v in value.items()}
+        else:
+            self._delta_timestamps = None
+
+    def _make_data_spec(self, data_spec: dict[str, Any], buffer_capacity: int) -> dict[str, dict[str, Any]]:
+        """Makes the data spec for np.memmap."""
+        if any(k.startswith("_") for k in data_spec):
+            raise ValueError(
+                "data_spec keys should not start with '_'. This prefix is reserved for internal logic."
+            )
+        preset_keys = {
+            OnlineBuffer.INDEX_KEY,
+            OnlineBuffer.FRAME_INDEX_KEY,
+            OnlineBuffer.EPISODE_INDEX_KEY,
+            OnlineBuffer.TIMESTAMP_KEY,
+        }
+        if len(intersection := set(data_spec).intersection(preset_keys)) > 0:
+            raise ValueError(
+                f"data_spec should not contain any of {preset_keys} as these are handled internally. "
+                f"The provided data_spec has {intersection}."
+            )
+        complete_data_spec = {
+            # _next_index will be a pointer to the next index that we should start filling from when we add
+            # more data.
+            OnlineBuffer.NEXT_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": ()},
+            # Since the memmap is initialized with all-zeros, this keeps track of which indices are occupied
+            # with real data rather than the dummy initialization.
+            OnlineBuffer.OCCUPANCY_MASK_KEY: {"dtype": np.dtype("?"), "shape": (buffer_capacity,)},
+            OnlineBuffer.INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.FRAME_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.EPISODE_INDEX_KEY: {"dtype": np.dtype("int64"), "shape": (buffer_capacity,)},
+            OnlineBuffer.TIMESTAMP_KEY: {"dtype": np.dtype("float64"), "shape": (buffer_capacity,)},
+        }
+        for k, v in data_spec.items():
+            complete_data_spec[k] = {"dtype": v["dtype"], "shape": (buffer_capacity, *v["shape"])}
+        return complete_data_spec
+
+    def add_data(self, data: dict[str, np.ndarray]):
+        """Add new data to the buffer, which could potentially mean shifting old data out.
+
+        The new data should contain all the frames (in order) of any number of episodes. The indices should
+        start from 0 (note to the developer: this can easily be generalized). See the `rollout` and
+        `eval_policy` functions in `eval.py` for more information on how the data is constructed.
+
+        Shift the incoming data index and episode_index to continue on from the last frame. Note that this
+        will be done in place!
+        """
+        if len(missing_keys := (set(self.data_keys).difference(set(data)))) > 0:
+            raise ValueError(f"Missing data keys: {missing_keys}")
+        new_data_length = len(data[self.data_keys[0]])
+        if not all(len(data[k]) == new_data_length for k in self.data_keys):
+            raise ValueError("All data items should have the same length")
+
+        next_index = self._data[OnlineBuffer.NEXT_INDEX_KEY]
+
+        # Sanity check to make sure that the new data indices start from 0.
+        assert data[OnlineBuffer.EPISODE_INDEX_KEY][0].item() == 0
+        assert data[OnlineBuffer.INDEX_KEY][0].item() == 0
+
+        # Shift the incoming indices if necessary.
+        if self.num_frames > 0:
+            last_episode_index = self._data[OnlineBuffer.EPISODE_INDEX_KEY][next_index - 1]
+            last_data_index = self._data[OnlineBuffer.INDEX_KEY][next_index - 1]
+            data[OnlineBuffer.EPISODE_INDEX_KEY] += last_episode_index + 1
+            data[OnlineBuffer.INDEX_KEY] += last_data_index + 1
+
+        # Insert the new data starting from next_index. It may be necessary to wrap around to the start.
+        n_surplus = max(0, new_data_length - (self._buffer_capacity - next_index))
+        for k in self.data_keys:
+            if n_surplus == 0:
+                slc = slice(next_index, next_index + new_data_length)
+                self._data[k][slc] = data[k]
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][slc] = True
+            else:
+                self._data[k][next_index:] = data[k][:-n_surplus]
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY][next_index:] = True
+                self._data[k][:n_surplus] = data[k][-n_surplus:]
+        if n_surplus == 0:
+            self._data[OnlineBuffer.NEXT_INDEX_KEY] = next_index + new_data_length
+        else:
+            self._data[OnlineBuffer.NEXT_INDEX_KEY] = n_surplus
+
+    @property
+    def data_keys(self) -> list[str]:
+        keys = set(self._data)
+        keys.remove(OnlineBuffer.OCCUPANCY_MASK_KEY)
+        keys.remove(OnlineBuffer.NEXT_INDEX_KEY)
+        return sorted(keys)
+
+    @property
+    def fps(self) -> float | None:
+        return self._fps
+
+    @property
+    def num_episodes(self) -> int:
+        return len(
+            np.unique(self._data[OnlineBuffer.EPISODE_INDEX_KEY][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
+        )
+
+    @property
+    def num_frames(self) -> int:
+        return np.count_nonzero(self._data[OnlineBuffer.OCCUPANCY_MASK_KEY])
+
+    def __len__(self):
+        return self.num_frames
+
+    def _item_to_tensors(self, item: dict) -> dict:
+        item_ = {}
+        for k, v in item.items():
+            if isinstance(v, torch.Tensor):
+                item_[k] = v
+            elif isinstance(v, np.ndarray):
+                item_[k] = torch.from_numpy(v)
+            else:
+                item_[k] = torch.tensor(v)
+        return item_
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        if idx >= len(self) or idx < -len(self):
+            raise IndexError
+
+        item = {k: v[idx] for k, v in self._data.items() if not k.startswith("_")}
+
+        if self.delta_timestamps is None:
+            return self._item_to_tensors(item)
+
+        episode_index = item[OnlineBuffer.EPISODE_INDEX_KEY]
+        current_ts = item[OnlineBuffer.TIMESTAMP_KEY]
+        episode_data_indices = np.where(
+            np.bitwise_and(
+                self._data[OnlineBuffer.EPISODE_INDEX_KEY] == episode_index,
+                self._data[OnlineBuffer.OCCUPANCY_MASK_KEY],
+            )
+        )[0]
+        episode_timestamps = self._data[OnlineBuffer.TIMESTAMP_KEY][episode_data_indices]
+
+        for data_key in self.delta_timestamps:
+            # Note: The logic in this loop is copied from `load_previous_and_future_frames`.
+            # Get timestamps used as query to retrieve data of previous/future frames.
+            query_ts = current_ts + self.delta_timestamps[data_key]
+
+            # Compute distances between each query timestamp and all timestamps of all the frames belonging to
+            # the episode.
+            dist = np.abs(query_ts[:, None] - episode_timestamps[None, :])
+            argmin_ = np.argmin(dist, axis=1)
+            min_ = dist[np.arange(dist.shape[0]), argmin_]
+
+            is_pad = min_ > self.tolerance_s
+
+            # Check violated query timestamps are all outside the episode range.
+            assert (
+                (query_ts[is_pad] < episode_timestamps[0]) | (episode_timestamps[-1] < query_ts[is_pad])
+            ).all(), (
+                f"One or several timestamps unexpectedly violate the tolerance ({min_} > {self.tolerance_s=}"
+                ") inside the episode range."
+            )
+
+            # Load frames for this data key.
+            item[data_key] = self._data[data_key][episode_data_indices[argmin_]]
+
+            item[f"{data_key}{OnlineBuffer.IS_PAD_POSTFIX}"] = is_pad
+
+        return self._item_to_tensors(item)
+
+    def get_data_by_key(self, key: str) -> torch.Tensor:
+        """Returns all data for a given data key as a Tensor."""
+        return torch.from_numpy(self._data[key][self._data[OnlineBuffer.OCCUPANCY_MASK_KEY]])
+
+
+def compute_sampler_weights(
+    offline_dataset: LeRobotDataset,
+    offline_drop_n_last_frames: int = 0,
+    online_dataset: OnlineBuffer | None = None,
+    online_sampling_ratio: float | None = None,
+    online_drop_n_last_frames: int = 0,
+) -> torch.Tensor:
+    """Compute the sampling weights for the online training dataloader in train.py.
+
+    Args:
+        offline_dataset: The LeRobotDataset used for offline pre-training.
+        online_drop_n_last_frames: Number of frames to drop from the end of each offline dataset episode.
+        online_dataset: The OnlineBuffer used in online training.
+        online_sampling_ratio: The proportion of data that should be sampled from the online dataset. If an
+            online dataset is provided, this value must also be provided.
+        online_drop_n_first_frames: See `offline_drop_n_last_frames`. This is the same, but for the online
+            dataset.
+    Returns:
+        Tensor of weights for [offline_dataset; online_dataset], normalized to 1.
+
+    Notes to maintainers:
+        - This duplicates some logic from EpisodeAwareSampler. We should consider converging to one approach.
+        - When used with `torch.utils.data.WeightedRandomSampler`, it could completely replace
+          `EpisodeAwareSampler` as the online dataset related arguments are optional. The only missing feature
+          is the ability to turn shuffling off.
+        - Options `drop_first_n_frames` and `episode_indices_to_use` can be added easily. They were not
+          included here to avoid adding complexity.
+    """
+    if len(offline_dataset) == 0 and (online_dataset is None or len(online_dataset) == 0):
+        raise ValueError("At least one of `offline_dataset` or `online_dataset` should be contain data.")
+    if (online_dataset is None) ^ (online_sampling_ratio is None):
+        raise ValueError(
+            "`online_dataset` and `online_sampling_ratio` must be provided together or not at all."
+        )
+    offline_sampling_ratio = 0 if online_sampling_ratio is None else 1 - online_sampling_ratio
+
+    weights = []
+
+    if len(offline_dataset) > 0:
+        offline_data_mask_indices = []
+        for start_index, end_index in zip(
+            offline_dataset.episode_data_index["from"],
+            offline_dataset.episode_data_index["to"],
+            strict=True,
+        ):
+            offline_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - offline_drop_n_last_frames)
+            )
+        offline_data_mask = torch.zeros(len(offline_dataset), dtype=torch.bool)
+        offline_data_mask[torch.tensor(offline_data_mask_indices)] = True
+        weights.append(
+            torch.full(
+                size=(len(offline_dataset),),
+                fill_value=offline_sampling_ratio / offline_data_mask.sum(),
+            )
+            * offline_data_mask
+        )
+
+    if online_dataset is not None and len(online_dataset) > 0:
+        online_data_mask_indices = []
+        episode_indices = online_dataset.get_data_by_key("episode_index")
+        for episode_idx in torch.unique(episode_indices):
+            where_episode = torch.where(episode_indices == episode_idx)
+            start_index = where_episode[0][0]
+            end_index = where_episode[0][-1] + 1
+            online_data_mask_indices.extend(
+                range(start_index.item(), end_index.item() - online_drop_n_last_frames)
+            )
+        online_data_mask = torch.zeros(len(online_dataset), dtype=torch.bool)
+        online_data_mask[torch.tensor(online_data_mask_indices)] = True
+        weights.append(
+            torch.full(
+                size=(len(online_dataset),),
+                fill_value=online_sampling_ratio / online_data_mask.sum(),
+            )
+            * online_data_mask
+        )
+
+    weights = torch.cat(weights)
+
+    if weights.sum() == 0:
+        weights += 1 / len(weights)
+    else:
+        weights /= weights.sum()
+
+    return weights

lerobot/common/datasets/push_dataset_to_hub/CODEBASE_VERSION.md

@@ -0,0 +1,56 @@
+## Using / Updating `CODEBASE_VERSION` (for maintainers)
+
+Since our dataset pushed to the hub are decoupled with the evolution of this repo, we ensure compatibility of
+the datasets with our code, we use a `CODEBASE_VERSION` (defined in
+lerobot/common/datasets/lerobot_dataset.py) variable.
+
+For instance, [`lerobot/pusht`](https://huggingface.co/datasets/lerobot/pusht) has many versions to maintain backward compatibility between LeRobot codebase versions:
+- [v1.0](https://huggingface.co/datasets/lerobot/pusht/tree/v1.0)
+- [v1.1](https://huggingface.co/datasets/lerobot/pusht/tree/v1.1)
+- [v1.2](https://huggingface.co/datasets/lerobot/pusht/tree/v1.2)
+- [v1.3](https://huggingface.co/datasets/lerobot/pusht/tree/v1.3)
+- [v1.4](https://huggingface.co/datasets/lerobot/pusht/tree/v1.4)
+- [v1.5](https://huggingface.co/datasets/lerobot/pusht/tree/v1.5)
+- [v1.6](https://huggingface.co/datasets/lerobot/pusht/tree/v1.6) <-- last version
+- [main](https://huggingface.co/datasets/lerobot/pusht/tree/main) <-- points to the last version
+
+Starting with v1.6, every dataset pushed to the hub or saved locally also have this version number in their
+`info.json` metadata.
+
+### Uploading a new dataset
+If you are pushing a new dataset, you don't need to worry about any of the instructions below, nor to be
+compatible with previous codebase versions. The `push_dataset_to_hub.py` script will automatically tag your
+dataset with the current `CODEBASE_VERSION`.
+
+### Updating an existing dataset
+If you want to update an existing dataset, you need to change the `CODEBASE_VERSION` from `lerobot_dataset.py`
+before running `push_dataset_to_hub.py`. This is especially useful if you introduce a breaking change
+intentionally or not (i.e. something not backward compatible such as modifying the reward functions used,
+deleting some frames at the end of an episode, etc.). That way, people running a previous version of the
+codebase won't be affected by your change and backward compatibility is maintained.
+
+However, you will need to update the version of ALL the other datasets so that they have the new
+`CODEBASE_VERSION` as a branch in their hugging face dataset repository. Don't worry, there is an easy way
+that doesn't require to run `push_dataset_to_hub.py`. You can just "branch-out" from the `main` branch on HF
+dataset repo by running this script which corresponds to a `git checkout -b` (so no copy or upload needed):
+
+```python
+from huggingface_hub import HfApi
+
+from lerobot import available_datasets
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+
+api = HfApi()
+
+for repo_id in available_datasets:
+    dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
+    branches = [b.name for b in dataset_info.branches]
+    if CODEBASE_VERSION in branches:
+        print(f"{repo_id} already @{CODEBASE_VERSION}, skipping.")
+        continue
+    else:
+        # Now create a branch named after the new version by branching out from "main"
+        # which is expected to be the preceding version
+        api.create_branch(repo_id, repo_type="dataset", branch=CODEBASE_VERSION, revision="main")
+        print(f"{repo_id} successfully updated @{CODEBASE_VERSION}")
+```

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_cabinet.txt

@@ -0,0 +1,85 @@
+https://drive.google.com/file/d/1_SOJkgfP5yZyVjMhTt3nwhvyUjcnlI51/view?usp=drive_link
+https://drive.google.com/file/d/1rmgN8UUzph1qwJnzG1d-uOafodn-gLvb/view?usp=drive_link
+https://drive.google.com/file/d/1NYQ-XxsBVinB6dUoZmVWweT83367P3i2/view?usp=drive_link
+https://drive.google.com/file/d/1oAv_j74zxxCJieMG7r5Vl2BeHK1__3s3/view?usp=drive_link
+https://drive.google.com/file/d/1wFUJQROsrTJt64YRuIeExhFjr2wnK5uu/view?usp=drive_link
+https://drive.google.com/file/d/1KzL3Tt0Le7jVl58XVRUcmigmXjyiuhbK/view?usp=drive_link
+https://drive.google.com/file/d/1qy_YBladeHtianSSGtgAPSHtMin7msvf/view?usp=drive_link
+https://drive.google.com/file/d/1rA_F0V_qL_nyuC_0aBKCisF4-0TIkF2Y/view?usp=drive_link
+https://drive.google.com/file/d/1hw-8qMpz9VgSt62XoASqNRuPECpCwJQP/view?usp=drive_link
+https://drive.google.com/file/d/1BpHOl9rKMzdvNGka6js7C0s40hH6vnDA/view?usp=drive_link
+https://drive.google.com/file/d/1PazhkhiDnJ-OUMyDVDFxEZNKQQqHiNWS/view?usp=drive_link
+https://drive.google.com/file/d/1lZ665R6ATl57dypxH4dGJ2NSt6XYnbuz/view?usp=drive_link
+https://drive.google.com/file/d/1V9HzLaf-tlG15wUzT7KrTDCS_z1vi5NV/view?usp=drive_link
+https://drive.google.com/file/d/1aKauWiXoKqbNwn_2xs4MrmLlaNYlVNmO/view?usp=drive_link
+https://drive.google.com/file/d/1WVD5DFhriO1YmmOgiVHhacR6HWoTPxav/view?usp=drive_link
+https://drive.google.com/file/d/1_X43WgeBAsfkhH9EmpyPki8U9joMeAGC/view?usp=drive_link
+https://drive.google.com/file/d/1t8x0GqWoNKWtnBsB7_D40Z34nL9ak4kf/view?usp=drive_link
+https://drive.google.com/file/d/15V_f26WaKOXjKnq2T3HRWAmtQUi4lbu2/view?usp=drive_link
+https://drive.google.com/file/d/11VFIAsiSDsMOBANgrOcZBpKB9AFWnLy7/view?usp=drive_link
+https://drive.google.com/file/d/1M0NS7vVaxJv3FHnuRYtdwTFYF7We4LxP/view?usp=drive_link
+https://drive.google.com/file/d/1mR0OItTNqFnVLoczcyKYlm6drAy778lO/view?usp=drive_link
+https://drive.google.com/file/d/1NbVFWDQAh-z4JJ4D-Zw6Lps9kdvpqh2j/view?usp=drive_link
+https://drive.google.com/file/d/1JQoZGBzl4W3QG26-n39tefcGN0fDRMbB/view?usp=drive_link
+https://drive.google.com/file/d/1VBjHl-TvZpncopvasIP5G9gecbB2a5f6/view?usp=drive_link
+https://drive.google.com/file/d/1VzSf6zaB21nahm7MsPwroXbJ84NIwq0b/view?usp=drive_link
+https://drive.google.com/file/d/1OtNnfMEydNtZOcivs4k6E_uJSpf8PkGy/view?usp=drive_link
+https://drive.google.com/file/d/14nVvpvsrFr_03Pa_N7MKzwnRwibOUYM6/view?usp=drive_link
+https://drive.google.com/file/d/1M8li6duiO2r3lv_9HhF_XJn0oZUIEK5F/view?usp=drive_link
+https://drive.google.com/file/d/1Cpzea6fO14lxAaNfSBifqoa4ekhCiLD1/view?usp=drive_link
+https://drive.google.com/file/d/1mbxRTm5vlbsY9UJ0jfjM6j9D7kPJjBpG/view?usp=drive_link
+https://drive.google.com/file/d/1RXD1i6IfWsHRlCxVmG04h2h5Ycm_WwZN/view?usp=drive_link
+https://drive.google.com/file/d/1QFqFSwDGOk1BkgGmqgCcc2BRWnJ6R3MA/view?usp=drive_link
+https://drive.google.com/file/d/1bFqWR8DQM0ZUxxtS2bl-RANQvukeFLzp/view?usp=drive_link
+https://drive.google.com/file/d/1pR-rH3yNGoyPdD4hJ6-3lXQ-PstBx9du/view?usp=drive_link
+https://drive.google.com/file/d/107OAwLY-hva9HeQLIK7VCh-ytdDabVjr/view?usp=drive_link
+https://drive.google.com/file/d/1Tpl08QOaSZ37GTO4awFWSdD8wBR9xdlT/view?usp=drive_link
+https://drive.google.com/file/d/1MR164AOM-0S1T6RX8xKTV2IHyaCvpqAW/view?usp=drive_link
+https://drive.google.com/file/d/1_wknJfVnStIhJ82lU_QtcrwahsqYIsr8/view?usp=drive_link
+https://drive.google.com/file/d/1ZuEktWrbYkTx0l5pj3WiZ2CJrfbDOHNo/view?usp=drive_link
+https://drive.google.com/file/d/15G_10hkkkq6yxvyI5NGZirlF-RzduR2F/view?usp=drive_link
+https://drive.google.com/file/d/1DBKxg3ONqh7dhLuX6oh1Yyo2x383V1Hp/view?usp=drive_link
+https://drive.google.com/file/d/1B5iDBkTUr5vopDddV_fHud18SqAHhauS/view?usp=drive_link
+https://drive.google.com/file/d/1acwFV0eenRkki1QcjSKH5xqOtys-P3Pr/view?usp=drive_link
+https://drive.google.com/file/d/1S47BI83xyrh-FKXsvAQqer98Biu_p8XK/view?usp=drive_link
+https://drive.google.com/file/d/1JL6DmBZl3uyq9dyLfgSqtGF06e7E9JwM/view?usp=drive_link
+https://drive.google.com/file/d/16WvRS4Kjog8Pxgr0E3sGGnI01YwL9Uql/view?usp=drive_link
+https://drive.google.com/file/d/12ttGqL33IPWg0-s1SD44rr22M6LiSQBr/view?usp=drive_link
+https://drive.google.com/file/d/1OyZqqnldTU_DliRbr6x0C4a_iWPwIN7j/view?usp=drive_link
+https://drive.google.com/file/d/1oYk00IpLnR9fesLfD15Ebe7nVBffEbcS/view?usp=drive_link
+https://drive.google.com/file/d/1eyE2-MQduCEqCd-5_kl5zsoOEERAzpZD/view?usp=drive_link
+https://drive.google.com/file/d/1ir1Ya-vO0d97pfvbePlUeuKTTRc0qIMU/view?usp=drive_link
+https://drive.google.com/file/d/1hOi-JnqlMt47gVnLZHMTqeojyYVErohl/view?usp=drive_link
+https://drive.google.com/file/d/1NFFw5_PqigQ7xGqsL-MNq2B1r5yAscCf/view?usp=drive_link
+https://drive.google.com/file/d/1uftq1-Zlh8d2sNLWrlVcKYQUwZTD7o24/view?usp=drive_link
+https://drive.google.com/file/d/1-ax19dSLPacVgk000T-m3l4flPcg07pM/view?usp=drive_link
+https://drive.google.com/file/d/126y-lgn86-ZmCz8hooF1THKJGGObw3OB/view?usp=drive_link
+https://drive.google.com/file/d/1JiDniK0VmDIkk92AbBILb8J2Ba59PWML/view?usp=drive_link
+https://drive.google.com/file/d/1kr8nPIRljiU0R4J9SMgj80o1FPQxzu9z/view?usp=drive_link
+https://drive.google.com/file/d/1bbThWRij1pKBh_kFgV8FwK0sXtTHBoLX/view?usp=drive_link
+https://drive.google.com/file/d/1WenzDW6lxk1xkOFm-OiGFfc0ROskAuKU/view?usp=drive_link
+https://drive.google.com/file/d/1MiKRzuzUn1yN-k_6kPJJzIGy7dT-nnsD/view?usp=drive_link
+https://drive.google.com/file/d/17rRg2tcmB-gNhQ0KoZJQmNfyFeoij1jH/view?usp=drive_link
+https://drive.google.com/file/d/11mokBpvrY3ld6sY5WztREtJ1jgqfQV70/view?usp=drive_link
+https://drive.google.com/file/d/1Il_6IOx9NDp1bX_KHizJfBwzTufTmn86/view?usp=drive_link
+https://drive.google.com/file/d/1KswtJGsxJ7eeBDAmNA_aeLjOxcH6MIxa/view?usp=drive_link
+https://drive.google.com/file/d/1gzMhi5uWu4C3Y6WbQ3L-08V96GxTZrRR/view?usp=drive_link
+https://drive.google.com/file/d/1nRQFtaBxfUCYc2W90Qibh0kHCt6YQCfc/view?usp=drive_link
+https://drive.google.com/file/d/1vs-gyW-KheqHbUATwAhA2mmR9GOGw7f_/view?usp=drive_link
+https://drive.google.com/file/d/1MuxzGOA2fgLaHryq82KkQumtuRJGcUOC/view?usp=drive_link
+https://drive.google.com/file/d/1IIwxZnGlqrXLUXqG6yMO0r7uhCvhpk9e/view?usp=drive_link
+https://drive.google.com/file/d/1vE7XPyaFcXP4DtTY5Y9WKIt7zWgmX-Cr/view?usp=drive_link
+https://drive.google.com/file/d/1j-bIV09gr21RC3-x1N_pK4RPLV3fmWKz/view?usp=drive_link
+https://drive.google.com/file/d/1t3nW1rD3S-EL0Oymb5U7ZAj5UMkydkln/view?usp=drive_link
+https://drive.google.com/file/d/14hbfHCdMKtJZ41F9CQReMec2jeRFTOqR/view?usp=drive_link
+https://drive.google.com/file/d/1x-hUyOSne5BW0AzQ3W6_Pf4g5yXQWi9M/view?usp=drive_link
+https://drive.google.com/file/d/1sw9JqRg6E-3P84I3ZhzTrJMu0vuiaMmP/view?usp=drive_link
+https://drive.google.com/file/d/1LuqhQlL4MGZhB_6THmkovRxrlP26BbdC/view?usp=drive_link
+https://drive.google.com/file/d/15C5K6v_lkjnMSmUvVyqHQKwh2N166e7K/view?usp=drive_link
+https://drive.google.com/file/d/1ns_9eSsQeeoZ10nlbkLy8tu0GmJFSnkt/view?usp=drive_link
+https://drive.google.com/file/d/1NpzWJeK6CqjxzjIMYe6aYdX8xGsQwD4o/view?usp=drive_link
+https://drive.google.com/file/d/1NMLezwufKJ9_8xTc9KQThSzVVD71B9Ui/view?usp=drive_link
+https://drive.google.com/file/d/1aa71DCUqs6oXlIxX35jgsmsgm-NlDxPV/view?usp=drive_link
+https://drive.google.com/file/d/1UJzkIZzAL0j-D5YQBnoq7mHvttASy12O/view?usp=drive_link
+https://drive.google.com/file/d/1nPgx36HIJFb7oI94VbRzWjpPP2GANxzG/view?usp=drive_link
+https://drive.google.com/file/d/1NovAP-KVJjqcuvWy3d6G4ptGGAIDqcCx/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_chair.txt

@@ -0,0 +1,55 @@
+https://drive.google.com/file/d/11M3Ye0r5agMaaicPbVGD0q2Hb3rGklbb/view?usp=drive_link
+https://drive.google.com/file/d/1-tx7SvYYgSvXCvnf_EI2OVdwK-CkFY6S/view?usp=drive_link
+https://drive.google.com/file/d/1EWJunmOpMHaU1hE106wwpbkGYcjQXYAF/view?usp=drive_link
+https://drive.google.com/file/d/1IDn95Z7FSiCckrSENtGV4u3RyFHNQSDY/view?usp=drive_link
+https://drive.google.com/file/d/1CwzvWj1i7QOtqrZvsCZ6BdZaKNDfpN32/view?usp=drive_link
+https://drive.google.com/file/d/1HvAvlhm77nAD3Td24QPSeq8lw-Rl_aOh/view?usp=drive_link
+https://drive.google.com/file/d/1t-suKYOPhXH666RpAYNRp2QU_DOy3AeM/view?usp=drive_link
+https://drive.google.com/file/d/18xpKgWh7RWyjMN5PkLTOo-AxsAadAuRw/view?usp=drive_link
+https://drive.google.com/file/d/1oci5Eto-ztv-AQNz8EnwZveBIhxvk-xJ/view?usp=drive_link
+https://drive.google.com/file/d/1Y-t_4vxdE6NpHO0DLJR8f3mD0Q-Wj5-c/view?usp=drive_link
+https://drive.google.com/file/d/1lylRqbbbB8bgtpsBWMPACmHJreuKmllv/view?usp=drive_link
+https://drive.google.com/file/d/1yliSyMig_NXShWfQx6qyW7Ijf2Y5lFK6/view?usp=drive_link
+https://drive.google.com/file/d/1XXhwJsJbeb7KXAooGvJapnm9bjnGUmxS/view?usp=drive_link
+https://drive.google.com/file/d/1_xs1f3hW2JArKyvfF7UWubWjyROGTLs6/view?usp=drive_link
+https://drive.google.com/file/d/1WVEHpr6EqKCZbkHapQSTXJq4xE4SWFT-/view?usp=drive_link
+https://drive.google.com/file/d/1RqOHv9pEQGvW8NUA7ynffFmG999TL_Az/view?usp=drive_link
+https://drive.google.com/file/d/1cu5AgD2gh-uA3PFJmzxxzNaF3qOSlYY1/view?usp=drive_link
+https://drive.google.com/file/d/1SsrXqiPclNrnYToPZ9Uq-k3y0C4qdHT1/view?usp=drive_link
+https://drive.google.com/file/d/1-J7EXf0vjkLIfSqT8ICEsP6CTjzSLBop/view?usp=drive_link
+https://drive.google.com/file/d/11O7ewUmoZXfyyKjy_6B5RW4DpjICxqBT/view?usp=drive_link
+https://drive.google.com/file/d/1iic44kZoCsjNsfAz2cMstZ9-WQvAhblF/view?usp=drive_link
+https://drive.google.com/file/d/1yLV1lVX-2WnWQldGlnQZ0x7QBuDiVkL3/view?usp=drive_link
+https://drive.google.com/file/d/1Tybp9ru98TTbGn4eyROpUQwDFuALWXmk/view?usp=drive_link
+https://drive.google.com/file/d/13E9OTMiipVJByDs5-J19oWwAz7l94LTN/view?usp=drive_link
+https://drive.google.com/file/d/1EeTpJQdMSliw4JzSMtJ6CyTvVdexjM4M/view?usp=drive_link
+https://drive.google.com/file/d/1NHyNwoFqzeAu-1_PSpq5JfxaiD_xbpn9/view?usp=drive_link
+https://drive.google.com/file/d/1fJcS0phDp4xm_FyGaJ5wr9Pe4KqtHaxD/view?usp=drive_link
+https://drive.google.com/file/d/12AqrLUaewDPEcFRqPZeZFb_TQ0Lfi3At/view?usp=drive_link
+https://drive.google.com/file/d/1x_hd4Qsq1oJS-aj2t3qM7WbbV7KZj05b/view?usp=drive_link
+https://drive.google.com/file/d/14OUSUArmsB068hs6BuEIXQhI1Cyz8Sf0/view?usp=drive_link
+https://drive.google.com/file/d/16zlzh1T5zeUJQnFf382NXkFEKEnDub4O/view?usp=drive_link
+https://drive.google.com/file/d/1IbDltmN-NEFCNtr1TO4ILxEgQ94rtjWv/view?usp=drive_link
+https://drive.google.com/file/d/15gmlf8Gx9455pZ1AlqcCSwh3nDPxMzSr/view?usp=drive_link
+https://drive.google.com/file/d/1qHpRL1oZfIMo_vxnm8qfwQ-7l0BZIVva/view?usp=drive_link
+https://drive.google.com/file/d/1H1xskIgiFZivkYn23rMzH3xePGOh3VTC/view?usp=drive_link
+https://drive.google.com/file/d/1avls6Pv0kYiCMNVknbc1zQsgy64MUDMM/view?usp=drive_link
+https://drive.google.com/file/d/1MmWVgCj5khc8KMIifmt3EzF1o-CtPyyn/view?usp=drive_link
+https://drive.google.com/file/d/1U0kCc_xqW0WNppf4sbnK14euWKdPZtzB/view?usp=drive_link
+https://drive.google.com/file/d/16CaEyQscOuhLj23PEGDTL9DeyNkohkMn/view?usp=drive_link
+https://drive.google.com/file/d/1Iu8uM6UUJ0zW8tvN-9UiOe_4oSNzEutg/view?usp=drive_link
+https://drive.google.com/file/d/1UImqiBaIxCR-1DNJaZhHqeHhaySOtVIr/view?usp=drive_link
+https://drive.google.com/file/d/1VpU2V_leIoRIyv_lAvE7eLHBG8DxCTnp/view?usp=drive_link
+https://drive.google.com/file/d/1_Q8J27OT3Xby7QY6yHvIJauFRWEMxkRm/view?usp=drive_link
+https://drive.google.com/file/d/1bantmVo1L9Xz4tbiNw_a1UC2Z_HPO1wT/view?usp=drive_link
+https://drive.google.com/file/d/1IRIXMJMCBDkBjbaHvAlEiBogSvZ1jK_3/view?usp=drive_link
+https://drive.google.com/file/d/1mAHXKjiFbjwydypW2t5Lv8_H5x6nHegl/view?usp=drive_link
+https://drive.google.com/file/d/1SfyY796fLrBCMY39OcyuxZafqSCRZPZk/view?usp=drive_link
+https://drive.google.com/file/d/1X-44sZ8CcfzIskc0dvSx882o1yFhHaZB/view?usp=drive_link
+https://drive.google.com/file/d/1BOIWCCCk6DLD4Bmvc75ZbbLi9AQm-1ao/view?usp=drive_link
+https://drive.google.com/file/d/1RuyDtRE1kk76sw-wP8vx5SgLoPF3PA_H/view?usp=drive_link
+https://drive.google.com/file/d/1c4eoQiBbGuy3CTAQDUSkd84Ponh1roAQ/view?usp=drive_link
+https://drive.google.com/file/d/19PXB9z4Ljq6dsbf9TqcOrrP5SRbw2Tc_/view?usp=drive_link
+https://drive.google.com/file/d/1nn1VVZVoIXWdYDozR7XHXE4mPLQG80PQ/view?usp=drive_link
+https://drive.google.com/file/d/1MBdFGOKPV8GUhwoSsJ_Ky3qAMLM2Bv3K/view?usp=drive_link
+https://drive.google.com/file/d/1of3k_M-7Nh3I1TndcWedxK4ca9dn8Sc5/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_elevator.txt

@@ -0,0 +1,20 @@
+https://drive.google.com/file/d/12ctkOAdkCNGN1JLbZb5ww3XTBn2LFpGI/view?usp=drive_link
+https://drive.google.com/file/d/1G_Vd46_4fq6O64gHHjUbJX5Ld44ZZx0y/view?usp=drive_link
+https://drive.google.com/file/d/1uKgUy73B3xBogQAOUhfZjO0X5qZGsi2c/view?usp=drive_link
+https://drive.google.com/file/d/1fu9cIrfI-fE2LhdGUxbx7-8Ci_PF8Ypm/view?usp=drive_link
+https://drive.google.com/file/d/1Ygk9ZPJzx8xw2A9JF3NHbJ44TqnvSTQR/view?usp=drive_link
+https://drive.google.com/file/d/18m5xPuccNsEB20WPshm3zhxmXc6k63ED/view?usp=drive_link
+https://drive.google.com/file/d/1DiqqxC44rriviRQpqogcv0-EB-Y6nr9g/view?usp=drive_link
+https://drive.google.com/file/d/1qPdaoTVDizJXkfXLioWU7iJ8hqCXSyOQ/view?usp=drive_link
+https://drive.google.com/file/d/1Fj9kIA_mG7f67WFfACJEaZ7izcHG7vUm/view?usp=drive_link
+https://drive.google.com/file/d/1WpYehZnI2P7dUdJPfkE-ij1rqCnjZEbB/view?usp=drive_link
+https://drive.google.com/file/d/1_zwWkT4jPyzB38STWb6whlzsPzXmfA9r/view?usp=drive_link
+https://drive.google.com/file/d/1U6-J4I_fPlSFFGfhZPxS5_YzKXwXIZYp/view?usp=drive_link
+https://drive.google.com/file/d/1pRhxxcTfZp5tQo_EScvJUwfc3amiS6Vk/view?usp=drive_link
+https://drive.google.com/file/d/1lWLntqra83RlYU_gN7Vostnfydf6gutd/view?usp=drive_link
+https://drive.google.com/file/d/1vIBKo0x-NYEHV1FvRpco1lQMpRdAWAIL/view?usp=drive_link
+https://drive.google.com/file/d/1pdrLV3JTQou_XH0Aap61Ssf60iVKm1jJ/view?usp=drive_link
+https://drive.google.com/file/d/1QTsLoQ7SwmKdQHjBGVDaR2uTwfFwtrOf/view?usp=drive_link
+https://drive.google.com/file/d/1Gytai8M_12J36GY6L_TulEcOC-035jwS/view?usp=drive_link
+https://drive.google.com/file/d/14LJudNc629NT-i8xreXtzl27ce_DxOFJ/view?usp=drive_link
+https://drive.google.com/file/d/1sBvPCODbzxGAI0S3lgN5cSG9Go3lRi00/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_shrimp.txt

@@ -0,0 +1,18 @@
+https://drive.google.com/file/d/1MJn9GbC8p9lN4gC9KDMLEkTkP_gGpXj0/view?usp=drive_link
+https://drive.google.com/file/d/1-4LXgjl7ZCOgp-8GCJmFRD8OeqN5Jf7-/view?usp=drive_link
+https://drive.google.com/file/d/1Ho06Ce0SPbqU3juaMxNUwAt3zCRLGC8W/view?usp=drive_link
+https://drive.google.com/file/d/1ivHoj7_7olBSxH-Y8kqXEW7ttITK-45j/view?usp=drive_link
+https://drive.google.com/file/d/1qjY4hM_IvZ8cq2II_n9MeJbvyeuN4oBP/view?usp=drive_link
+https://drive.google.com/file/d/1rKVhO_f92-7sw13T8hTVrza3B9oAVgoy/view?usp=drive_link
+https://drive.google.com/file/d/1pcLPHO8fBkc1-CRa88tyQtEueE4xiXNi/view?usp=drive_link
+https://drive.google.com/file/d/1Vev_chCsIeEdvQ8poEYNsOJFGy_QU8kZ/view?usp=drive_link
+https://drive.google.com/file/d/1l5G4zpRkxSLCQjvGPYSN4zfCvVRQuzMz/view?usp=drive_link
+https://drive.google.com/file/d/14vgthE1eoakXkr2-DRw50E6lAqYOiUuE/view?usp=drive_link
+https://drive.google.com/file/d/17nPSmKKmgQ2B7zkzWrZYiLM3RBuFod82/view?usp=drive_link
+https://drive.google.com/file/d/1QcDsxplVvb_ID9BVrihl5FvlC-j7waXi/view?usp=drive_link
+https://drive.google.com/file/d/18pEejBpI-eEVaWAAjBCyC0vgbX3T1Esj/view?usp=drive_link
+https://drive.google.com/file/d/1H8eH6_IRODtEFT6WoM77ltR5OoOrqXmI/view?usp=drive_link
+https://drive.google.com/file/d/1IWlpFRZhoxyG4nS13CWK4leZVk5wbNx4/view?usp=drive_link
+https://drive.google.com/file/d/1PbZA8_OCGmMLxNP9xbkLRSChniL4uGxl/view?usp=drive_link
+https://drive.google.com/file/d/1p9XAdmG2f_WeflNO4DIJ_tr1rK6M9B4B/view?usp=drive_link
+https://drive.google.com/file/d/1nS59Et1cNAvKo3Y4SeSGRuZD5TvBbCF3/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_wash_pan.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1S8eFg98IaGAIKVZ8QFWG1bx4mHa-O204

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/mobile_wipe_wine.txt

@@ -0,0 +1,4 @@
+https://drive.google.com/drive/folders/1tC_g1AJ8lglBLY-fjsQrG6DMBa3Ucp-0
+https://drive.google.com/file/d/1fG_Yi2MJrFjiUVN3XoiWXLtTxHlwwaDv/view?usp=drive_link
+https://drive.google.com/file/d/1WX32VWfzzX3Blmd06DRxLwFbMJfVe7P4/view?usp=drive_link
+https://drive.google.com/file/d/18onsX3vXg3xkFwP5bVUCjdV4n9TRn0C9/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_insertion_human.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1RgyD0JgTX30H4IM5XZn8I3zSV_mr8pyF
+https://drive.google.com/file/d/18Cudl6nikDtgRolea7je8iF_gGKzynOP/view?usp=drive_link
+https://drive.google.com/file/d/1C1kZYyROzs-PrLc0SkDgUgMi4-L3lauE/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_insertion_scripted.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1TsojQQSXtHEoGnqgJ3gmpPQR2DPLtS2N
+https://drive.google.com/file/d/1wfMSZ24oOh5KR_0aaP3Cnu_c4ZCveduB/view?usp=drive_link
+https://drive.google.com/file/d/17EuCUWS6uCCr6yyNzpXdcdE-_TTNCKtf/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_transfer_cube_human.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1sc-E4QYW7A0o23m1u2VWNGVq5smAsfCo
+https://drive.google.com/file/d/18smMymtr8tIxaNUQ61gW6dG50pt3MvGq/view?usp=drive_link
+https://drive.google.com/file/d/1Nk7l53d9sJoGDBKAOnNrExX5nLacATc6/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/sim_transfer_cube_scripted.txt

@@ -0,0 +1,3 @@
+https://drive.google.com/drive/folders/1aRyoOhQwxhyt1J8XgEig4s6kzaw__LXj
+https://drive.google.com/file/d/1pnGIOd-E4-rhz2P3VxpknMKRZCoKt6eI/view?usp=drive_link
+https://drive.google.com/file/d/1GKReZHrXU73NMiC5zKCq_UtqPVtYq8eo/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_battery.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/19qS_n7vKgDcPeTMnvDHQ5-n73xEbJz5D
+https://drive.google.com/file/d/1oC31By0A2bsBeHyUwBdQw1z4ng6yi9Za/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_candy.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1m5rQ6UVH8Q9RQp_6c0CxkQ88-L-ScO7q
+https://drive.google.com/file/d/1wHz2qcmwcVG0C0CZ9MjQDQcmj4OY9_a3/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_coffee.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1seQGay470nGQ-knBI5TjsTr8iL9Qws5q
+https://drive.google.com/file/d/1T89hSX5U99wLGvGTE7yUBaQPOpyj6Sai/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_coffee_new.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1t3eDc5Rg0DveyRe8oTm6Dia_FYU5mXyf
+https://drive.google.com/file/d/1TXFaduTakvS0ZWJqKCX-HIvYglum_5CY/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_cups_open.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1Z9X3DNzd6LS0FFjQemNUMoMA5yk5VQOh
+https://drive.google.com/file/d/1Wlyc0vTkjXuWB6zbaVOWhEfD7BmPgUV_/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_fork_pick_up.txt

@@ -0,0 +1,53 @@
+https://drive.google.com/drive/folders/1DYgB4ifX4uIid9m9jnC0Zdz8Nf7ZC0fc
+https://drive.google.com/file/d/1Eb-NRNk_FmVleCbU_Ng5Y4dfcjTKN7Rv/view?usp=drive_link
+https://drive.google.com/file/d/1dkhjEADakT-44l9jf-nK4x89kr4yG_qb/view?usp=drive_link
+https://drive.google.com/file/d/14hDhgcZkVqNExGb4tIXpSjMshhqZETch/view?usp=drive_link
+https://drive.google.com/file/d/1zVMEHpHbuNyP5A_lYU7RPSLB-4V0yfZw/view?usp=drive_link
+https://drive.google.com/file/d/1JtgDjBvy7FnRpFzrx_foC3quorYQFAR-/view?usp=drive_link
+https://drive.google.com/file/d/1EHdneB6F-PP0dQlX8qPaXbxmKoBy_YwO/view?usp=drive_link
+https://drive.google.com/file/d/17Z0jjVBy1OPKREPu77_n_rQzorDiapji/view?usp=drive_link
+https://drive.google.com/file/d/1F4i23qPJ_qTf5jWjfLo4ARGJChznYWt3/view?usp=drive_link
+https://drive.google.com/file/d/1kZtXWM3uS0-rLblydBfJ0mMcVnMMXw9w/view?usp=drive_link
+https://drive.google.com/file/d/1mNODox87xFfY5Z_o5mcLsr8SHb39jDik/view?usp=drive_link
+https://drive.google.com/file/d/1Ob44VdmEUA93FKDECiRb5Ogz2xQg5IWp/view?usp=drive_link
+https://drive.google.com/file/d/1fdQLdjj3Cwv33R1wZhfrLz9Del8mqgHb/view?usp=drive_link
+https://drive.google.com/file/d/1Yu3L3ft21zP__XL8pCfhb788ZleuW1n5/view?usp=drive_link
+https://drive.google.com/file/d/1ozBBWXVZ9hXDh9ooHUNroHdYm8UDqnhJ/view?usp=drive_link
+https://drive.google.com/file/d/1o0TGqvfWw_Lunxb5ubKDS21Lr_WC0h75/view?usp=drive_link
+https://drive.google.com/file/d/1jZnd5eP5L6BH5l98BPN6OnoQx3fu8e9n/view?usp=drive_link
+https://drive.google.com/file/d/1S5sYbz8wcLYp0V67v13i4PRcBxodn4Hg/view?usp=drive_link
+https://drive.google.com/file/d/1rFeg_x6ftJYwPtBv34D3h2L2cpDLeR4G/view?usp=drive_link
+https://drive.google.com/file/d/1GvS3lcm4o6nm_scUk0XxKeVFNmzjucDZ/view?usp=drive_link
+https://drive.google.com/file/d/1-9i0riphC7NhhDahcQfD1QoBXP5gF90A/view?usp=drive_link
+https://drive.google.com/file/d/15p_IqGsMbKuvzMS872THAZr-3SBtb1Fr/view?usp=drive_link
+https://drive.google.com/file/d/1ToyYcBfJL8gbQn0q_59zPLsFmm7dmMJo/view?usp=drive_link
+https://drive.google.com/file/d/1e_7PNH7CYafE4pAebP7ZdI7XFbmEcy_i/view?usp=drive_link
+https://drive.google.com/file/d/1JoabvGVsIQdug2xOhUIhetEIyDM91y_Y/view?usp=drive_link
+https://drive.google.com/file/d/1kOMw1y0lmnVaCjwZICfzCsx6e0Z8MNGR/view?usp=drive_link
+https://drive.google.com/file/d/16it_wd1JOevUQTK2_CvF_pBACTgpIPgM/view?usp=drive_link
+https://drive.google.com/file/d/1IRcCj9HnJSfbyMgr5XEERGlEnWeZQwOc/view?usp=drive_link
+https://drive.google.com/file/d/1Z2dIJfq_S3liGmPN9Rphvkmucnmw7tlb/view?usp=drive_link
+https://drive.google.com/file/d/1J3NoAjzndGx9yNyaBOJHdNny1epzUoBt/view?usp=drive_link
+https://drive.google.com/file/d/18nOvxV1k8FSmBrhT4TPo2sKKSZXougyx/view?usp=drive_link
+https://drive.google.com/file/d/1CT8FxclafFMjSd7gCWVw3VSeryeiF04i/view?usp=drive_link
+https://drive.google.com/file/d/16M9KVqQMFfSsXfypK0bocFft8Nz3j2Rt/view?usp=drive_link
+https://drive.google.com/file/d/18QPVkw6bj6HW8LTPrQLWrrUX4R6RcF42/view?usp=drive_link
+https://drive.google.com/file/d/1hQTVtA5hBTE_StXpJafTZJ3tgt2VQQ_t/view?usp=drive_link
+https://drive.google.com/file/d/1Dn-d5g69H6EgAWgsFdrcbJKtz7ySsCQ8/view?usp=drive_link
+https://drive.google.com/file/d/13hMr16483P7ALYv73yMRUN37fJdVQM62/view?usp=drive_link
+https://drive.google.com/file/d/1848yN3XMN5zJMEgApt6KzrWgfRPfimtv/view?usp=drive_link
+https://drive.google.com/file/d/1oAD9kSnS0fTgj-CjD4u9VdZ5X67IOIMa/view?usp=drive_link
+https://drive.google.com/file/d/1ilzIWLCCG5b_KgF5s0wdN2I5-lFNpwC1/view?usp=drive_link
+https://drive.google.com/file/d/1rjsT2YBjnidxod1s9s-myAYz8boHr-WB/view?usp=drive_link
+https://drive.google.com/file/d/18Gg48HTub15bd8qzbhiCUufbVy0fbN5G/view?usp=drive_link
+https://drive.google.com/file/d/1WsSnQSqmMTVSRwrhT1Y-v782My2zcjLm/view?usp=drive_link
+https://drive.google.com/file/d/1ea9ZCvoyc-xqiFXgeDcA_mOWsw7VUuoi/view?usp=drive_link
+https://drive.google.com/file/d/1wv1v3-XhPgbNzp62BXbJTDzMPu2tlDUc/view?usp=drive_link
+https://drive.google.com/file/d/18-ikzt8LoZ83Gi3goKCELs4U4z8hrRoF/view?usp=drive_link
+https://drive.google.com/file/d/16Bjhp7JNCXkGuLvyNcZowAx3W-Y-15DV/view?usp=drive_link
+https://drive.google.com/file/d/1Gc-KRI-xwcp1fMR55ugbrLg_5y3SPde-/view?usp=drive_link
+https://drive.google.com/file/d/1oP72Q386Z4Sy5MMm-t5yNogIe5Van_9k/view?usp=drive_link
+https://drive.google.com/file/d/112T90eDUDVH-SyOV7UnZl5bscAH2hcfq/view?usp=drive_link
+https://drive.google.com/file/d/1y-uKOesRRhjgDtFbG_j65f4SGg0v8XDg/view?usp=drive_link
+https://drive.google.com/file/d/1LOP05OagoI3km-ZKQBrS204A85UVk7Ok/view?usp=drive_link
+https://drive.google.com/file/d/1QkHQKgasVzWsmdPvkXgGhWyQ84d93_Az/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_pingpong_test.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1Ut2cv6o6Pkfgg46DgwVUM7Z5PkNG8eJ-

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_pro_pencil.txt

@@ -0,0 +1 @@
+https://drive.google.com/drive/folders/1FqxPV0PgvgIu8XFjtvZSPSExuNcxVVAY

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_screw_driver.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1SKtG0ct9q0nVdYssJNMWSOjikcXliT58
+https://drive.google.com/file/d/1nchD21O30B3i3LDoqramo1zgW5YvpJIN/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_tape.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1_4DHf2cma0xsChLQFghwigX6Ukti5-zQ
+https://drive.google.com/file/d/1_8vS4hDNDgUQY-SmekrNaa7dF67QJYU-/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_thread_velcro.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1_4DHf2cma0xsChLQFghwigX6Ukti5-zQ
+https://drive.google.com/file/d/1_8vS4hDNDgUQY-SmekrNaa7dF67QJYU-/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_towel.txt

@@ -0,0 +1,2 @@
+https://drive.google.com/drive/folders/1fAD7vkyTGTFB_nGXIKofCU1U05oE3MFv
+https://drive.google.com/file/d/1XzyQ2B6LLvcurIonOpEu4nij2qwNWshH/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_vinh_cup.txt

@@ -0,0 +1,53 @@
+https://drive.google.com/drive/folders/13EQsVsnxT86K20QAoyE_YpsFbQ7fZQdu
+https://drive.google.com/file/d/1-W_JHghZG65FNTVhw1SXhtQrazdLL3Ue/view?usp=drive_link
+https://drive.google.com/file/d/1VwRJgdWUo-2nQaNM7Bs77-fsm8iwUxEo/view?usp=drive_link
+https://drive.google.com/file/d/1wFzGRo5iYA13WLi6IV1ry64RyahQBFio/view?usp=drive_link
+https://drive.google.com/file/d/1IKtQzQ-n-UTv64hYpReu2R4cqUvmNQqD/view?usp=drive_link
+https://drive.google.com/file/d/1GicVci9OiuuZZH79i5Mg7AtWod94MzwT/view?usp=drive_link
+https://drive.google.com/file/d/1JVnIoR7EIQp70T4eAf9RX65JcTrzsjQc/view?usp=drive_link
+https://drive.google.com/file/d/1W2xr4h23ucjPrc-mBEeqnACsfaImpc0p/view?usp=drive_link
+https://drive.google.com/file/d/10xj_0V7A07o3uCa7v5omUrTC0YlPW8H3/view?usp=drive_link
+https://drive.google.com/file/d/1FOc3EMaCy8Mb0_a7PuXLAwKwvxkbKmwU/view?usp=drive_link
+https://drive.google.com/file/d/143PgDXBcf2GQ0Q07ZPMVMfBgZDd5sLJG/view?usp=drive_link
+https://drive.google.com/file/d/1pE5Tyj0LlGbGWvUzuhixp86Ibu55Ez3I/view?usp=drive_link
+https://drive.google.com/file/d/141668b1VzX80ncrVJPzhkoAeIFB4MEK9/view?usp=drive_link
+https://drive.google.com/file/d/1bw12lo37p1ZvRvErHsll7cEYi2OxscvZ/view?usp=drive_link
+https://drive.google.com/file/d/1zfnMFvbgBjl6SzYhksbaOzfbwLrCN6tb/view?usp=drive_link
+https://drive.google.com/file/d/1-GIszA6mUJMaNB-tdh9r9skc77SWA0VX/view?usp=drive_link
+https://drive.google.com/file/d/1fTB0zWFYU6zh4IIUFT2zX_OkwYqmElwY/view?usp=drive_link
+https://drive.google.com/file/d/1gPIPNKGmrO9c7gKF7SP0SuUYbIBBq8z1/view?usp=drive_link
+https://drive.google.com/file/d/12JeJ-dQd5lYyn6PlDOGdE-ChVeiZ-Uv0/view?usp=drive_link
+https://drive.google.com/file/d/100_20cgCqerU6qoh3TfTbwLy9mlDAFEG/view?usp=drive_link
+https://drive.google.com/file/d/111oAGJ76ku_pYgbBoIdZAC1_XEQcPI__/view?usp=drive_link
+https://drive.google.com/file/d/1UhC8L-354ZQ2gblPFGI35EMsVwfpuKa0/view?usp=drive_link
+https://drive.google.com/file/d/1sIXQSgUR_xdrNtGrL6QGBnkLMKErsIp1/view?usp=drive_link
+https://drive.google.com/file/d/16Ax77bDSIXnsn4GFL8XYKKT1P6bPpfMd/view?usp=drive_link
+https://drive.google.com/file/d/1pgRVYwwVIsWq_qsWqZpe1UBzZfF5Fa9D/view?usp=drive_link
+https://drive.google.com/file/d/1jtimaZkWsY1P5gC2bbS64H_WCUU7HXN2/view?usp=drive_link
+https://drive.google.com/file/d/1N6Bh02P-RiTEgtx1YH1Db_X3TGpP-X_r/view?usp=drive_link
+https://drive.google.com/file/d/14Fy8EwJ8d9Vh97Yt1VOvUChSCrfIjBij/view?usp=drive_link
+https://drive.google.com/file/d/1IRuv42dvIMPuKhcMZmuXaBjJ-lPFOmQd/view?usp=drive_link
+https://drive.google.com/file/d/16XWzNY2D8ucVVn5geBgsVdhm3ppO4que/view?usp=drive_link
+https://drive.google.com/file/d/1xsVOoQgthK_L_SDrmq_JvQgUpAvPEAY8/view?usp=drive_link
+https://drive.google.com/file/d/1bZbw66DyEMvnJnzkdUUNbKjvNKg8KFYM/view?usp=drive_link
+https://drive.google.com/file/d/1CyTVkdrNGGpouCXr4CfhKbMzE6Ah3oo3/view?usp=drive_link
+https://drive.google.com/file/d/1hDRyeM-XEDpHXpptbT8LvNnlQUR3PWOh/view?usp=drive_link
+https://drive.google.com/file/d/1XhHWxbra8Iy5irQZ83IvxwaJqHq9x4s1/view?usp=drive_link
+https://drive.google.com/file/d/1haZcn6aM1o4JlmP9tJj3x2enrxiPaDSD/view?usp=drive_link
+https://drive.google.com/file/d/1ypDyuUTbljaBZ34f-t7lj3O_0bRmyX2n/view?usp=drive_link
+https://drive.google.com/file/d/1ILEEZo_tA9_ChIAprr2mPaNVKZi5vXsO/view?usp=drive_link
+https://drive.google.com/file/d/1U7nVYFaGE8vVTfLCW33D74xOjDcqfgyJ/view?usp=drive_link
+https://drive.google.com/file/d/1rZ93_rmCov5SMDxPkfM3qthcRELZrQX6/view?usp=drive_link
+https://drive.google.com/file/d/1mYO1b_csddtyE3qT6cwLiw-m2w2_1Lxh/view?usp=drive_link
+https://drive.google.com/file/d/1xz7Q5x2jikY8wJQjMRQpRws6AnfWlHm5/view?usp=drive_link
+https://drive.google.com/file/d/1OO8GaO-0FrSZRd1kxMYwBmubyiLOWnbl/view?usp=drive_link
+https://drive.google.com/file/d/1EXn4NVDmf-4_HCy34mYwT-vwK2CFI9ev/view?usp=drive_link
+https://drive.google.com/file/d/10hH70XhXRL9C5SnAG4toHtfHqfJUJo4H/view?usp=drive_link
+https://drive.google.com/file/d/18tiBcxea0guUai4lwsXQvt0q2LZ8ZnnJ/view?usp=drive_link
+https://drive.google.com/file/d/1Q8R8qv37vk5PQ5kQ2ibx6BFLOySD0VpX/view?usp=drive_link
+https://drive.google.com/file/d/17aNriHzjhdibCyuUjQoMFZqjybJZtggG/view?usp=drive_link
+https://drive.google.com/file/d/1LVjEYHSdeKm6CotU1QguIeNEPaIaFl_1/view?usp=drive_link
+https://drive.google.com/file/d/1ufAhE_EkgJ85slg2EW8aW_grOzE_Lmxd/view?usp=drive_link
+https://drive.google.com/file/d/1wtzLtXrkw9eXRGESTPIOlpl1tInu-b2m/view?usp=drive_link
+https://drive.google.com/file/d/1Mk5qvVtD_QHwGOUApRq76TUw2T5THu6f/view?usp=drive_link
+https://drive.google.com/file/d/1y1WQ3hboWVJ68KEYQQ3OhreGuaUpSgwc/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_vinh_cup_left.txt

@@ -0,0 +1,52 @@
+https://drive.google.com/drive/folders/1dxWh6YFZUDt6qXIoxgD9bla3CiFjZ11C
+https://drive.google.com/file/d/1hNBJN00SCAlOl0ZEgm7RRGbAGDjyBs0p/view?usp=drive_link
+https://drive.google.com/file/d/17He0CVwXGeoMmXg4SHKo-osNn7YPKVL7/view?usp=drive_link
+https://drive.google.com/file/d/1laNKUVID1x2CV6a2O2WQjwFewKu4lidL/view?usp=drive_link
+https://drive.google.com/file/d/1pNf36xbZJGRArYLmNAvRj5y6CoqdC6kB/view?usp=drive_link
+https://drive.google.com/file/d/1_4E1-y3JXk5I0ebycLYM70YDPK9g52gZ/view?usp=drive_link
+https://drive.google.com/file/d/1PHfzhGPdbolKyOpS3FnR2w7Q8zUlJXSk/view?usp=drive_link
+https://drive.google.com/file/d/17ls2PPN-Pi3tEuK059cwV2_iDT8aGhOO/view?usp=drive_link
+https://drive.google.com/file/d/1LWsg6PmCT00Kv_N_slrmcwKmQPGoBT3k/view?usp=drive_link
+https://drive.google.com/file/d/12LckrchoHTUVH7rxi8J7zD9dA19GXvoW/view?usp=drive_link
+https://drive.google.com/file/d/1VqrJKjAIkj5gtFXL69grdSeu9CyaqnSw/view?usp=drive_link
+https://drive.google.com/file/d/1g5rQYDBZvW-kUtYPeyF3qmd53v6k7kXu/view?usp=drive_link
+https://drive.google.com/file/d/10kUgaSJ0TS7teaG83G3Rf_DG4XGrBt6A/view?usp=drive_link
+https://drive.google.com/file/d/1je9XmneZQZvTma5adMJICUPDovW3ppei/view?usp=drive_link
+https://drive.google.com/file/d/1v28r6bedwZGbUPVVTVImXhK-42XdtGfj/view?usp=drive_link
+https://drive.google.com/file/d/1-TEEx9sGVvzMMaNXYfQMtY2JJ6cvl0dT/view?usp=drive_link
+https://drive.google.com/file/d/1YdBKdJFP9rJWBUX7qrOYL_gfUA8o6J9M/view?usp=drive_link
+https://drive.google.com/file/d/1X9vffwQHNUSKLXr2RlYNtbWDIFCIDfdF/view?usp=drive_link
+https://drive.google.com/file/d/11hqesqa5kvEe5FABUnZRcvmOhR373cYM/view?usp=drive_link
+https://drive.google.com/file/d/1ltTTECjEcbQPgS3UPRgMzaE2x9n6H7dC/view?usp=drive_link
+https://drive.google.com/file/d/1Zxqfa29JdwT-bfMpivi6IG2vz34d21dD/view?usp=drive_link
+https://drive.google.com/file/d/11LQlVxS5hz494dYUJ_PNRPx2NHIJbQns/view?usp=drive_link
+https://drive.google.com/file/d/1i1JhNtnZpO_E8rAv8gxBP3ZTZRvcvsZi/view?usp=drive_link
+https://drive.google.com/file/d/11jOXAr2EULUO4Qkm748634lg4UUFho5U/view?usp=drive_link
+https://drive.google.com/file/d/1rj67wur8DdB_Pipwx24bY43xu4X1eQ5e/view?usp=drive_link
+https://drive.google.com/file/d/15ZTm6lO6f_JQy_4SNfrOu3iPYn1Ro8mh/view?usp=drive_link
+https://drive.google.com/file/d/1q4gBtqWPJtCwXEvknGgN0WHGp7Vfn1b9/view?usp=drive_link
+https://drive.google.com/file/d/1t17keyre47AYqm8GgXiQ7EcvcUkeSiDQ/view?usp=drive_link
+https://drive.google.com/file/d/1OYUPGxtZgOF86Ng_BEOTXm_XOYpuQPsO/view?usp=drive_link
+https://drive.google.com/file/d/1cBjbGHi3dwWHtx6r9EQJi0JT_CE3LuHt/view?usp=drive_link
+https://drive.google.com/file/d/14qaMyF0mcbCB-fCYKNyo5_2NahSC6D5u/view?usp=drive_link
+https://drive.google.com/file/d/12FgX86eA7Y5co9ULBVK80XMsiKQSs-Ri/view?usp=drive_link
+https://drive.google.com/file/d/1yvoHWidf-jdBVw6qCCXOFfkVwKj_2hPk/view?usp=drive_link
+https://drive.google.com/file/d/1a2SugsSDlC8UtUrFzp-_KAwyZckQOvdQ/view?usp=drive_link
+https://drive.google.com/file/d/1l8pILBFSAosypWJMza2K09Vm7rug9axm/view?usp=drive_link
+https://drive.google.com/file/d/1hfPQ8dBCk97PnOhq6_MIISm3IEzcOxJG/view?usp=drive_link
+https://drive.google.com/file/d/1PPAUwlJCFKpms8cqF_k1v2_fCgDBOc3S/view?usp=drive_link
+https://drive.google.com/file/d/1lVKQZeqFfK3amEmLuFhYLUFQ2eyE8rOW/view?usp=drive_link
+https://drive.google.com/file/d/1K9iPMLfDowcIFoyzpvgn88dQ6x6kVwNG/view?usp=drive_link
+https://drive.google.com/file/d/1PNvMqG9tL7QxeLaYBGHiWYR6SYb5iIct/view?usp=drive_link
+https://drive.google.com/file/d/1xkRtzbvIkUsylx9hrFLGQsJn0h1EYu-5/view?usp=drive_link
+https://drive.google.com/file/d/1nxMRrJlSayjDIfr5CmHO1NzAw3COhsLi/view?usp=drive_link
+https://drive.google.com/file/d/1Qs3WEyMGrmagiHIkkFEueWNnJhkUeR1s/view?usp=drive_link
+https://drive.google.com/file/d/1D-G2_Q0SS3M8zyJbg_XzkF2ANPw1HTuX/view?usp=drive_link
+https://drive.google.com/file/d/1mdmJsDGO-YtJAOF_yPKl6lq4PJOIbQhT/view?usp=drive_link
+https://drive.google.com/file/d/11m9bwfop_sPmnQr_8amB6EEsrbAeG_z5/view?usp=drive_link
+https://drive.google.com/file/d/19tyYt5FMn5kru0g9o2nMJhKPnsDqkIZv/view?usp=drive_link
+https://drive.google.com/file/d/1XvTpUdsVTZ-vydvdYYmynbma--HfUGSl/view?usp=drive_link
+https://drive.google.com/file/d/1MO3hFu68J6NohTzr9aB_fY02VA6QSOqj/view?usp=drive_link
+https://drive.google.com/file/d/1Lh-UjwAk__04YOTWINF_QGVU8SjetVaY/view?usp=drive_link
+https://drive.google.com/file/d/1jkSOUwZV5GJ7rZlVeErjcu0DBQs8Np0d/view?usp=drive_link
+https://drive.google.com/file/d/1VIN1eLI-93WrVQwCjsv6XQr353DqqBYA/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls/static_ziploc_slide.txt

@@ -0,0 +1,8 @@
+https://drive.google.com/drive/folders/1EgKar7rWBmTIRmeJYZciSwjZx3uP2mHO
+https://drive.google.com/file/d/12eYWQO15atK2hBjXhynPJd9MKAj_42pz/view?usp=drive_link
+https://drive.google.com/file/d/1Ul4oEeICJDjgfYTl4H1uaisTzVYIM6wd/view?usp=drive_link
+https://drive.google.com/file/d/1WSF-OG8lKSe2wVYCv5D1aJNipxpgddk-/view?usp=drive_link
+https://drive.google.com/file/d/1_ppD5j5sFh26aWW0JmhLzJMeNB-lCArk/view?usp=drive_link
+https://drive.google.com/file/d/1WUp846dgWXYhu4oJfhHxiU6YL_7N6s4W/view?usp=drive_link
+https://drive.google.com/file/d/1HRZNAIoAQw_uYiPwnBvtBioQoqiqoXdA/view?usp=drive_link
+https://drive.google.com/file/d/1hedGq-QDMnIn8GlXXBC3GiEJ_Y-LTxyt/view?usp=drive_link

lerobot/common/datasets/push_dataset_to_hub/_diffusion_policy_replay_buffer.py

@@ -1,3 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 """Helper code for loading PushT dataset from Diffusion Policy (https://diffusion-policy.cs.columbia.edu/)
 
 Copied from the original Diffusion Policy repository and used in our `download_and_upload_dataset.py` script.

lerobot/common/datasets/push_dataset_to_hub/_download_raw.py

@@ -1,179 +1,202 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 """
-This file contains all obsolete download scripts. They are centralized here to not have to load
-useless dependencies when using datasets.
+This file contains download scripts for raw datasets.
+
+Example of usage:
+```
+python lerobot/common/datasets/push_dataset_to_hub/_download_raw.py \
+--raw-dir data/lerobot-raw/pusht_raw \
+--repo-id lerobot-raw/pusht_raw
+```
 """
 
-import io
+import argparse
+import logging
+import warnings
 from pathlib import Path
 
-import tqdm
+from huggingface_hub import snapshot_download
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
+
+# {raw_repo_id: raw_format}
+AVAILABLE_RAW_REPO_IDS = {
+    "lerobot-raw/aloha_mobile_cabinet_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_chair_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_elevator_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_shrimp_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_wash_pan_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_mobile_wipe_wine_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_insertion_human_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_insertion_scripted_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_transfer_cube_human_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_sim_transfer_cube_scripted_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_battery_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_candy_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_coffee_new_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_coffee_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_cups_open_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_fork_pick_up_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_pingpong_test_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_pro_pencil_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_screw_driver_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_tape_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_thread_velcro_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_towel_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_vinh_cup_left_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_vinh_cup_raw": "aloha_hdf5",
+    "lerobot-raw/aloha_static_ziploc_slide_raw": "aloha_hdf5",
+    "lerobot-raw/umi_cup_in_the_wild_raw": "umi_zarr",
+    "lerobot-raw/pusht_raw": "pusht_zarr",
+    "lerobot-raw/unitreeh1_fold_clothes_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_rearrange_objects_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_two_robot_greeting_raw": "aloha_hdf5",
+    "lerobot-raw/unitreeh1_warehouse_raw": "aloha_hdf5",
+    "lerobot-raw/xarm_lift_medium_raw": "xarm_pkl",
+    "lerobot-raw/xarm_lift_medium_replay_raw": "xarm_pkl",
+    "lerobot-raw/xarm_push_medium_raw": "xarm_pkl",
+    "lerobot-raw/xarm_push_medium_replay_raw": "xarm_pkl",
+    "lerobot-raw/fractal20220817_data_raw": "openx_rlds.fractal20220817_data",
+    "lerobot-raw/kuka_raw": "openx_rlds.kuka",
+    "lerobot-raw/bridge_openx_raw": "openx_rlds.bridge_openx",
+    "lerobot-raw/taco_play_raw": "openx_rlds.taco_play",
+    "lerobot-raw/jaco_play_raw": "openx_rlds.jaco_play",
+    "lerobot-raw/berkeley_cable_routing_raw": "openx_rlds.berkeley_cable_routing",
+    "lerobot-raw/roboturk_raw": "openx_rlds.roboturk",
+    "lerobot-raw/nyu_door_opening_surprising_effectiveness_raw": "openx_rlds.nyu_door_opening_surprising_effectiveness",
+    "lerobot-raw/viola_raw": "openx_rlds.viola",
+    "lerobot-raw/berkeley_autolab_ur5_raw": "openx_rlds.berkeley_autolab_ur5",
+    "lerobot-raw/toto_raw": "openx_rlds.toto",
+    "lerobot-raw/language_table_raw": "openx_rlds.language_table",
+    "lerobot-raw/columbia_cairlab_pusht_real_raw": "openx_rlds.columbia_cairlab_pusht_real",
+    "lerobot-raw/stanford_kuka_multimodal_dataset_raw": "openx_rlds.stanford_kuka_multimodal_dataset",
+    "lerobot-raw/nyu_rot_dataset_raw": "openx_rlds.nyu_rot_dataset",
+    "lerobot-raw/io_ai_tech_raw": "openx_rlds.io_ai_tech",
+    "lerobot-raw/stanford_hydra_dataset_raw": "openx_rlds.stanford_hydra_dataset",
+    "lerobot-raw/austin_buds_dataset_raw": "openx_rlds.austin_buds_dataset",
+    "lerobot-raw/nyu_franka_play_dataset_raw": "openx_rlds.nyu_franka_play_dataset",
+    "lerobot-raw/maniskill_dataset_raw": "openx_rlds.maniskill_dataset",
+    "lerobot-raw/furniture_bench_dataset_raw": "openx_rlds.furniture_bench_dataset",
+    "lerobot-raw/cmu_franka_exploration_dataset_raw": "openx_rlds.cmu_franka_exploration_dataset",
+    "lerobot-raw/ucsd_kitchen_dataset_raw": "openx_rlds.ucsd_kitchen_dataset",
+    "lerobot-raw/ucsd_pick_and_place_dataset_raw": "openx_rlds.ucsd_pick_and_place_dataset",
+    "lerobot-raw/spoc_raw": "openx_rlds.spoc",
+    "lerobot-raw/austin_sailor_dataset_raw": "openx_rlds.austin_sailor_dataset",
+    "lerobot-raw/austin_sirius_dataset_raw": "openx_rlds.austin_sirius_dataset",
+    "lerobot-raw/bc_z_raw": "openx_rlds.bc_z",
+    "lerobot-raw/utokyo_pr2_opening_fridge_raw": "openx_rlds.utokyo_pr2_opening_fridge",
+    "lerobot-raw/utokyo_pr2_tabletop_manipulation_raw": "openx_rlds.utokyo_pr2_tabletop_manipulation",
+    "lerobot-raw/utokyo_xarm_pick_and_place_raw": "openx_rlds.utokyo_xarm_pick_and_place",
+    "lerobot-raw/utokyo_xarm_bimanual_raw": "openx_rlds.utokyo_xarm_bimanual",
+    "lerobot-raw/utokyo_saytap_raw": "openx_rlds.utokyo_saytap",
+    "lerobot-raw/robo_net_raw": "openx_rlds.robo_net",
+    "lerobot-raw/robo_set_raw": "openx_rlds.robo_set",
+    "lerobot-raw/berkeley_mvp_raw": "openx_rlds.berkeley_mvp",
+    "lerobot-raw/berkeley_rpt_raw": "openx_rlds.berkeley_rpt",
+    "lerobot-raw/kaist_nonprehensile_raw": "openx_rlds.kaist_nonprehensile",
+    "lerobot-raw/stanford_mask_vit_raw": "openx_rlds.stanford_mask_vit",
+    "lerobot-raw/tokyo_u_lsmo_raw": "openx_rlds.tokyo_u_lsmo",
+    "lerobot-raw/dlr_sara_pour_raw": "openx_rlds.dlr_sara_pour",
+    "lerobot-raw/dlr_sara_grid_clamp_raw": "openx_rlds.dlr_sara_grid_clamp",
+    "lerobot-raw/dlr_edan_shared_control_raw": "openx_rlds.dlr_edan_shared_control",
+    "lerobot-raw/asu_table_top_raw": "openx_rlds.asu_table_top",
+    "lerobot-raw/stanford_robocook_raw": "openx_rlds.stanford_robocook",
+    "lerobot-raw/imperialcollege_sawyer_wrist_cam_raw": "openx_rlds.imperialcollege_sawyer_wrist_cam",
+    "lerobot-raw/iamlab_cmu_pickup_insert_raw": "openx_rlds.iamlab_cmu_pickup_insert",
+    "lerobot-raw/uiuc_d3field_raw": "openx_rlds.uiuc_d3field",
+    "lerobot-raw/utaustin_mutex_raw": "openx_rlds.utaustin_mutex",
+    "lerobot-raw/berkeley_fanuc_manipulation_raw": "openx_rlds.berkeley_fanuc_manipulation",
+    "lerobot-raw/cmu_playing_with_food_raw": "openx_rlds.cmu_playing_with_food",
+    "lerobot-raw/cmu_play_fusion_raw": "openx_rlds.cmu_play_fusion",
+    "lerobot-raw/cmu_stretch_raw": "openx_rlds.cmu_stretch",
+    "lerobot-raw/berkeley_gnm_recon_raw": "openx_rlds.berkeley_gnm_recon",
+    "lerobot-raw/berkeley_gnm_cory_hall_raw": "openx_rlds.berkeley_gnm_cory_hall",
+    "lerobot-raw/berkeley_gnm_sac_son_raw": "openx_rlds.berkeley_gnm_sac_son",
+    "lerobot-raw/droid_raw": "openx_rlds.droid",
+    "lerobot-raw/droid_100_raw": "openx_rlds.droid100",
+    "lerobot-raw/fmb_raw": "openx_rlds.fmb",
+    "lerobot-raw/dobbe_raw": "openx_rlds.dobbe",
+    "lerobot-raw/usc_cloth_sim_raw": "openx_rlds.usc_cloth_sim",
+    "lerobot-raw/plex_robosuite_raw": "openx_rlds.plex_robosuite",
+    "lerobot-raw/conq_hose_manipulation_raw": "openx_rlds.conq_hose_manipulation",
+    "lerobot-raw/vima_raw": "openx_rlds.vima",
+    "lerobot-raw/robot_vqa_raw": "openx_rlds.robot_vqa",
+    "lerobot-raw/mimic_play_raw": "openx_rlds.mimic_play",
+    "lerobot-raw/tidybot_raw": "openx_rlds.tidybot",
+    "lerobot-raw/eth_agent_affordances_raw": "openx_rlds.eth_agent_affordances",
+}
 
 
-def download_raw(root, dataset_id) -> Path:
-    if "pusht" in dataset_id:
-        return download_pusht(root=root, dataset_id=dataset_id)
-    elif "xarm" in dataset_id:
-        return download_xarm(root=root, dataset_id=dataset_id)
-    elif "aloha" in dataset_id:
-        return download_aloha(root=root, dataset_id=dataset_id)
-    elif "umi" in dataset_id:
-        return download_umi(root=root, dataset_id=dataset_id)
-    else:
-        raise ValueError(dataset_id)
+def download_raw(raw_dir: Path, repo_id: str):
+    check_repo_id(repo_id)
+    user_id, dataset_id = repo_id.split("/")
+
+    if not dataset_id.endswith("_raw"):
+        warnings.warn(
+            f"""`dataset_id` ({dataset_id}) doesn't end with '_raw' (e.g. 'lerobot/pusht_raw'). Following this
+             naming convention by renaming your repository is advised, but not mandatory.""",
+            stacklevel=1,
+        )
+
+    # Send warning if raw_dir isn't well formated
+    if raw_dir.parts[-2] != user_id or raw_dir.parts[-1] != dataset_id:
+        warnings.warn(
+            f"""`raw_dir` ({raw_dir}) doesn't contain a community or user id `/` the name of the dataset that
+             match the `repo_id` (e.g. 'data/lerobot/pusht_raw'). Following this naming convention is advised,
+             but not mandatory.""",
+            stacklevel=1,
+        )
+    raw_dir.mkdir(parents=True, exist_ok=True)
+
+    logging.info(f"Start downloading from huggingface.co/{user_id} for {dataset_id}")
+    snapshot_download(repo_id, repo_type="dataset", local_dir=raw_dir)
+    logging.info(f"Finish downloading from huggingface.co/{user_id} for {dataset_id}")
 
 
-def download_and_extract_zip(url: str, destination_folder: Path) -> bool:
-    import zipfile
-
-    import requests
-
-    print(f"downloading from {url}")
-    response = requests.get(url, stream=True)
-    if response.status_code == 200:
-        total_size = int(response.headers.get("content-length", 0))
-        progress_bar = tqdm.tqdm(total=total_size, unit="B", unit_scale=True)
-
-        zip_file = io.BytesIO()
-        for chunk in response.iter_content(chunk_size=1024):
-            if chunk:
-                zip_file.write(chunk)
-                progress_bar.update(len(chunk))
-
-        progress_bar.close()
-
-        zip_file.seek(0)
-
-        with zipfile.ZipFile(zip_file, "r") as zip_ref:
-            zip_ref.extractall(destination_folder)
-        return True
-    else:
-        return False
+def download_all_raw_datasets(data_dir: Path | None = None):
+    if data_dir is None:
+        data_dir = Path("data")
+    for repo_id in AVAILABLE_RAW_REPO_IDS:
+        raw_dir = data_dir / repo_id
+        download_raw(raw_dir, repo_id)
 
 
-def download_pusht(root: str, dataset_id: str = "pusht", fps: int = 10) -> Path:
-    pusht_url = "https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip"
-    pusht_zarr = Path("pusht/pusht_cchi_v7_replay.zarr")
+def main():
+    parser = argparse.ArgumentParser(
+        description=f"""A script to download raw datasets from Hugging Face hub to a local directory. Here is a
+            non exhaustive list of available repositories to use in `--repo-id`: {list(AVAILABLE_RAW_REPO_IDS.keys())}""",
+    )
 
-    root = Path(root)
-    raw_dir: Path = root / f"{dataset_id}_raw"
-    zarr_path: Path = (raw_dir / pusht_zarr).resolve()
-    if not zarr_path.is_dir():
-        raw_dir.mkdir(parents=True, exist_ok=True)
-        download_and_extract_zip(pusht_url, raw_dir)
-    return zarr_path
-
-
-def download_xarm(root: str, dataset_id: str, fps: int = 15) -> Path:
-    root = Path(root)
-    raw_dir: Path = root / "xarm_datasets_raw"
-    if not raw_dir.exists():
-        import zipfile
-
-        import gdown
-
-        raw_dir.mkdir(parents=True, exist_ok=True)
-        # from https://github.com/fyhMer/fowm/blob/main/scripts/download_datasets.py
-        url = "https://drive.google.com/uc?id=1nhxpykGtPDhmQKm-_B8zBSywVRdgeVya"
-        zip_path = raw_dir / "data.zip"
-        gdown.download(url, str(zip_path), quiet=False)
-        print("Extracting...")
-        with zipfile.ZipFile(str(zip_path), "r") as zip_f:
-            for member in zip_f.namelist():
-                if member.startswith("data/xarm") and member.endswith(".pkl"):
-                    print(member)
-                    zip_f.extract(member=member)
-        zip_path.unlink()
-
-    dataset_path: Path = root / f"{dataset_id}"
-    return dataset_path
-
-
-def download_aloha(root: str, dataset_id: str) -> Path:
-    folder_urls = {
-        "aloha_sim_insertion_human": "https://drive.google.com/drive/folders/1RgyD0JgTX30H4IM5XZn8I3zSV_mr8pyF",
-        "aloha_sim_insertion_scripted": "https://drive.google.com/drive/folders/1TsojQQSXtHEoGnqgJ3gmpPQR2DPLtS2N",
-        "aloha_sim_transfer_cube_human": "https://drive.google.com/drive/folders/1sc-E4QYW7A0o23m1u2VWNGVq5smAsfCo",
-        "aloha_sim_transfer_cube_scripted": "https://drive.google.com/drive/folders/1aRyoOhQwxhyt1J8XgEig4s6kzaw__LXj",
-    }
-
-    ep48_urls = {
-        "aloha_sim_insertion_human": "https://drive.google.com/file/d/18Cudl6nikDtgRolea7je8iF_gGKzynOP/view?usp=drive_link",
-        "aloha_sim_insertion_scripted": "https://drive.google.com/file/d/1wfMSZ24oOh5KR_0aaP3Cnu_c4ZCveduB/view?usp=drive_link",
-        "aloha_sim_transfer_cube_human": "https://drive.google.com/file/d/18smMymtr8tIxaNUQ61gW6dG50pt3MvGq/view?usp=drive_link",
-        "aloha_sim_transfer_cube_scripted": "https://drive.google.com/file/d/1pnGIOd-E4-rhz2P3VxpknMKRZCoKt6eI/view?usp=drive_link",
-    }
-
-    ep49_urls = {
-        "aloha_sim_insertion_human": "https://drive.google.com/file/d/1C1kZYyROzs-PrLc0SkDgUgMi4-L3lauE/view?usp=drive_link",
-        "aloha_sim_insertion_scripted": "https://drive.google.com/file/d/17EuCUWS6uCCr6yyNzpXdcdE-_TTNCKtf/view?usp=drive_link",
-        "aloha_sim_transfer_cube_human": "https://drive.google.com/file/d/1Nk7l53d9sJoGDBKAOnNrExX5nLacATc6/view?usp=drive_link",
-        "aloha_sim_transfer_cube_scripted": "https://drive.google.com/file/d/1GKReZHrXU73NMiC5zKCq_UtqPVtYq8eo/view?usp=drive_link",
-    }
-    num_episodes = {  # noqa: F841 # we keep this for reference
-        "aloha_sim_insertion_human": 50,
-        "aloha_sim_insertion_scripted": 50,
-        "aloha_sim_transfer_cube_human": 50,
-        "aloha_sim_transfer_cube_scripted": 50,
-    }
-
-    episode_len = {  # noqa: F841 # we keep this for reference
-        "aloha_sim_insertion_human": 500,
-        "aloha_sim_insertion_scripted": 400,
-        "aloha_sim_transfer_cube_human": 400,
-        "aloha_sim_transfer_cube_scripted": 400,
-    }
-
-    cameras = {  # noqa: F841 # we keep this for reference
-        "aloha_sim_insertion_human": ["top"],
-        "aloha_sim_insertion_scripted": ["top"],
-        "aloha_sim_transfer_cube_human": ["top"],
-        "aloha_sim_transfer_cube_scripted": ["top"],
-    }
-    root = Path(root)
-    raw_dir: Path = root / f"{dataset_id}_raw"
-    if not raw_dir.is_dir():
-        import gdown
-
-        assert dataset_id in folder_urls
-        assert dataset_id in ep48_urls
-        assert dataset_id in ep49_urls
-
-        raw_dir.mkdir(parents=True, exist_ok=True)
-
-        gdown.download_folder(folder_urls[dataset_id], output=str(raw_dir))
-
-        # because of the 50 files limit per directory, two files episode 48 and 49 were missing
-        gdown.download(ep48_urls[dataset_id], output=str(raw_dir / "episode_48.hdf5"), fuzzy=True)
-        gdown.download(ep49_urls[dataset_id], output=str(raw_dir / "episode_49.hdf5"), fuzzy=True)
-    return raw_dir
-
-
-def download_umi(root: str, dataset_id: str) -> Path:
-    url_cup_in_the_wild = "https://real.stanford.edu/umi/data/zarr_datasets/cup_in_the_wild.zarr.zip"
-    cup_in_the_wild_zarr = Path("umi/cup_in_the_wild/cup_in_the_wild.zarr")
-
-    root = Path(root)
-    raw_dir: Path = root / f"{dataset_id}_raw"
-    zarr_path: Path = (raw_dir / cup_in_the_wild_zarr).resolve()
-    if not zarr_path.is_dir():
-        raw_dir.mkdir(parents=True, exist_ok=True)
-        download_and_extract_zip(url_cup_in_the_wild, zarr_path)
-    return zarr_path
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        required=True,
+        help="Directory containing input raw datasets (e.g. `data/aloha_mobile_chair_raw` or `data/pusht_raw).",
+    )
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="""Repositery identifier on Hugging Face: a community or a user name `/` the name of
+        the dataset (e.g. `lerobot/pusht_raw`, `cadene/aloha_sim_insertion_human_raw`).""",
+    )
+    args = parser.parse_args()
+    download_raw(**vars(args))
 
 
 if __name__ == "__main__":
-    root = "data"
-    dataset_ids = [
-        "pusht",
-        "xarm_lift_medium",
-        "xarm_lift_medium_replay",
-        "xarm_push_medium",
-        "xarm_push_medium_replay",
-        "aloha_sim_insertion_human",
-        "aloha_sim_insertion_scripted",
-        "aloha_sim_transfer_cube_human",
-        "aloha_sim_transfer_cube_scripted",
-        "umi_cup_in_the_wild",
-    ]
-    for dataset_id in dataset_ids:
-        download_raw(root=root, dataset_id=dataset_id)
+    main()

lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py

@@ -0,0 +1,184 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Use this script to batch encode lerobot dataset from their raw format to LeRobotDataset and push their updated
+version to the hub. Under the hood, this script reuses 'push_dataset_to_hub.py'. It assumes that you already
+downloaded raw datasets, which you can do with the related '_download_raw.py' script.
+
+For instance, for codebase_version = 'v1.6', the following command was run, assuming raw datasets from
+lerobot-raw were downloaded in 'raw/datasets/directory':
+```bash
+python lerobot/common/datasets/push_dataset_to_hub/_encode_datasets.py \
+  --raw-dir raw/datasets/directory \
+  --raw-repo-ids lerobot-raw \
+  --local-dir push/datasets/directory \
+  --tests-data-dir tests/data \
+  --push-repo lerobot \
+  --vcodec libsvtav1 \
+  --pix-fmt yuv420p \
+  --g 2 \
+  --crf 30
+```
+"""
+
+import argparse
+from pathlib import Path
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub._download_raw import AVAILABLE_RAW_REPO_IDS
+from lerobot.common.datasets.push_dataset_to_hub.utils import check_repo_id
+from lerobot.scripts.push_dataset_to_hub import push_dataset_to_hub
+
+
+def get_push_repo_id_from_raw(raw_repo_id: str, push_repo: str) -> str:
+    dataset_id_raw = raw_repo_id.split("/")[1]
+    dataset_id = dataset_id_raw.removesuffix("_raw")
+    return f"{push_repo}/{dataset_id}"
+
+
+def encode_datasets(
+    raw_dir: Path,
+    raw_repo_ids: list[str],
+    push_repo: str,
+    vcodec: str,
+    pix_fmt: str,
+    g: int,
+    crf: int,
+    local_dir: Path | None = None,
+    tests_data_dir: Path | None = None,
+    raw_format: str | None = None,
+    dry_run: bool = False,
+) -> None:
+    if len(raw_repo_ids) == 1 and raw_repo_ids[0].lower() == "lerobot-raw":
+        raw_repo_ids_format = AVAILABLE_RAW_REPO_IDS
+    else:
+        if raw_format is None:
+            raise ValueError(raw_format)
+        raw_repo_ids_format = {id_: raw_format for id_ in raw_repo_ids}
+
+    for raw_repo_id, repo_raw_format in raw_repo_ids_format.items():
+        check_repo_id(raw_repo_id)
+        dataset_repo_id_push = get_push_repo_id_from_raw(raw_repo_id, push_repo)
+        dataset_raw_dir = raw_dir / raw_repo_id
+        dataset_dir = local_dir / dataset_repo_id_push if local_dir is not None else None
+        encoding = {
+            "vcodec": vcodec,
+            "pix_fmt": pix_fmt,
+            "g": g,
+            "crf": crf,
+        }
+
+        if not (dataset_raw_dir).is_dir():
+            raise NotADirectoryError(dataset_raw_dir)
+
+        if not dry_run:
+            push_dataset_to_hub(
+                dataset_raw_dir,
+                raw_format=repo_raw_format,
+                repo_id=dataset_repo_id_push,
+                local_dir=dataset_dir,
+                resume=True,
+                encoding=encoding,
+                tests_data_dir=tests_data_dir,
+            )
+        else:
+            print(
+                f"DRY RUN: {dataset_raw_dir}  -->  {dataset_dir}  -->  {dataset_repo_id_push}@{CODEBASE_VERSION}"
+            )
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        default=Path("data"),
+        help="Directory where raw datasets are located.",
+    )
+    parser.add_argument(
+        "--raw-repo-ids",
+        type=str,
+        nargs="*",
+        default=["lerobot-raw"],
+        help="""Raw dataset repo ids. if 'lerobot-raw', the keys from `AVAILABLE_RAW_REPO_IDS` will be
+            used and raw datasets will be fetched from the 'lerobot-raw/' repo and pushed with their
+            associated format. It is assumed that each dataset is located at `raw_dir / raw_repo_id` """,
+    )
+    parser.add_argument(
+        "--raw-format",
+        type=str,
+        default=None,
+        help="""Raw format to use for the raw repo-ids. Must be specified if --raw-repo-ids is not
+            'lerobot-raw'""",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="""When provided, writes the dataset converted to LeRobotDataset format in this directory
+        (e.g. `data/lerobot/aloha_mobile_chair`).""",
+    )
+    parser.add_argument(
+        "--push-repo",
+        type=str,
+        default="lerobot",
+        help="Repo to upload datasets to",
+    )
+    parser.add_argument(
+        "--vcodec",
+        type=str,
+        default="libsvtav1",
+        help="Codec to use for encoding videos",
+    )
+    parser.add_argument(
+        "--pix-fmt",
+        type=str,
+        default="yuv420p",
+        help="Pixel formats (chroma subsampling) to be used for encoding",
+    )
+    parser.add_argument(
+        "--g",
+        type=int,
+        default=2,
+        help="Group of pictures sizes to be used for encoding.",
+    )
+    parser.add_argument(
+        "--crf",
+        type=int,
+        default=30,
+        help="Constant rate factors to be used for encoding.",
+    )
+    parser.add_argument(
+        "--tests-data-dir",
+        type=Path,
+        default=None,
+        help=(
+            "When provided, save tests artifacts into the given directory "
+            "(e.g. `--tests-data-dir tests/data` will save to tests/data/{--repo-id})."
+        ),
+    )
+    parser.add_argument(
+        "--dry-run",
+        type=int,
+        default=0,
+        help="If not set to 0, this script won't download or upload anything.",
+    )
+    args = parser.parse_args()
+    encode_datasets(**vars(args))
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/push_dataset_to_hub/_umi_imagecodecs_numcodecs.py

@@ -1,3 +1,18 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 # imagecodecs/numcodecs.py
 
 # Copyright (c) 2021-2022, Christoph Gohlke

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -0,0 +1,233 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Contains utilities to process raw data format of HDF5 files like in: https://github.com/tonyzhaozh/act
+"""
+
+import gc
+import shutil
+from pathlib import Path
+
+import h5py
+import numpy as np
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def get_cameras(hdf5_data):
+    # ignore depth channel, not currently handled
+    # TODO(rcadene): add depth
+    rgb_cameras = [key for key in hdf5_data["/observations/images"].keys() if "depth" not in key]  # noqa: SIM118
+    return rgb_cameras
+
+
+def check_format(raw_dir) -> bool:
+    # only frames from simulation are uncompressed
+    compressed_images = "sim" not in raw_dir.name
+
+    hdf5_paths = list(raw_dir.glob("episode_*.hdf5"))
+    assert len(hdf5_paths) != 0
+    for hdf5_path in hdf5_paths:
+        with h5py.File(hdf5_path, "r") as data:
+            assert "/action" in data
+            assert "/observations/qpos" in data
+
+            assert data["/action"].ndim == 2
+            assert data["/observations/qpos"].ndim == 2
+
+            num_frames = data["/action"].shape[0]
+            assert num_frames == data["/observations/qpos"].shape[0]
+
+            for camera in get_cameras(data):
+                assert num_frames == data[f"/observations/images/{camera}"].shape[0]
+
+                if compressed_images:
+                    assert data[f"/observations/images/{camera}"].ndim == 2
+                else:
+                    assert data[f"/observations/images/{camera}"].ndim == 4
+                    b, h, w, c = data[f"/observations/images/{camera}"].shape
+                    assert c < h and c < w, f"Expect (h,w,c) image format but ({h=},{w=},{c=}) provided."
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # only frames from simulation are uncompressed
+    compressed_images = "sim" not in raw_dir.name
+
+    hdf5_files = sorted(raw_dir.glob("episode_*.hdf5"))
+    num_episodes = len(hdf5_files)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx in tqdm.tqdm(ep_ids):
+        ep_path = hdf5_files[ep_idx]
+        with h5py.File(ep_path, "r") as ep:
+            num_frames = ep["/action"].shape[0]
+
+            # last step of demonstration is considered done
+            done = torch.zeros(num_frames, dtype=torch.bool)
+            done[-1] = True
+
+            state = torch.from_numpy(ep["/observations/qpos"][:])
+            action = torch.from_numpy(ep["/action"][:])
+            if "/observations/qvel" in ep:
+                velocity = torch.from_numpy(ep["/observations/qvel"][:])
+            if "/observations/effort" in ep:
+                effort = torch.from_numpy(ep["/observations/effort"][:])
+
+            ep_dict = {}
+
+            for camera in get_cameras(ep):
+                img_key = f"observation.images.{camera}"
+
+                if compressed_images:
+                    import cv2
+
+                    # load one compressed image after the other in RAM and uncompress
+                    imgs_array = []
+                    for data in ep[f"/observations/images/{camera}"]:
+                        imgs_array.append(cv2.imdecode(data, 1))
+                    imgs_array = np.array(imgs_array)
+
+                else:
+                    # load all images in RAM
+                    imgs_array = ep[f"/observations/images/{camera}"][:]
+
+                if video:
+                    # save png images in temporary directory
+                    tmp_imgs_dir = videos_dir / "tmp_images"
+                    save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                    # encode images to a mp4 video
+                    fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                    video_path = videos_dir / fname
+                    encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                    # clean temporary images directory
+                    shutil.rmtree(tmp_imgs_dir)
+
+                    # store the reference to the video frame
+                    ep_dict[img_key] = [
+                        {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                    ]
+                else:
+                    ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+            ep_dict["observation.state"] = state
+            if "/observations/velocity" in ep:
+                ep_dict["observation.velocity"] = velocity
+            if "/observations/effort" in ep:
+                ep_dict["observation.effort"] = effort
+            ep_dict["action"] = action
+            ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+            ep_dict["next.done"] = done
+            # TODO(rcadene): add reward and success by computing them in sim
+
+            assert isinstance(ep_idx, int)
+            ep_dicts.append(ep_dict)
+
+        gc.collect()
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        if video:
+            features[key] = VideoFrame()
+        else:
+            features[key] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if "observation.velocity" in data_dict:
+        features["observation.velocity"] = Sequence(
+            length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    if "observation.effort" in data_dict:
+        features["observation.effort"] = Sequence(
+            length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 50
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/aloha_processor.py

@@ -1,199 +0,0 @@
-import re
-from pathlib import Path
-
-import h5py
-import torch
-import tqdm
-from datasets import Dataset, Features, Image, Sequence, Value
-from PIL import Image as PILImage
-
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes
-from lerobot.common.datasets.utils import (
-    hf_transform_to_torch,
-)
-
-
-class AlohaProcessor:
-    """
-    Process HDF5 files formatted like in: https://github.com/tonyzhaozh/act
-
-    Attributes:
-        folder_path (Path): Path to the directory containing HDF5 files.
-        cameras (list[str]): List of camera identifiers to check in the files.
-        fps (int): Frames per second used in timestamp calculations.
-
-    Methods:
-        is_valid() -> bool:
-            Validates if each HDF5 file within the folder contains all required datasets.
-        preprocess() -> dict:
-            Processes the files and returns structured data suitable for further analysis.
-        to_hf_dataset(data_dict: dict) -> Dataset:
-            Converts processed data into a Hugging Face Dataset object.
-    """
-
-    def __init__(self, folder_path: Path, cameras: list[str] | None = None, fps: int | None = None):
-        """
-        Initializes the AlohaProcessor with a specified directory path containing HDF5 files,
-        an optional list of cameras, and a frame rate.
-
-        Args:
-            folder_path (Path): The directory path where HDF5 files are stored.
-            cameras (list[str] | None): Optional list of cameras to validate within the files. Defaults to ['top'] if None.
-            fps (int): Frame rate for the datasets, used in time calculations. Default is 50.
-
-        Examples:
-            >>> processor = AlohaProcessor(Path("path_to_hdf5_directory"), ["camera1", "camera2"])
-            >>> processor.is_valid()
-            True
-        """
-        self.folder_path = folder_path
-        if cameras is None:
-            cameras = ["top"]
-        self.cameras = cameras
-        if fps is None:
-            fps = 50
-        self._fps = fps
-
-    @property
-    def fps(self) -> int:
-        return self._fps
-
-    def is_valid(self) -> bool:
-        """
-        Validates the HDF5 files in the specified folder to ensure they contain the required datasets
-        for actions, positions, and images for each specified camera.
-
-        Returns:
-            bool: True if all files are valid HDF5 files with all required datasets, False otherwise.
-        """
-        hdf5_files: list[Path] = list(self.folder_path.glob("episode_*.hdf5"))
-        if len(hdf5_files) == 0:
-            return False
-        try:
-            hdf5_files = sorted(
-                hdf5_files, key=lambda x: int(re.search(r"episode_(\d+).hdf5", x.name).group(1))
-            )
-        except AttributeError:
-            # All file names must contain a numerical identifier matching 'episode_(\\d+).hdf5
-            return False
-
-        # Check if the sequence is consecutive eg episode_0, episode_1, episode_2, etc.
-        # If not, return False
-        previous_number = None
-        for file in hdf5_files:
-            current_number = int(re.search(r"episode_(\d+).hdf5", file.name).group(1))
-            if previous_number is not None and current_number - previous_number != 1:
-                return False
-            previous_number = current_number
-
-        for file in hdf5_files:
-            try:
-                with h5py.File(file, "r") as file:
-                    # Check for the expected datasets within the HDF5 file
-                    required_datasets = ["/action", "/observations/qpos"]
-                    # Add camera-specific image datasets to the required datasets
-                    camera_datasets = [f"/observations/images/{cam}" for cam in self.cameras]
-                    required_datasets.extend(camera_datasets)
-
-                    if not all(dataset in file for dataset in required_datasets):
-                        return False
-            except OSError:
-                return False
-        return True
-
-    def preprocess(self):
-        """
-        Collects episode data from the HDF5 file and returns it as an AlohaStep named tuple.
-
-        Returns:
-            AlohaStep: Named tuple containing episode data.
-
-        Raises:
-            ValueError: If the file is not valid.
-        """
-        if not self.is_valid():
-            raise ValueError("The HDF5 file is invalid or does not contain the required datasets.")
-
-        hdf5_files = list(self.folder_path.glob("*.hdf5"))
-        hdf5_files = sorted(hdf5_files, key=lambda x: int(re.search(r"episode_(\d+)", x.name).group(1)))
-        ep_dicts = []
-        episode_data_index = {"from": [], "to": []}
-
-        id_from = 0
-
-        for ep_path in tqdm.tqdm(hdf5_files):
-            with h5py.File(ep_path, "r") as ep:
-                ep_id = int(re.search(r"episode_(\d+)", ep_path.name).group(1))
-                num_frames = ep["/action"].shape[0]
-
-                # last step of demonstration is considered done
-                done = torch.zeros(num_frames, dtype=torch.bool)
-                done[-1] = True
-
-                state = torch.from_numpy(ep["/observations/qpos"][:])
-                action = torch.from_numpy(ep["/action"][:])
-
-                ep_dict = {}
-
-                for cam in self.cameras:
-                    image = torch.from_numpy(ep[f"/observations/images/{cam}"][:])  # b h w c
-                    ep_dict[f"observation.images.{cam}"] = [PILImage.fromarray(x.numpy()) for x in image]
-
-                ep_dict.update(
-                    {
-                        "observation.state": state,
-                        "action": action,
-                        "episode_index": torch.tensor([ep_id] * num_frames),
-                        "frame_index": torch.arange(0, num_frames, 1),
-                        "timestamp": torch.arange(0, num_frames, 1) / self.fps,
-                        # TODO(rcadene): compute reward and success
-                        # "next.reward": reward,
-                        "next.done": done,
-                        # "next.success": success,
-                    }
-                )
-
-                assert isinstance(ep_id, int)
-                ep_dicts.append(ep_dict)
-
-                episode_data_index["from"].append(id_from)
-                episode_data_index["to"].append(id_from + num_frames)
-
-            id_from += num_frames
-
-        data_dict = concatenate_episodes(ep_dicts)
-        return data_dict, episode_data_index
-
-    def to_hf_dataset(self, data_dict) -> Dataset:
-        """
-        Converts a dictionary of data into a Hugging Face Dataset object.
-
-        Args:
-            data_dict (dict): A dictionary containing the data to be converted.
-
-        Returns:
-            Dataset: The converted Hugging Face Dataset object.
-        """
-        image_features = {f"observation.images.{cam}": Image() for cam in self.cameras}
-        features = {
-            "observation.state": Sequence(
-                length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
-            ),
-            "action": Sequence(length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)),
-            "episode_index": Value(dtype="int64", id=None),
-            "frame_index": Value(dtype="int64", id=None),
-            "timestamp": Value(dtype="float32", id=None),
-            # "next.reward": Value(dtype="float32", id=None),
-            "next.done": Value(dtype="bool", id=None),
-            # "next.success": Value(dtype="bool", id=None),
-            "index": Value(dtype="int64", id=None),
-        }
-        update_features = {**image_features, **features}
-        features = Features(update_features)
-        hf_dataset = Dataset.from_dict(data_dict, features=features)
-        hf_dataset.set_transform(hf_transform_to_torch)
-
-        return hf_dataset
-
-    def cleanup(self):
-        pass

lerobot/common/datasets/push_dataset_to_hub/cam_png_format.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Contains utilities to process raw data format of png images files recorded with capture_camera_feed.py
+"""
+
+from pathlib import Path
+
+import torch
+from datasets import Dataset, Features, Image, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+)
+from lerobot.common.datasets.utils import hf_transform_to_torch
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def check_format(raw_dir: Path) -> bool:
+    image_paths = list(raw_dir.glob("frame_*.png"))
+    if len(image_paths) == 0:
+        raise ValueError
+
+
+def load_from_raw(raw_dir: Path, fps: int, episodes: list[int] | None = None):
+    if episodes is not None:
+        # TODO(aliberts): add support for multi-episodes.
+        raise NotImplementedError()
+
+    ep_dict = {}
+    ep_idx = 0
+
+    image_paths = sorted(raw_dir.glob("frame_*.png"))
+    num_frames = len(image_paths)
+
+    ep_dict["observation.image"] = [PILImage.open(x) for x in image_paths]
+    ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+    ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+    ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+
+    ep_dicts = [ep_dict]
+    data_dict = concatenate_episodes(ep_dicts)
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    if video or episodes or encoding is not None:
+        # TODO(aliberts): support this
+        raise NotImplementedError
+
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/dora_parquet_format.py

@@ -0,0 +1,233 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Contains utilities to process raw data format from dora-record
+"""
+
+import re
+import warnings
+from pathlib import Path
+
+import pandas as pd
+import torch
+from datasets import Dataset, Features, Image, Sequence, Value
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import calculate_episode_data_index
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def check_format(raw_dir) -> bool:
+    assert raw_dir.exists()
+
+    leader_file = list(raw_dir.glob("*.parquet"))
+    if len(leader_file) == 0:
+        raise ValueError(f"Missing parquet files in '{raw_dir}'")
+    return True
+
+
+def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episodes: list[int] | None = None):
+    # Load data stream that will be used as reference for the timestamps synchronization
+    reference_files = list(raw_dir.glob("observation.images.cam_*.parquet"))
+    if len(reference_files) == 0:
+        raise ValueError(f"Missing reference files for camera, starting with  in '{raw_dir}'")
+    # select first camera in alphanumeric order
+    reference_key = sorted(reference_files)[0].stem
+    reference_df = pd.read_parquet(raw_dir / f"{reference_key}.parquet")
+    reference_df = reference_df[["timestamp_utc", reference_key]]
+
+    # Merge all data stream using nearest backward strategy
+    df = reference_df
+    for path in raw_dir.glob("*.parquet"):
+        key = path.stem  # action or observation.state or ...
+        if key == reference_key:
+            continue
+        if "failed_episode_index" in key:
+            # TODO(rcadene): add support for removing episodes that are tagged as "failed"
+            continue
+        modality_df = pd.read_parquet(path)
+        modality_df = modality_df[["timestamp_utc", key]]
+        df = pd.merge_asof(
+            df,
+            modality_df,
+            on="timestamp_utc",
+            # "nearest" is the best option over "backward", since the latter can desynchronizes camera timestamps by
+            # matching timestamps that are too far appart, in order to fit the backward constraints. It's not the case for "nearest".
+            # However, note that "nearest" might synchronize the reference camera with other cameras on slightly future timestamps.
+            # are too far appart.
+            direction="nearest",
+            tolerance=pd.Timedelta(f"{1/fps} seconds"),
+        )
+    # Remove rows with episode_index -1 which indicates data that correspond to in-between episodes
+    df = df[df["episode_index"] != -1]
+
+    image_keys = [key for key in df if "observation.images." in key]
+
+    def get_episode_index(row):
+        episode_index_per_cam = {}
+        for key in image_keys:
+            path = row[key][0]["path"]
+            match = re.search(r"_(\d{6}).mp4", path)
+            if not match:
+                raise ValueError(path)
+            episode_index = int(match.group(1))
+            episode_index_per_cam[key] = episode_index
+        if len(set(episode_index_per_cam.values())) != 1:
+            raise ValueError(
+                f"All cameras are expected to belong to the same episode, but getting {episode_index_per_cam}"
+            )
+        return episode_index
+
+    df["episode_index"] = df.apply(get_episode_index, axis=1)
+
+    # dora only use arrays, so single values are encapsulated into a list
+    df["frame_index"] = df.groupby("episode_index").cumcount()
+    df = df.reset_index()
+    df["index"] = df.index
+
+    # set 'next.done' to True for the last frame of each episode
+    df["next.done"] = False
+    df.loc[df.groupby("episode_index").tail(1).index, "next.done"] = True
+
+    df["timestamp"] = df["timestamp_utc"].map(lambda x: x.timestamp())
+    # each episode starts with timestamp 0 to match the ones from the video
+    df["timestamp"] = df.groupby("episode_index")["timestamp"].transform(lambda x: x - x.iloc[0])
+
+    del df["timestamp_utc"]
+
+    # sanity check
+    has_nan = df.isna().any().any()
+    if has_nan:
+        raise ValueError("Dataset contains Nan values.")
+
+    # sanity check episode indices go from 0 to n-1
+    ep_ids = [ep_idx for ep_idx, _ in df.groupby("episode_index")]
+    expected_ep_ids = list(range(df["episode_index"].max() + 1))
+    if ep_ids != expected_ep_ids:
+        raise ValueError(f"Episodes indices go from {ep_ids} instead of {expected_ep_ids}")
+
+    # Create symlink to raw videos directory (that needs to be absolute not relative)
+    videos_dir.parent.mkdir(parents=True, exist_ok=True)
+    videos_dir.symlink_to((raw_dir / "videos").absolute())
+
+    # sanity check the video paths are well formated
+    for key in df:
+        if "observation.images." not in key:
+            continue
+        for ep_idx in ep_ids:
+            video_path = videos_dir / f"{key}_episode_{ep_idx:06d}.mp4"
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
+
+    data_dict = {}
+    for key in df:
+        # is video frame
+        if "observation.images." in key:
+            # we need `[0] because dora only use arrays, so single values are encapsulated into a list.
+            # it is the case for video_frame dictionary = [{"path": ..., "timestamp": ...}]
+            data_dict[key] = [video_frame[0] for video_frame in df[key].values]
+
+            # sanity check the video path is well formated
+            video_path = videos_dir.parent / data_dict[key][0]["path"]
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
+        # is number
+        elif df[key].iloc[0].ndim == 0 or df[key].iloc[0].shape[0] == 1:
+            data_dict[key] = torch.from_numpy(df[key].values)
+        # is vector
+        elif df[key].iloc[0].shape[0] > 1:
+            data_dict[key] = torch.stack([torch.from_numpy(x.copy()) for x in df[key].values])
+        else:
+            raise ValueError(key)
+
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        if video:
+            features[key] = VideoFrame()
+        else:
+            features[key] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if "observation.velocity" in data_dict:
+        features["observation.velocity"] = Sequence(
+            length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    if "observation.effort" in data_dict:
+        features["observation.effort"] = Sequence(
+            length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+    else:
+        raise NotImplementedError()
+
+    if not video:
+        raise NotImplementedError()
+
+    if encoding is not None:
+        warnings.warn(
+            "Video encoding is currently done outside of LeRobot for the dora_parquet format.",
+            stacklevel=1,
+        )
+
+    data_df = load_from_raw(raw_dir, videos_dir, fps, episodes)
+    hf_dataset = to_hf_dataset(data_df, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = "unknown"
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/openx_rlds_format.py

@@ -0,0 +1,312 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+For all datasets in the RLDS format.
+For https://github.com/google-deepmind/open_x_embodiment (OPENX) datasets.
+
+NOTE: You need to install tensorflow and tensorflow_datsets before running this script.
+
+Example:
+    python lerobot/scripts/push_dataset_to_hub.py \
+        --raw-dir /path/to/data/bridge_dataset/1.0.0/ \
+        --repo-id your_hub/sampled_bridge_data_v2 \
+        --raw-format rlds \
+        --episodes 3 4 5 8 9
+
+Exact dataset fps defined in openx/config.py, obtained from:
+    https://docs.google.com/spreadsheets/d/1rPBD77tk60AEIGZrGSODwyyzs5FgCU9Uz3h-3_t2A9g/edit?gid=0#gid=0&range=R:R
+"""
+
+import shutil
+from pathlib import Path
+
+import numpy as np
+import tensorflow as tf
+import tensorflow_datasets as tfds
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+np.set_printoptions(precision=2)
+
+
+def tf_to_torch(data):
+    return torch.from_numpy(data.numpy())
+
+
+def tf_img_convert(img):
+    if img.dtype == tf.string:
+        img = tf.io.decode_image(img, expand_animations=False, dtype=tf.uint8)
+    elif img.dtype != tf.uint8:
+        raise ValueError(f"Unsupported image dtype: found with dtype {img.dtype}")
+    return img.numpy()
+
+
+def _broadcast_metadata_rlds(i: tf.Tensor, traj: dict) -> dict:
+    """
+    In the RLDS format, each trajectory has some top-level metadata that is explicitly separated out, and a "steps"
+    entry. This function moves the "steps" entry to the top level, broadcasting any metadata to the length of the
+    trajectory. This function also adds the extra metadata fields `_len`, `_traj_index`, and `_frame_index`.
+
+    NOTE: adapted from DLimp library https://github.com/kvablack/dlimp/
+    """
+    steps = traj.pop("steps")
+
+    traj_len = tf.shape(tf.nest.flatten(steps)[0])[0]
+
+    # broadcast metadata to the length of the trajectory
+    metadata = tf.nest.map_structure(lambda x: tf.repeat(x, traj_len), traj)
+
+    # put steps back in
+    assert "traj_metadata" not in steps
+    traj = {**steps, "traj_metadata": metadata}
+
+    assert "_len" not in traj
+    assert "_traj_index" not in traj
+    assert "_frame_index" not in traj
+    traj["_len"] = tf.repeat(traj_len, traj_len)
+    traj["_traj_index"] = tf.repeat(i, traj_len)
+    traj["_frame_index"] = tf.range(traj_len)
+
+    return traj
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    """
+    Args:
+        raw_dir (Path): _description_
+        videos_dir (Path): _description_
+        fps (int): _description_
+        video (bool): _description_
+        episodes (list[int] | None, optional): _description_. Defaults to None.
+    """
+    ds_builder = tfds.builder_from_directory(str(raw_dir))
+    dataset = ds_builder.as_dataset(
+        split="all",
+        decoders={"steps": tfds.decode.SkipDecoding()},
+    )
+
+    dataset_info = ds_builder.info
+    print("dataset_info: ", dataset_info)
+
+    ds_length = len(dataset)
+    dataset = dataset.take(ds_length)
+    # "flatten" the dataset as such we can apply trajectory level map() easily
+    # each [obs][key] has a shape of (frame_size, ...)
+    dataset = dataset.enumerate().map(_broadcast_metadata_rlds)
+
+    # we will apply the standardization transform if the dataset_name is provided
+    # if the dataset name is not provided and the goal is to convert any rlds formatted dataset
+    # search for 'image' keys in the observations
+    image_keys = []
+    state_keys = []
+    observation_info = dataset_info.features["steps"]["observation"]
+    for key in observation_info:
+        # check whether the key is for an image or a vector observation
+        if len(observation_info[key].shape) == 3:
+            # only adding uint8 images discards depth images
+            if observation_info[key].dtype == tf.uint8:
+                image_keys.append(key)
+        else:
+            state_keys.append(key)
+
+    lang_key = "language_instruction" if "language_instruction" in dataset.element_spec else None
+
+    print(" - image_keys: ", image_keys)
+    print(" - lang_key: ", lang_key)
+
+    it = iter(dataset)
+
+    ep_dicts = []
+    # Init temp path to save ep_dicts in case of crash
+    tmp_ep_dicts_dir = videos_dir.parent.joinpath("ep_dicts")
+    tmp_ep_dicts_dir.mkdir(parents=True, exist_ok=True)
+
+    # check if ep_dicts have already been saved in /tmp
+    starting_ep_idx = 0
+    saved_ep_dicts = [ep.__str__() for ep in tmp_ep_dicts_dir.iterdir()]
+    if len(saved_ep_dicts) > 0:
+        saved_ep_dicts.sort()
+        # get last ep_idx number
+        starting_ep_idx = int(saved_ep_dicts[-1][-13:-3]) + 1
+        for i in range(starting_ep_idx):
+            episode = next(it)
+            ep_dicts.append(torch.load(saved_ep_dicts[i]))
+
+    # if we user specified episodes, skip the ones not in the list
+    if episodes is not None:
+        if ds_length == 0:
+            raise ValueError("No episodes found.")
+        # convert episodes index to sorted list
+        episodes = sorted(episodes)
+
+    for ep_idx in tqdm.tqdm(range(starting_ep_idx, ds_length)):
+        episode = next(it)
+
+        # if user specified episodes, skip the ones not in the list
+        if episodes is not None:
+            if len(episodes) == 0:
+                break
+            if ep_idx == episodes[0]:
+                # process this episode
+                print(" selecting episode idx: ", ep_idx)
+                episodes.pop(0)
+            else:
+                continue  # skip
+
+        num_frames = episode["action"].shape[0]
+
+        ep_dict = {}
+        for key in state_keys:
+            ep_dict[f"observation.{key}"] = tf_to_torch(episode["observation"][key])
+
+        ep_dict["action"] = tf_to_torch(episode["action"])
+        ep_dict["next.reward"] = tf_to_torch(episode["reward"]).float()
+        ep_dict["next.done"] = tf_to_torch(episode["is_last"])
+        ep_dict["is_terminal"] = tf_to_torch(episode["is_terminal"])
+        ep_dict["is_first"] = tf_to_torch(episode["is_first"])
+        ep_dict["discount"] = tf_to_torch(episode["discount"])
+
+        # If lang_key is present, convert the entire tensor at once
+        if lang_key is not None:
+            ep_dict["language_instruction"] = [x.numpy().decode("utf-8") for x in episode[lang_key]]
+
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+
+        image_array_dict = {key: [] for key in image_keys}
+
+        for im_key in image_keys:
+            imgs = episode["observation"][im_key]
+            image_array_dict[im_key] = [tf_img_convert(img) for img in imgs]
+
+        # loop through all cameras
+        for im_key in image_keys:
+            img_key = f"observation.images.{im_key}"
+            imgs_array = image_array_dict[im_key]
+            imgs_array = np.array(imgs_array)
+            if video:
+                # save png images in temporary directory
+                tmp_imgs_dir = videos_dir / "tmp_images"
+                save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                # encode images to a mp4 video
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                # clean temporary images directory
+                shutil.rmtree(tmp_imgs_dir)
+
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+            else:
+                ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        path_ep_dict = tmp_ep_dicts_dir.joinpath(
+            "ep_dict_" + "0" * (10 - len(str(ep_idx))) + str(ep_idx) + ".pt"
+        )
+        torch.save(ep_dict, path_ep_dict)
+
+        ep_dicts.append(ep_dict)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    for key in data_dict:
+        # check if vector state obs
+        if key.startswith("observation.") and "observation.images." not in key:
+            features[key] = Sequence(length=data_dict[key].shape[1], feature=Value(dtype="float32", id=None))
+        # check if image obs
+        elif "observation.images." in key:
+            if video:
+                features[key] = VideoFrame()
+            else:
+                features[key] = Image()
+
+    if "language_instruction" in data_dict:
+        features["language_instruction"] = Value(dtype="string", id=None)
+
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+
+    features["is_terminal"] = Value(dtype="bool", id=None)
+    features["is_first"] = Value(dtype="bool", id=None)
+    features["discount"] = Value(dtype="float32", id=None)
+
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/pusht_processor.py

@@ -1,180 +0,0 @@
-from pathlib import Path
-
-import numpy as np
-import torch
-import tqdm
-import zarr
-from datasets import Dataset, Features, Image, Sequence, Value
-from PIL import Image as PILImage
-
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes
-from lerobot.common.datasets.utils import (
-    hf_transform_to_torch,
-)
-
-
-class PushTProcessor:
-    """ Process zarr files formatted like in: https://github.com/real-stanford/diffusion_policy
-    """
-    def __init__(self, folder_path: Path, fps: int | None = None):
-        self.zarr_path = folder_path
-        if fps is None:
-            fps = 10
-        self._fps = fps
-
-    @property
-    def fps(self) -> int:
-        return self._fps
-
-    def is_valid(self):
-        try:
-            zarr_data = zarr.open(self.zarr_path, mode="r")
-        except Exception:
-            # TODO (azouitine): Handle the exception properly
-            return False
-        required_datasets = {
-            "data/action",
-            "data/img",
-            "data/keypoint",
-            "data/n_contacts",
-            "data/state",
-            "meta/episode_ends",
-        }
-        for dataset in required_datasets:
-            if dataset not in zarr_data:
-                return False
-        nb_frames = zarr_data["data/img"].shape[0]
-
-        required_datasets.remove("meta/episode_ends")
-
-        return all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
-
-    def preprocess(self):
-        try:
-            import pymunk
-            from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
-
-            from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
-                ReplayBuffer as DiffusionPolicyReplayBuffer,
-            )
-        except ModuleNotFoundError as e:
-            print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
-            raise e
-
-        # as define in env
-        success_threshold = 0.95  # 95% coverage,
-
-        dataset_dict = DiffusionPolicyReplayBuffer.copy_from_path(
-            self.zarr_path
-        )  # , keys=['img', 'state', 'action'])
-
-        episode_ids = torch.from_numpy(dataset_dict.get_episode_idxs())
-        num_episodes = dataset_dict.meta["episode_ends"].shape[0]
-        assert len(
-            {dataset_dict[key].shape[0] for key in dataset_dict.keys()}  # noqa: SIM118
-        ), "Some data type dont have the same number of total frames."
-
-        # TODO: verify that goal pose is expected to be fixed
-        goal_pos_angle = np.array([256, 256, np.pi / 4])  # x, y, theta (in radians)
-        goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
-
-        imgs = torch.from_numpy(dataset_dict["img"])  # b h w c
-        states = torch.from_numpy(dataset_dict["state"])
-        actions = torch.from_numpy(dataset_dict["action"])
-
-        ep_dicts = []
-        episode_data_index = {"from": [], "to": []}
-
-        id_from = 0
-        for episode_id in tqdm.tqdm(range(num_episodes)):
-            id_to = dataset_dict.meta["episode_ends"][episode_id]
-
-            num_frames = id_to - id_from
-
-            assert (episode_ids[id_from:id_to] == episode_id).all()
-
-            image = imgs[id_from:id_to]
-            assert image.min() >= 0.0
-            assert image.max() <= 255.0
-            image = image.type(torch.uint8)
-
-            state = states[id_from:id_to]
-            agent_pos = state[:, :2]
-            block_pos = state[:, 2:4]
-            block_angle = state[:, 4]
-
-            reward = torch.zeros(num_frames)
-            success = torch.zeros(num_frames, dtype=torch.bool)
-            done = torch.zeros(num_frames, dtype=torch.bool)
-            for i in range(num_frames):
-                space = pymunk.Space()
-                space.gravity = 0, 0
-                space.damping = 0
-
-                # Add walls.
-                walls = [
-                    PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
-                    PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
-                    PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
-                    PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
-                ]
-                space.add(*walls)
-
-                block_body = PushTEnv.add_tee(space, block_pos[i].tolist(), block_angle[i].item())
-                goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
-                block_geom = pymunk_to_shapely(block_body, block_body.shapes)
-                intersection_area = goal_geom.intersection(block_geom).area
-                goal_area = goal_geom.area
-                coverage = intersection_area / goal_area
-                reward[i] = np.clip(coverage / success_threshold, 0, 1)
-                success[i] = coverage > success_threshold
-
-            # last step of demonstration is considered done
-            done[-1] = True
-
-            ep_dict = {
-                "observation.image": [PILImage.fromarray(x.numpy()) for x in image],
-                "observation.state": agent_pos,
-                "action": actions[id_from:id_to],
-                "episode_index": torch.tensor([episode_id] * num_frames, dtype=torch.int),
-                "frame_index": torch.arange(0, num_frames, 1),
-                "timestamp": torch.arange(0, num_frames, 1) / self.fps,
-                # "next.observation.image": image[1:],
-                # "next.observation.state": agent_pos[1:],
-                # TODO(rcadene): verify that reward and done are aligned with image and agent_pos
-                "next.reward": torch.cat([reward[1:], reward[[-1]]]),
-                "next.done": torch.cat([done[1:], done[[-1]]]),
-                "next.success": torch.cat([success[1:], success[[-1]]]),
-            }
-            ep_dicts.append(ep_dict)
-
-            episode_data_index["from"].append(id_from)
-            episode_data_index["to"].append(id_from + num_frames)
-
-            id_from += num_frames
-
-        data_dict = concatenate_episodes(ep_dicts)
-        return data_dict, episode_data_index
-
-    def to_hf_dataset(self, data_dict):
-        features = {
-            "observation.image": Image(),
-            "observation.state": Sequence(
-                length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
-            ),
-            "action": Sequence(length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)),
-            "episode_index": Value(dtype="int64", id=None),
-            "frame_index": Value(dtype="int64", id=None),
-            "timestamp": Value(dtype="float32", id=None),
-            "next.reward": Value(dtype="float32", id=None),
-            "next.done": Value(dtype="bool", id=None),
-            "next.success": Value(dtype="bool", id=None),
-            "index": Value(dtype="int64", id=None),
-        }
-        features = Features(features)
-        hf_dataset = Dataset.from_dict(data_dict, features=features)
-        hf_dataset.set_transform(hf_transform_to_torch)
-        return hf_dataset
-
-    def cleanup(self):
-        pass

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -0,0 +1,275 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Process zarr files formatted like in: https://github.com/real-stanford/diffusion_policy"""
+
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir):
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    required_datasets = {
+        "data/action",
+        "data/img",
+        "data/keypoint",
+        "data/n_contacts",
+        "data/state",
+        "meta/episode_ends",
+    }
+    for dataset in required_datasets:
+        assert dataset in zarr_data
+    nb_frames = zarr_data["data/img"].shape[0]
+
+    required_datasets.remove("meta/episode_ends")
+
+    assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    keypoints_instead_of_image: bool = False,
+    encoding: dict | None = None,
+):
+    try:
+        import pymunk
+        from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
+
+        from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
+            ReplayBuffer as DiffusionPolicyReplayBuffer,
+        )
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+    # as define in gmy-pusht env: https://github.com/huggingface/gym-pusht/blob/e0684ff988d223808c0a9dcfaba9dc4991791370/gym_pusht/envs/pusht.py#L174
+    success_threshold = 0.95  # 95% coverage,
+
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
+
+    episode_ids = torch.from_numpy(zarr_data.get_episode_idxs())
+    assert len(
+        {zarr_data[key].shape[0] for key in zarr_data.keys()}  # noqa: SIM118
+    ), "Some data type dont have the same number of total frames."
+
+    # TODO(rcadene): verify that goal pose is expected to be fixed
+    goal_pos_angle = np.array([256, 256, np.pi / 4])  # x, y, theta (in radians)
+    goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
+
+    imgs = torch.from_numpy(zarr_data["img"])  # b h w c
+    states = torch.from_numpy(zarr_data["state"])
+    actions = torch.from_numpy(zarr_data["action"])
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx = 0
+    for to_idx in zarr_data.meta["episode_ends"]:
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    num_episodes = len(from_ids)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        from_idx = from_ids[selected_ep_idx]
+        to_idx = to_ids[selected_ep_idx]
+        num_frames = to_idx - from_idx
+
+        # sanity check
+        assert (episode_ids[from_idx:to_idx] == ep_idx).all()
+
+        # get image
+        if not keypoints_instead_of_image:
+            image = imgs[from_idx:to_idx]
+            assert image.min() >= 0.0
+            assert image.max() <= 255.0
+            image = image.type(torch.uint8)
+
+        # get state
+        state = states[from_idx:to_idx]
+        agent_pos = state[:, :2]
+        block_pos = state[:, 2:4]
+        block_angle = state[:, 4]
+
+        # get reward, success, done, and (maybe) keypoints
+        reward = torch.zeros(num_frames)
+        success = torch.zeros(num_frames, dtype=torch.bool)
+        if keypoints_instead_of_image:
+            keypoints = torch.zeros(num_frames, 16)  # 8 keypoints each with 2 coords
+        done = torch.zeros(num_frames, dtype=torch.bool)
+        for i in range(num_frames):
+            space = pymunk.Space()
+            space.gravity = 0, 0
+            space.damping = 0
+
+            # Add walls.
+            walls = [
+                PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
+                PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
+                PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
+                PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
+            ]
+            space.add(*walls)
+
+            block_body, block_shapes = PushTEnv.add_tee(space, block_pos[i].tolist(), block_angle[i].item())
+            goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
+            block_geom = pymunk_to_shapely(block_body, block_body.shapes)
+            intersection_area = goal_geom.intersection(block_geom).area
+            goal_area = goal_geom.area
+            coverage = intersection_area / goal_area
+            reward[i] = np.clip(coverage / success_threshold, 0, 1)
+            success[i] = coverage > success_threshold
+            if keypoints_instead_of_image:
+                keypoints[i] = torch.from_numpy(PushTEnv.get_keypoints(block_shapes).flatten())
+
+        # last step of demonstration is considered done
+        done[-1] = True
+
+        ep_dict = {}
+
+        if not keypoints_instead_of_image:
+            imgs_array = [x.numpy() for x in image]
+            img_key = "observation.image"
+            if video:
+                # save png images in temporary directory
+                tmp_imgs_dir = videos_dir / "tmp_images"
+                save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                # encode images to a mp4 video
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                # clean temporary images directory
+                shutil.rmtree(tmp_imgs_dir)
+
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+            else:
+                ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = agent_pos
+        if keypoints_instead_of_image:
+            ep_dict["observation.environment_state"] = keypoints
+        ep_dict["action"] = actions[from_idx:to_idx]
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        # ep_dict["next.observation.image"] = image[1:],
+        # ep_dict["next.observation.state"] = agent_pos[1:],
+        # TODO(rcadene)] = verify that reward and done are aligned with image and agent_pos
+        ep_dict["next.reward"] = torch.cat([reward[1:], reward[[-1]]])
+        ep_dict["next.done"] = torch.cat([done[1:], done[[-1]]])
+        ep_dict["next.success"] = torch.cat([success[1:], success[[-1]]])
+        ep_dicts.append(ep_dict)
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video, keypoints_instead_of_image: bool = False):
+    features = {}
+
+    if not keypoints_instead_of_image:
+        if video:
+            features["observation.image"] = VideoFrame()
+        else:
+            features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    if keypoints_instead_of_image:
+        features["observation.environment_state"] = Sequence(
+            length=data_dict["observation.environment_state"].shape[1],
+            feature=Value(dtype="float32", id=None),
+        )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["next.success"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # Manually change this to True to use keypoints of the T instead of an image observation (but don't merge
+    # with True). Also make sure to use video = 0 in the `push_dataset_to_hub.py` script.
+    keypoints_instead_of_image = False
+
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 10
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, keypoints_instead_of_image, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video, keypoints_instead_of_image)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video if not keypoints_instead_of_image else 0,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/umi_processor.py

@@ -1,280 +0,0 @@
-import os
-import re
-import shutil
-from glob import glob
-
-import numpy as np
-import torch
-import tqdm
-import zarr
-from datasets import Dataset, Features, Image, Sequence, Value
-from PIL import Image as PILImage
-
-from lerobot.common.datasets.push_dataset_to_hub._umi_imagecodecs_numcodecs import register_codecs
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes
-from lerobot.common.datasets.utils import (
-    hf_transform_to_torch,
-)
-
-
-class UmiProcessor:
-    """
-     Process UMI (Universal Manipulation Interface) data stored in Zarr format like in: https://github.com/real-stanford/universal_manipulation_interface
-
-    Attributes:
-        folder_path (str): The path to the folder containing Zarr datasets.
-        fps (int): Frames per second, used to calculate timestamps for frames.
-
-    """
-
-    def __init__(self, folder_path: str, fps: int | None = None):
-        self.zarr_path = folder_path
-        if fps is None:
-            # TODO (azouitine): Add reference to the paper
-            fps = 15
-        self._fps = fps
-        register_codecs()
-
-    @property
-    def fps(self) -> int:
-        return self._fps
-
-    def is_valid(self) -> bool:
-        """
-        Validates the Zarr folder to ensure it contains all required datasets with consistent frame counts.
-
-        Returns:
-            bool: True if all required datasets are present and have consistent frame counts, False otherwise.
-        """
-        # Check if the Zarr folder is valid
-        try:
-            zarr_data = zarr.open(self.zarr_path, mode="r")
-        except Exception:
-            # TODO (azouitine): Handle the exception properly
-            return False
-        required_datasets = {
-            "data/robot0_demo_end_pose",
-            "data/robot0_demo_start_pose",
-            "data/robot0_eef_pos",
-            "data/robot0_eef_rot_axis_angle",
-            "data/robot0_gripper_width",
-            "meta/episode_ends",
-            "data/camera0_rgb",
-        }
-        for dataset in required_datasets:
-            if dataset not in zarr_data:
-                return False
-        nb_frames = zarr_data["data/camera0_rgb"].shape[0]
-
-        required_datasets.remove("meta/episode_ends")
-
-        return all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
-
-    def preprocess(self):
-        """
-        Collects and processes all episodes from the Zarr dataset into structured data dictionaries.
-
-        Returns:
-            Tuple[Dict, Dict]: A tuple containing the structured episode data and episode index mappings.
-        """
-        zarr_data = zarr.open(self.zarr_path, mode="r")
-
-        # We process the image data separately because it is too large to fit in memory
-        end_pose = torch.from_numpy(zarr_data["data/robot0_demo_end_pose"][:])
-        start_pos = torch.from_numpy(zarr_data["data/robot0_demo_start_pose"][:])
-        eff_pos = torch.from_numpy(zarr_data["data/robot0_eef_pos"][:])
-        eff_rot_axis_angle = torch.from_numpy(zarr_data["data/robot0_eef_rot_axis_angle"][:])
-        gripper_width = torch.from_numpy(zarr_data["data/robot0_gripper_width"][:])
-
-        states_pos = torch.cat([eff_pos, eff_rot_axis_angle], dim=1)
-        states = torch.cat([states_pos, gripper_width], dim=1)
-
-        episode_ends = zarr_data["meta/episode_ends"][:]
-        num_episodes: int = episode_ends.shape[0]
-
-        episode_ids = torch.from_numpy(self.get_episode_idxs(episode_ends))
-
-        # We convert it in torch tensor later because the jit function does not support torch tensors
-        episode_ends = torch.from_numpy(episode_ends)
-
-        ep_dicts = []
-        episode_data_index = {"from": [], "to": []}
-        id_from = 0
-
-        for episode_id in tqdm.tqdm(range(num_episodes)):
-            id_to = episode_ends[episode_id]
-
-            num_frames = id_to - id_from
-
-            assert (
-                episode_ids[id_from:id_to] == episode_id
-            ).all(), f"episode_ids[{id_from}:{id_to}] != {episode_id}"
-
-            state = states[id_from:id_to]
-            ep_dict = {
-                # observation.image will be filled later
-                "observation.state": state,
-                "episode_index": torch.tensor([episode_id] * num_frames, dtype=torch.int),
-                "frame_index": torch.arange(0, num_frames, 1),
-                "timestamp": torch.arange(0, num_frames, 1) / self.fps,
-                "episode_data_index_from": torch.tensor([id_from] * num_frames),
-                "episode_data_index_to": torch.tensor([id_from + num_frames] * num_frames),
-                "end_pose": end_pose[id_from:id_to],
-                "start_pos": start_pos[id_from:id_to],
-                "gripper_width": gripper_width[id_from:id_to],
-            }
-            ep_dicts.append(ep_dict)
-            episode_data_index["from"].append(id_from)
-            episode_data_index["to"].append(id_from + num_frames)
-            id_from += num_frames
-
-        data_dict = concatenate_episodes(ep_dicts)
-
-        total_frames = id_from
-        data_dict["index"] = torch.arange(0, total_frames, 1)
-
-        print("Saving images to disk in temporary folder...")
-        # datasets.Image() can take a list of paths to images, so we save the images to a temporary folder
-        # to avoid loading them all in memory
-        _save_images_concurrently(
-            data=zarr_data, image_key="data/camera0_rgb", folder_path="tmp_umi_images", max_workers=12
-        )
-        print("Saving images to disk in temporary folder... Done")
-
-        # Sort files by number eg. 1.png, 2.png, 3.png, 9.png, 10.png instead of 1.png, 10.png, 2.png, 3.png, 9.png
-        # to correctly match the images with the data
-        images_path = sorted(
-            glob("tmp_umi_images/*"), key=lambda x: int(re.search(r"(\d+)\.png$", x).group(1))
-        )
-        data_dict["observation.image"] = images_path
-        print("Images saved to disk, do not forget to delete the folder tmp_umi_images/")
-
-        # Cleanup
-        return data_dict, episode_data_index
-
-    def to_hf_dataset(self, data_dict):
-        """
-        Converts the processed data dictionary into a Hugging Face dataset with defined features.
-
-        Args:
-            data_dict (Dict): The data dictionary containing tensors and episode information.
-
-        Returns:
-            Dataset: A Hugging Face dataset constructed from the provided data dictionary.
-        """
-        features = {
-            "observation.image": Image(),
-            "observation.state": Sequence(
-                length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
-            ),
-            "episode_index": Value(dtype="int64", id=None),
-            "frame_index": Value(dtype="int64", id=None),
-            "timestamp": Value(dtype="float32", id=None),
-            "index": Value(dtype="int64", id=None),
-            "episode_data_index_from": Value(dtype="int64", id=None),
-            "episode_data_index_to": Value(dtype="int64", id=None),
-            # `start_pos` and `end_pos` respectively represent the positions of the end-effector
-            # at the beginning and the end of the episode.
-            # `gripper_width` indicates the distance between the grippers, and this value is included
-            # in the state vector, which comprises the concatenation of the end-effector position
-            # and gripper width.
-            "end_pose": Sequence(
-                length=data_dict["end_pose"].shape[1], feature=Value(dtype="float32", id=None)
-            ),
-            "start_pos": Sequence(
-                length=data_dict["start_pos"].shape[1], feature=Value(dtype="float32", id=None)
-            ),
-            "gripper_width": Sequence(
-                length=data_dict["gripper_width"].shape[1], feature=Value(dtype="float32", id=None)
-            ),
-        }
-        features = Features(features)
-        hf_dataset = Dataset.from_dict(data_dict, features=features)
-        hf_dataset.set_transform(hf_transform_to_torch)
-
-        return hf_dataset
-
-    def cleanup(self):
-        # Cleanup
-        if os.path.exists("tmp_umi_images"):
-            print("Removing temporary images folder")
-            shutil.rmtree("tmp_umi_images")
-            print("Cleanup done")
-
-    @classmethod
-    def get_episode_idxs(cls, episode_ends: np.ndarray) -> np.ndarray:
-        # Optimized and simplified version of this function: https://github.com/real-stanford/universal_manipulation_interface/blob/298776ce251f33b6b3185a98d6e7d1f9ad49168b/diffusion_policy/common/replay_buffer.py#L374
-        from numba import jit
-
-        @jit(nopython=True)
-        def _get_episode_idxs(episode_ends):
-            result = np.zeros((episode_ends[-1],), dtype=np.int64)
-            start_idx = 0
-            for episode_number, end_idx in enumerate(episode_ends):
-                result[start_idx:end_idx] = episode_number
-                start_idx = end_idx
-            return result
-
-        return _get_episode_idxs(episode_ends)
-
-
-def _clear_folder(folder_path: str):
-    """
-    Clears all the content of the specified folder. Creates the folder if it does not exist.
-
-    Args:
-    folder_path (str): Path to the folder to clear.
-
-    Examples:
-    >>> import os
-    >>> os.makedirs('example_folder', exist_ok=True)
-    >>> with open('example_folder/temp_file.txt', 'w') as f:
-    ...     f.write('example')
-    >>> clear_folder('example_folder')
-    >>> os.listdir('example_folder')
-    []
-    """
-    if os.path.exists(folder_path):
-        for filename in os.listdir(folder_path):
-            file_path = os.path.join(folder_path, filename)
-            try:
-                if os.path.isfile(file_path) or os.path.islink(file_path):
-                    os.unlink(file_path)
-                elif os.path.isdir(file_path):
-                    shutil.rmtree(file_path)
-            except Exception as e:
-                print(f"Failed to delete {file_path}. Reason: {e}")
-    else:
-        os.makedirs(folder_path)
-
-
-def _save_image(img_array: np.array, i: int, folder_path: str):
-    """
-    Saves a single image to the specified folder.
-
-    Args:
-    img_array (ndarray): The numpy array of the image.
-    i (int): Index of the image, used for naming.
-    folder_path (str): Path to the folder where the image will be saved.
-    """
-    img = PILImage.fromarray(img_array)
-    img_format = "PNG" if img_array.dtype == np.uint8 else "JPEG"
-    img.save(os.path.join(folder_path, f"{i}.{img_format.lower()}"), quality=100)
-
-
-def _save_images_concurrently(data: dict, image_key: str, folder_path: str, max_workers: int = 4):
-    from concurrent.futures import ThreadPoolExecutor
-
-    """
-    Saves images from the zarr_data to the specified folder using multithreading.
-
-    Args:
-    zarr_data (dict): A dictionary containing image data in an array format.
-    folder_path (str): Path to the folder where images will be saved.
-    max_workers (int): The maximum number of threads to use for saving images.
-    """
-    num_images = len(data["data/camera0_rgb"])
-    _clear_folder(folder_path)  # Clear or create folder first
-
-    with ThreadPoolExecutor(max_workers=max_workers) as executor:
-        [executor.submit(_save_image, data[image_key][i], i, folder_path) for i in range(num_images)]

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -0,0 +1,234 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Process UMI (Universal Manipulation Interface) data stored in Zarr format like in: https://github.com/real-stanford/universal_manipulation_interface"""
+
+import logging
+import shutil
+from pathlib import Path
+
+import torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub._umi_imagecodecs_numcodecs import register_codecs
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir) -> bool:
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    required_datasets = {
+        "data/robot0_demo_end_pose",
+        "data/robot0_demo_start_pose",
+        "data/robot0_eef_pos",
+        "data/robot0_eef_rot_axis_angle",
+        "data/robot0_gripper_width",
+        "meta/episode_ends",
+        "data/camera0_rgb",
+    }
+    for dataset in required_datasets:
+        if dataset not in zarr_data:
+            return False
+
+    # mandatory to access zarr_data
+    register_codecs()
+    nb_frames = zarr_data["data/camera0_rgb"].shape[0]
+
+    required_datasets.remove("meta/episode_ends")
+    assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    # We process the image data separately because it is too large to fit in memory
+    end_pose = torch.from_numpy(zarr_data["data/robot0_demo_end_pose"][:])
+    start_pos = torch.from_numpy(zarr_data["data/robot0_demo_start_pose"][:])
+    eff_pos = torch.from_numpy(zarr_data["data/robot0_eef_pos"][:])
+    eff_rot_axis_angle = torch.from_numpy(zarr_data["data/robot0_eef_rot_axis_angle"][:])
+    gripper_width = torch.from_numpy(zarr_data["data/robot0_gripper_width"][:])
+
+    states_pos = torch.cat([eff_pos, eff_rot_axis_angle], dim=1)
+    states = torch.cat([states_pos, gripper_width], dim=1)
+
+    episode_ends = zarr_data["meta/episode_ends"][:]
+    num_episodes = episode_ends.shape[0]
+
+    # We convert it in torch tensor later because the jit function does not support torch tensors
+    episode_ends = torch.from_numpy(episode_ends)
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx = 0
+    for to_idx in episode_ends:
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    ep_dicts_dir = videos_dir / "ep_dicts"
+    ep_dicts_dir.mkdir(exist_ok=True, parents=True)
+    ep_dicts = []
+
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        ep_dict_path = ep_dicts_dir / f"{ep_idx}"
+        if not ep_dict_path.is_file():
+            from_idx = from_ids[selected_ep_idx]
+            to_idx = to_ids[selected_ep_idx]
+            num_frames = to_idx - from_idx
+
+            # TODO(rcadene): save temporary images of the episode?
+
+            state = states[from_idx:to_idx]
+
+            ep_dict = {}
+
+            # load 57MB of images in RAM (400x224x224x3 uint8)
+            imgs_array = zarr_data["data/camera0_rgb"][from_idx:to_idx]
+            img_key = "observation.image"
+            if video:
+                fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                video_path = videos_dir / fname
+                if not video_path.is_file():
+                    # save png images in temporary directory
+                    tmp_imgs_dir = videos_dir / "tmp_images"
+                    save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                    # encode images to a mp4 video
+                    encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+                    # clean temporary images directory
+                    shutil.rmtree(tmp_imgs_dir)
+
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+            else:
+                ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+            ep_dict["observation.state"] = state
+            ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+            ep_dict["episode_data_index_from"] = torch.tensor([from_idx] * num_frames)
+            ep_dict["episode_data_index_to"] = torch.tensor([from_idx + num_frames] * num_frames)
+            ep_dict["end_pose"] = end_pose[from_idx:to_idx]
+            ep_dict["start_pos"] = start_pos[from_idx:to_idx]
+            ep_dict["gripper_width"] = gripper_width[from_idx:to_idx]
+            torch.save(ep_dict, ep_dict_path)
+        else:
+            ep_dict = torch.load(ep_dict_path)
+
+        ep_dicts.append(ep_dict)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+    features["episode_data_index_from"] = Value(dtype="int64", id=None)
+    features["episode_data_index_to"] = Value(dtype="int64", id=None)
+    # `start_pos` and `end_pos` respectively represent the positions of the end-effector
+    # at the beginning and the end of the episode.
+    # `gripper_width` indicates the distance between the grippers, and this value is included
+    # in the state vector, which comprises the concatenation of the end-effector position
+    # and gripper width.
+    features["end_pose"] = Sequence(
+        length=data_dict["end_pose"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["start_pos"] = Sequence(
+        length=data_dict["start_pos"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["gripper_width"] = Sequence(
+        length=data_dict["gripper_width"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        # For umi cup in the wild: https://arxiv.org/pdf/2402.10329#table.caption.16
+        fps = 10
+
+    if not video:
+        logging.warning(
+            "Generating UMI dataset without `video=True` creates ~150GB on disk and requires ~80GB in RAM."
+        )
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -1,5 +1,30 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import inspect
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from typing import Dict
+
+import datasets
+import numpy
+import PIL
 import torch
 
+from lerobot.common.datasets.video_utils import encode_video_frames
+
 
 def concatenate_episodes(ep_dicts):
     data_dict = {}
@@ -18,3 +43,89 @@ def concatenate_episodes(ep_dicts):
     total_frames = data_dict["frame_index"].shape[0]
     data_dict["index"] = torch.arange(0, total_frames, 1)
     return data_dict
+
+
+def save_images_concurrently(imgs_array: numpy.array, out_dir: Path, max_workers: int = 4):
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    def save_image(img_array, i, out_dir):
+        img = PIL.Image.fromarray(img_array)
+        img.save(str(out_dir / f"frame_{i:06d}.png"), quality=100)
+
+    num_images = len(imgs_array)
+    with ThreadPoolExecutor(max_workers=max_workers) as executor:
+        [executor.submit(save_image, imgs_array[i], i, out_dir) for i in range(num_images)]
+
+
+def get_default_encoding() -> dict:
+    """Returns the default ffmpeg encoding parameters used by `encode_video_frames`."""
+    signature = inspect.signature(encode_video_frames)
+    return {
+        k: v.default
+        for k, v in signature.parameters.items()
+        if v.default is not inspect.Parameter.empty and k in ["vcodec", "pix_fmt", "g", "crf"]
+    }
+
+
+def check_repo_id(repo_id: str) -> None:
+    if len(repo_id.split("/")) != 2:
+        raise ValueError(
+            f"""`repo_id` is expected to contain a community or user id `/` the name of the dataset
+            (e.g. 'lerobot/pusht'), but contains '{repo_id}'."""
+        )
+
+
+# TODO(aliberts): remove
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torch.Tensor]:
+    """
+    Calculate episode data index for the provided HuggingFace Dataset. Relies on episode_index column of hf_dataset.
+
+    Parameters:
+    - hf_dataset (datasets.Dataset): A HuggingFace dataset containing the episode index.
+
+    Returns:
+    - episode_data_index: A dictionary containing the data index for each episode. The dictionary has two keys:
+        - "from": A tensor containing the starting index of each episode.
+        - "to": A tensor containing the ending index of each episode.
+    """
+    episode_data_index = {"from": [], "to": []}
+
+    current_episode = None
+    """
+    The episode_index is a list of integers, each representing the episode index of the corresponding example.
+    For instance, the following is a valid episode_index:
+      [0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2]
+
+    Below, we iterate through the episode_index and populate the episode_data_index dictionary with the starting and
+    ending index of each episode. For the episode_index above, the episode_data_index dictionary will look like this:
+        {
+            "from": [0, 3, 7],
+            "to": [3, 7, 12]
+        }
+    """
+    if len(hf_dataset) == 0:
+        episode_data_index = {
+            "from": torch.tensor([]),
+            "to": torch.tensor([]),
+        }
+        return episode_data_index
+    for idx, episode_idx in enumerate(hf_dataset["episode_index"]):
+        if episode_idx != current_episode:
+            # We encountered a new episode, so we append its starting location to the "from" list
+            episode_data_index["from"].append(idx)
+            # If this is not the first episode, we append the ending location of the previous episode to the "to" list
+            if current_episode is not None:
+                episode_data_index["to"].append(idx)
+            # Let's keep track of the current episode index
+            current_episode = episode_idx
+        else:
+            # We are still in the same episode, so there is nothing for us to do here
+            pass
+    # We have reached the end of the dataset, so we append the ending location of the last episode to the "to" list
+    episode_data_index["to"].append(idx + 1)
+
+    for k in ["from", "to"]:
+        episode_data_index[k] = torch.tensor(episode_data_index[k])
+
+    return episode_data_index

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py

@@ -0,0 +1,200 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Process pickle files formatted like in: https://github.com/fyhMer/fowm"""
+
+import pickle
+import shutil
+from pathlib import Path
+
+import einops
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION
+from lerobot.common.datasets.push_dataset_to_hub.utils import (
+    calculate_episode_data_index,
+    concatenate_episodes,
+    get_default_encoding,
+    save_images_concurrently,
+)
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir):
+    keys = {"actions", "rewards", "dones"}
+    nested_keys = {"observations": {"rgb", "state"}, "next_observations": {"rgb", "state"}}
+
+    xarm_files = list(raw_dir.glob("*.pkl"))
+    assert len(xarm_files) > 0
+
+    with open(xarm_files[0], "rb") as f:
+        dataset_dict = pickle.load(f)
+
+    assert isinstance(dataset_dict, dict)
+    assert all(k in dataset_dict for k in keys)
+
+    # Check for consistent lengths in nested keys
+    expected_len = len(dataset_dict["actions"])
+    assert all(len(dataset_dict[key]) == expected_len for key in keys if key in dataset_dict)
+
+    for key, subkeys in nested_keys.items():
+        nested_dict = dataset_dict.get(key, {})
+        assert all(len(nested_dict[subkey]) == expected_len for subkey in subkeys if subkey in nested_dict)
+
+
+def load_from_raw(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int,
+    video: bool,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    pkl_path = raw_dir / "buffer.pkl"
+
+    with open(pkl_path, "rb") as f:
+        pkl_data = pickle.load(f)
+
+    # load data indices from which each episode starts and ends
+    from_ids, to_ids = [], []
+    from_idx, to_idx = 0, 0
+    for done in pkl_data["dones"]:
+        to_idx += 1
+        if not done:
+            continue
+        from_ids.append(from_idx)
+        to_ids.append(to_idx)
+        from_idx = to_idx
+
+    num_episodes = len(from_ids)
+
+    ep_dicts = []
+    ep_ids = episodes if episodes else range(num_episodes)
+    for ep_idx, selected_ep_idx in tqdm.tqdm(enumerate(ep_ids)):
+        from_idx = from_ids[selected_ep_idx]
+        to_idx = to_ids[selected_ep_idx]
+        num_frames = to_idx - from_idx
+
+        image = torch.tensor(pkl_data["observations"]["rgb"][from_idx:to_idx])
+        image = einops.rearrange(image, "b c h w -> b h w c")
+        state = torch.tensor(pkl_data["observations"]["state"][from_idx:to_idx])
+        action = torch.tensor(pkl_data["actions"][from_idx:to_idx])
+        # TODO(rcadene): we have a missing last frame which is the observation when the env is done
+        # it is critical to have this frame for tdmpc to predict a "done observation/state"
+        # next_image = torch.tensor(pkl_data["next_observations"]["rgb"][from_idx:to_idx])
+        # next_state = torch.tensor(pkl_data["next_observations"]["state"][from_idx:to_idx])
+        next_reward = torch.tensor(pkl_data["rewards"][from_idx:to_idx])
+        next_done = torch.tensor(pkl_data["dones"][from_idx:to_idx])
+
+        ep_dict = {}
+
+        imgs_array = [x.numpy() for x in image]
+        img_key = "observation.image"
+        if video:
+            # save png images in temporary directory
+            tmp_imgs_dir = videos_dir / "tmp_images"
+            save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+            # encode images to a mp4 video
+            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+            video_path = videos_dir / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps, **(encoding or {}))
+
+            # clean temporary images directory
+            shutil.rmtree(tmp_imgs_dir)
+
+            # store the reference to the video frame
+            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+        else:
+            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = state
+        ep_dict["action"] = action
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        # ep_dict["next.observation.image"] = next_image
+        # ep_dict["next.observation.state"] = next_state
+        ep_dict["next.reward"] = next_reward
+        ep_dict["next.done"] = next_done
+        ep_dicts.append(ep_dict)
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = data_dict["frame_index"].shape[0]
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+    return data_dict
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+    # TODO(rcadene): add success
+    # features["next.success"] = Value(dtype='bool', id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(
+    raw_dir: Path,
+    videos_dir: Path,
+    fps: int | None = None,
+    video: bool = True,
+    episodes: list[int] | None = None,
+    encoding: dict | None = None,
+):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 15
+
+    data_dict = load_from_raw(raw_dir, videos_dir, fps, video, episodes, encoding)
+    hf_dataset = to_hf_dataset(data_dict, video)
+    episode_data_index = calculate_episode_data_index(hf_dataset)
+    info = {
+        "codebase_version": CODEBASE_VERSION,
+        "fps": fps,
+        "video": video,
+    }
+    if video:
+        info["encoding"] = get_default_encoding()
+
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/xarm_processor.py

@@ -1,145 +0,0 @@
-import pickle
-from pathlib import Path
-
-import einops
-import torch
-import tqdm
-from datasets import Dataset, Features, Image, Sequence, Value
-from PIL import Image as PILImage
-
-from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes
-from lerobot.common.datasets.utils import (
-    hf_transform_to_torch,
-)
-
-
-class XarmProcessor:
-    """Process pickle files formatted like in: https://github.com/fyhMer/fowm"""
-
-    def __init__(self, folder_path: str, fps: int | None = None):
-        self.folder_path = Path(folder_path)
-        self.keys = {"actions", "rewards", "dones", "masks"}
-        self.nested_keys = {"observations": {"rgb", "state"}, "next_observations": {"rgb", "state"}}
-        if fps is None:
-            fps = 15
-        self._fps = fps
-
-    @property
-    def fps(self) -> int:
-        return self._fps
-
-    def is_valid(self) -> bool:
-        # get all .pkl files
-        xarm_files = list(self.folder_path.glob("*.pkl"))
-        if len(xarm_files) != 1:
-            return False
-
-        try:
-            with open(xarm_files[0], "rb") as f:
-                dataset_dict = pickle.load(f)
-        except Exception:
-            return False
-
-        if not isinstance(dataset_dict, dict):
-            return False
-
-        if not all(k in dataset_dict for k in self.keys):
-            return False
-
-        # Check for consistent lengths in nested keys
-        try:
-            expected_len = len(dataset_dict["actions"])
-            if any(len(dataset_dict[key]) != expected_len for key in self.keys if key in dataset_dict):
-                return False
-
-            for key, subkeys in self.nested_keys.items():
-                nested_dict = dataset_dict.get(key, {})
-                if any(
-                    len(nested_dict[subkey]) != expected_len for subkey in subkeys if subkey in nested_dict
-                ):
-                    return False
-        except KeyError:  # If any expected key or subkey is missing
-            return False
-
-        return True  # All checks passed
-
-    def preprocess(self):
-        if not self.is_valid():
-            raise ValueError("The Xarm file is invalid or does not contain the required datasets.")
-
-        xarm_files = list(self.folder_path.glob("*.pkl"))
-
-        with open(xarm_files[0], "rb") as f:
-            dataset_dict = pickle.load(f)
-        ep_dicts = []
-        episode_data_index = {"from": [], "to": []}
-
-        id_from = 0
-        id_to = 0
-        episode_id = 0
-        total_frames = dataset_dict["actions"].shape[0]
-        for i in tqdm.tqdm(range(total_frames)):
-            id_to += 1
-
-            if not dataset_dict["dones"][i]:
-                continue
-
-            num_frames = id_to - id_from
-
-            image = torch.tensor(dataset_dict["observations"]["rgb"][id_from:id_to])
-            image = einops.rearrange(image, "b c h w -> b h w c")
-            state = torch.tensor(dataset_dict["observations"]["state"][id_from:id_to])
-            action = torch.tensor(dataset_dict["actions"][id_from:id_to])
-            # TODO(rcadene): we have a missing last frame which is the observation when the env is done
-            # it is critical to have this frame for tdmpc to predict a "done observation/state"
-            # next_image = torch.tensor(dataset_dict["next_observations"]["rgb"][id_from:id_to])
-            # next_state = torch.tensor(dataset_dict["next_observations"]["state"][id_from:id_to])
-            next_reward = torch.tensor(dataset_dict["rewards"][id_from:id_to])
-            next_done = torch.tensor(dataset_dict["dones"][id_from:id_to])
-
-            ep_dict = {
-                "observation.image": [PILImage.fromarray(x.numpy()) for x in image],
-                "observation.state": state,
-                "action": action,
-                "episode_index": torch.tensor([episode_id] * num_frames, dtype=torch.int),
-                "frame_index": torch.arange(0, num_frames, 1),
-                "timestamp": torch.arange(0, num_frames, 1) / self.fps,
-                # "next.observation.image": next_image,
-                # "next.observation.state": next_state,
-                "next.reward": next_reward,
-                "next.done": next_done,
-            }
-            ep_dicts.append(ep_dict)
-
-            episode_data_index["from"].append(id_from)
-            episode_data_index["to"].append(id_from + num_frames)
-
-            id_from = id_to
-            episode_id += 1
-
-        data_dict = concatenate_episodes(ep_dicts)
-        return data_dict, episode_data_index
-
-    def to_hf_dataset(self, data_dict):
-        features = {
-            "observation.image": Image(),
-            "observation.state": Sequence(
-                length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
-            ),
-            "action": Sequence(length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)),
-            "episode_index": Value(dtype="int64", id=None),
-            "frame_index": Value(dtype="int64", id=None),
-            "timestamp": Value(dtype="float32", id=None),
-            "next.reward": Value(dtype="float32", id=None),
-            "next.done": Value(dtype="bool", id=None),
-            #'next.success': Value(dtype='bool', id=None),
-            "index": Value(dtype="int64", id=None),
-        }
-        features = Features(features)
-        hf_dataset = Dataset.from_dict(data_dict, features=features)
-        hf_dataset.set_transform(hf_transform_to_torch)
-
-        return hf_dataset
-
-    def cleanup(self):
-        pass

lerobot/common/datasets/sampler.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Iterator, Union
+
+import torch
+
+
+class EpisodeAwareSampler:
+    def __init__(
+        self,
+        episode_data_index: dict,
+        episode_indices_to_use: Union[list, None] = None,
+        drop_n_first_frames: int = 0,
+        drop_n_last_frames: int = 0,
+        shuffle: bool = False,
+    ):
+        """Sampler that optionally incorporates episode boundary information.
+
+        Args:
+            episode_data_index: Dictionary with keys 'from' and 'to' containing the start and end indices of each episode.
+            episode_indices_to_use: List of episode indices to use. If None, all episodes are used.
+                                    Assumes that episodes are indexed from 0 to N-1.
+            drop_n_first_frames: Number of frames to drop from the start of each episode.
+            drop_n_last_frames: Number of frames to drop from the end of each episode.
+            shuffle: Whether to shuffle the indices.
+        """
+        indices = []
+        for episode_idx, (start_index, end_index) in enumerate(
+            zip(episode_data_index["from"], episode_data_index["to"], strict=True)
+        ):
+            if episode_indices_to_use is None or episode_idx in episode_indices_to_use:
+                indices.extend(
+                    range(start_index.item() + drop_n_first_frames, end_index.item() - drop_n_last_frames)
+                )
+
+        self.indices = indices
+        self.shuffle = shuffle
+
+    def __iter__(self) -> Iterator[int]:
+        if self.shuffle:
+            for i in torch.randperm(len(self.indices)):
+                yield self.indices[i]
+        else:
+            for i in self.indices:
+                yield i
+
+    def __len__(self) -> int:
+        return len(self.indices)

lerobot/common/datasets/transforms.py

@@ -0,0 +1,197 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import collections
+from typing import Any, Callable, Dict, Sequence
+
+import torch
+from torchvision.transforms import v2
+from torchvision.transforms.v2 import Transform
+from torchvision.transforms.v2 import functional as F  # noqa: N812
+
+
+class RandomSubsetApply(Transform):
+    """Apply a random subset of N transformations from a list of transformations.
+
+    Args:
+        transforms: list of transformations.
+        p: represents the multinomial probabilities (with no replacement) used for sampling the transform.
+            If the sum of the weights is not 1, they will be normalized. If ``None`` (default), all transforms
+            have the same probability.
+        n_subset: number of transformations to apply. If ``None``, all transforms are applied.
+            Must be in [1, len(transforms)].
+        random_order: apply transformations in a random order.
+    """
+
+    def __init__(
+        self,
+        transforms: Sequence[Callable],
+        p: list[float] | None = None,
+        n_subset: int | None = None,
+        random_order: bool = False,
+    ) -> None:
+        super().__init__()
+        if not isinstance(transforms, Sequence):
+            raise TypeError("Argument transforms should be a sequence of callables")
+        if p is None:
+            p = [1] * len(transforms)
+        elif len(p) != len(transforms):
+            raise ValueError(
+                f"Length of p doesn't match the number of transforms: {len(p)} != {len(transforms)}"
+            )
+
+        if n_subset is None:
+            n_subset = len(transforms)
+        elif not isinstance(n_subset, int):
+            raise TypeError("n_subset should be an int or None")
+        elif not (1 <= n_subset <= len(transforms)):
+            raise ValueError(f"n_subset should be in the interval [1, {len(transforms)}]")
+
+        self.transforms = transforms
+        total = sum(p)
+        self.p = [prob / total for prob in p]
+        self.n_subset = n_subset
+        self.random_order = random_order
+
+    def forward(self, *inputs: Any) -> Any:
+        needs_unpacking = len(inputs) > 1
+
+        selected_indices = torch.multinomial(torch.tensor(self.p), self.n_subset)
+        if not self.random_order:
+            selected_indices = selected_indices.sort().values
+
+        selected_transforms = [self.transforms[i] for i in selected_indices]
+
+        for transform in selected_transforms:
+            outputs = transform(*inputs)
+            inputs = outputs if needs_unpacking else (outputs,)
+
+        return outputs
+
+    def extra_repr(self) -> str:
+        return (
+            f"transforms={self.transforms}, "
+            f"p={self.p}, "
+            f"n_subset={self.n_subset}, "
+            f"random_order={self.random_order}"
+        )
+
+
+class SharpnessJitter(Transform):
+    """Randomly change the sharpness of an image or video.
+
+    Similar to a v2.RandomAdjustSharpness with p=1 and a sharpness_factor sampled randomly.
+    While v2.RandomAdjustSharpness applies — with a given probability — a fixed sharpness_factor to an image,
+    SharpnessJitter applies a random sharpness_factor each time. This is to have a more diverse set of
+    augmentations as a result.
+
+    A sharpness_factor of 0 gives a blurred image, 1 gives the original image while 2 increases the sharpness
+    by a factor of 2.
+
+    If the input is a :class:`torch.Tensor`,
+    it is expected to have [..., 1 or 3, H, W] shape, where ... means an arbitrary number of leading dimensions.
+
+    Args:
+        sharpness: How much to jitter sharpness. sharpness_factor is chosen uniformly from
+            [max(0, 1 - sharpness), 1 + sharpness] or the given
+            [min, max]. Should be non negative numbers.
+    """
+
+    def __init__(self, sharpness: float | Sequence[float]) -> None:
+        super().__init__()
+        self.sharpness = self._check_input(sharpness)
+
+    def _check_input(self, sharpness):
+        if isinstance(sharpness, (int, float)):
+            if sharpness < 0:
+                raise ValueError("If sharpness is a single number, it must be non negative.")
+            sharpness = [1.0 - sharpness, 1.0 + sharpness]
+            sharpness[0] = max(sharpness[0], 0.0)
+        elif isinstance(sharpness, collections.abc.Sequence) and len(sharpness) == 2:
+            sharpness = [float(v) for v in sharpness]
+        else:
+            raise TypeError(f"{sharpness=} should be a single number or a sequence with length 2.")
+
+        if not 0.0 <= sharpness[0] <= sharpness[1]:
+            raise ValueError(f"sharpnesss values should be between (0., inf), but got {sharpness}.")
+
+        return float(sharpness[0]), float(sharpness[1])
+
+    def _generate_value(self, left: float, right: float) -> float:
+        return torch.empty(1).uniform_(left, right).item()
+
+    def _transform(self, inpt: Any, params: Dict[str, Any]) -> Any:
+        sharpness_factor = self._generate_value(self.sharpness[0], self.sharpness[1])
+        return self._call_kernel(F.adjust_sharpness, inpt, sharpness_factor=sharpness_factor)
+
+
+def get_image_transforms(
+    brightness_weight: float = 1.0,
+    brightness_min_max: tuple[float, float] | None = None,
+    contrast_weight: float = 1.0,
+    contrast_min_max: tuple[float, float] | None = None,
+    saturation_weight: float = 1.0,
+    saturation_min_max: tuple[float, float] | None = None,
+    hue_weight: float = 1.0,
+    hue_min_max: tuple[float, float] | None = None,
+    sharpness_weight: float = 1.0,
+    sharpness_min_max: tuple[float, float] | None = None,
+    max_num_transforms: int | None = None,
+    random_order: bool = False,
+):
+    def check_value(name, weight, min_max):
+        if min_max is not None:
+            if len(min_max) != 2:
+                raise ValueError(
+                    f"`{name}_min_max` is expected to be a tuple of 2 dimensions, but {min_max} provided."
+                )
+            if weight < 0.0:
+                raise ValueError(
+                    f"`{name}_weight` is expected to be 0 or positive, but is negative ({weight})."
+                )
+
+    check_value("brightness", brightness_weight, brightness_min_max)
+    check_value("contrast", contrast_weight, contrast_min_max)
+    check_value("saturation", saturation_weight, saturation_min_max)
+    check_value("hue", hue_weight, hue_min_max)
+    check_value("sharpness", sharpness_weight, sharpness_min_max)
+
+    weights = []
+    transforms = []
+    if brightness_min_max is not None and brightness_weight > 0.0:
+        weights.append(brightness_weight)
+        transforms.append(v2.ColorJitter(brightness=brightness_min_max))
+    if contrast_min_max is not None and contrast_weight > 0.0:
+        weights.append(contrast_weight)
+        transforms.append(v2.ColorJitter(contrast=contrast_min_max))
+    if saturation_min_max is not None and saturation_weight > 0.0:
+        weights.append(saturation_weight)
+        transforms.append(v2.ColorJitter(saturation=saturation_min_max))
+    if hue_min_max is not None and hue_weight > 0.0:
+        weights.append(hue_weight)
+        transforms.append(v2.ColorJitter(hue=hue_min_max))
+    if sharpness_min_max is not None and sharpness_weight > 0.0:
+        weights.append(sharpness_weight)
+        transforms.append(SharpnessJitter(sharpness=sharpness_min_max))
+
+    n_subset = len(transforms)
+    if max_num_transforms is not None:
+        n_subset = min(n_subset, max_num_transforms)
+
+    if n_subset == 0:
+        return v2.Identity()
+    else:
+        # TODO(rcadene, aliberts): add v2.ToDtype float16?
+        return RandomSubsetApply(transforms, p=weights, n_subset=n_subset, random_order=random_order)

lerobot/common/datasets/utils.py

@@ -1,20 +1,70 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib.resources
 import json
-from copy import deepcopy
-from math import ceil
+import logging
+import textwrap
+from itertools import accumulate
 from pathlib import Path
+from pprint import pformat
+from typing import Any
 
 import datasets
-import einops
+import jsonlines
+import numpy as np
+import pyarrow.compute as pc
 import torch
-import tqdm
-from datasets import Image, load_dataset, load_from_disk
-from huggingface_hub import hf_hub_download
+from datasets.table import embed_table_storage
+from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from PIL import Image as PILImage
-from safetensors.torch import load_file
 from torchvision import transforms
 
+from lerobot.common.robot_devices.robots.utils import Robot
 
-def flatten_dict(d, parent_key="", sep="/"):
+DEFAULT_CHUNK_SIZE = 1000  # Max number of episodes per chunk
+
+INFO_PATH = "meta/info.json"
+EPISODES_PATH = "meta/episodes.jsonl"
+STATS_PATH = "meta/stats.json"
+TASKS_PATH = "meta/tasks.jsonl"
+
+DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
+DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+DEFAULT_IMAGE_PATH = "images/{image_key}/episode_{episode_index:06d}/frame_{frame_index:06d}.png"
+
+DATASET_CARD_TEMPLATE = """
+---
+# Metadata will go there
+---
+This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
+
+## {}
+
+"""
+
+DEFAULT_FEATURES = {
+    "timestamp": {"dtype": "float32", "shape": (1,), "names": None},
+    "frame_index": {"dtype": "int64", "shape": (1,), "names": None},
+    "episode_index": {"dtype": "int64", "shape": (1,), "names": None},
+    "index": {"dtype": "int64", "shape": (1,), "names": None},
+    "task_index": {"dtype": "int64", "shape": (1,), "names": None},
+}
+
+
+def flatten_dict(d: dict, parent_key: str = "", sep: str = "/") -> dict:
     """Flatten a nested dictionary structure by collapsing nested keys into one key with a separator.
 
     For example:
@@ -33,7 +83,7 @@ def flatten_dict(d, parent_key="", sep="/"):
     return dict(items)
 
 
-def unflatten_dict(d, sep="/"):
+def unflatten_dict(d: dict, sep: str = "/") -> dict:
     outdict = {}
     for key, value in d.items():
         parts = key.split(sep)
@@ -46,7 +96,83 @@ def unflatten_dict(d, sep="/"):
     return outdict
 
 
-def hf_transform_to_torch(items_dict):
+def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
+    serialized_dict = {key: value.tolist() for key, value in flatten_dict(stats).items()}
+    return unflatten_dict(serialized_dict)
+
+
+def write_parquet(dataset: datasets.Dataset, fpath: Path) -> None:
+    # Embed image bytes into the table before saving to parquet
+    format = dataset.format
+    dataset = dataset.with_format("arrow")
+    dataset = dataset.map(embed_table_storage, batched=False)
+    dataset = dataset.with_format(**format)
+    dataset.to_parquet(fpath)
+
+
+def load_json(fpath: Path) -> Any:
+    with open(fpath) as f:
+        return json.load(f)
+
+
+def write_json(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with open(fpath, "w") as f:
+        json.dump(data, f, indent=4, ensure_ascii=False)
+
+
+def load_jsonlines(fpath: Path) -> list[Any]:
+    with jsonlines.open(fpath, "r") as reader:
+        return list(reader)
+
+
+def write_jsonlines(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "w") as writer:
+        writer.write_all(data)
+
+
+def append_jsonlines(data: dict, fpath: Path) -> None:
+    fpath.parent.mkdir(exist_ok=True, parents=True)
+    with jsonlines.open(fpath, "a") as writer:
+        writer.write(data)
+
+
+def load_info(local_dir: Path) -> dict:
+    info = load_json(local_dir / INFO_PATH)
+    for ft in info["features"].values():
+        ft["shape"] = tuple(ft["shape"])
+    return info
+
+
+def load_stats(local_dir: Path) -> dict:
+    if not (local_dir / STATS_PATH).exists():
+        return None
+    stats = load_json(local_dir / STATS_PATH)
+    stats = {key: torch.tensor(value) for key, value in flatten_dict(stats).items()}
+    return unflatten_dict(stats)
+
+
+def load_tasks(local_dir: Path) -> dict:
+    tasks = load_jsonlines(local_dir / TASKS_PATH)
+    return {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
+
+
+def load_episodes(local_dir: Path) -> dict:
+    return load_jsonlines(local_dir / EPISODES_PATH)
+
+
+def load_image_as_numpy(fpath: str | Path, dtype="float32", channel_first: bool = True) -> np.ndarray:
+    img = PILImage.open(fpath).convert("RGB")
+    img_array = np.array(img, dtype=dtype)
+    if channel_first:  # (H, W, C) -> (C, H, W)
+        img_array = np.transpose(img_array, (2, 0, 1))
+    if "float" in dtype:
+        img_array /= 255.0
+    return img_array
+
+
+def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     """Get a transform function that convert items from Hugging Face dataset (pyarrow)
     to torch tensors. Importantly, images are converted from PIL, which corresponds to
     a channel last representation (h w c) of uint8 type, to a torch image representation
@@ -57,292 +183,260 @@ def hf_transform_to_torch(items_dict):
         if isinstance(first_item, PILImage.Image):
             to_tensor = transforms.ToTensor()
             items_dict[key] = [to_tensor(img) for img in items_dict[key]]
+        elif first_item is None:
+            pass
         else:
             items_dict[key] = [torch.tensor(x) for x in items_dict[key]]
     return items_dict
 
 
-def load_hf_dataset(repo_id, version, root, split) -> datasets.Dataset:
-    """hf_dataset contains all the observations, states, actions, rewards, etc."""
-    if root is not None:
-        hf_dataset = load_from_disk(str(Path(root) / repo_id / split))
-    else:
-        hf_dataset = load_dataset(repo_id, revision=version, split=split)
-    hf_dataset.set_transform(hf_transform_to_torch)
-    return hf_dataset
+def _get_major_minor(version: str) -> tuple[int]:
+    split = version.strip("v").split(".")
+    return int(split[0]), int(split[1])
 
 
-def load_episode_data_index(repo_id, version, root) -> dict[str, torch.Tensor]:
-    """episode_data_index contains the range of indices for each episode
+class BackwardCompatibilityError(Exception):
+    def __init__(self, repo_id, version):
+        message = textwrap.dedent(f"""
+            BackwardCompatibilityError: The dataset you requested ({repo_id}) is in {version} format.
 
-    Example:
-    ```python
-    from_id = episode_data_index["from"][episode_id].item()
-    to_id = episode_data_index["to"][episode_id].item()
-    episode_frames = [dataset[i] for i in range(from_id, to_id)]
-    ```
-    """
-    if root is not None:
-        path = Path(root) / repo_id / "meta_data" / "episode_data_index.safetensors"
-    else:
-        path = hf_hub_download(
-            repo_id, "meta_data/episode_data_index.safetensors", repo_type="dataset", revision=version
+            We introduced a new format since v2.0 which is not backward compatible with v1.x.
+            Please, use our conversion script. Modify the following command with your own task description:
+            ```
+            python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \\
+                --repo-id {repo_id} \\
+                --single-task "TASK DESCRIPTION."  # <---- /!\\ Replace TASK DESCRIPTION /!\\
+            ```
+
+            A few examples to replace TASK DESCRIPTION: "Pick up the blue cube and place it into the bin.",
+            "Insert the peg into the socket.", "Slide open the ziploc bag.", "Take the elevator to the 1st floor.",
+            "Open the top cabinet, store the pot inside it then close the cabinet.", "Push the T-shaped block onto the T-shaped target.",
+            "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.", "Fold the sweatshirt.", ...
+
+            If you encounter a problem, contact LeRobot maintainers on [Discord](https://discord.com/invite/s3KuuzsPFb)
+            or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
+        """)
+        super().__init__(message)
+
+
+def check_version_compatibility(
+    repo_id: str, version_to_check: str, current_version: str, enforce_breaking_major: bool = True
+) -> None:
+    current_major, _ = _get_major_minor(current_version)
+    major_to_check, _ = _get_major_minor(version_to_check)
+    if major_to_check < current_major and enforce_breaking_major:
+        raise BackwardCompatibilityError(repo_id, version_to_check)
+    elif float(version_to_check.strip("v")) < float(current_version.strip("v")):
+        logging.warning(
+            f"""The dataset you requested ({repo_id}) was created with a previous version ({version_to_check}) of the
+            codebase. The current codebase version is {current_version}. You should be fine since
+            backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
+            Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
         )
 
-    return load_file(path)
 
+def get_hub_safe_version(repo_id: str, version: str) -> str:
+    api = HfApi()
+    dataset_info = api.list_repo_refs(repo_id, repo_type="dataset")
+    branches = [b.name for b in dataset_info.branches]
+    if version not in branches:
+        num_version = float(version.strip("v"))
+        hub_num_versions = [float(v.strip("v")) for v in branches if v.startswith("v")]
+        if num_version >= 2.0 and all(v < 2.0 for v in hub_num_versions):
+            raise BackwardCompatibilityError(repo_id, version)
 
-def load_stats(repo_id, version, root) -> dict[str, dict[str, torch.Tensor]]:
-    """stats contains the statistics per modality computed over the full dataset, such as max, min, mean, std
-
-    Example:
-    ```python
-    normalized_action = (action - stats["action"]["mean"]) / stats["action"]["std"]
-    ```
-    """
-    if root is not None:
-        path = Path(root) / repo_id / "meta_data" / "stats.safetensors"
+        logging.warning(
+            f"""You are trying to load a dataset from {repo_id} created with a previous version of the
+            codebase. The following versions are available: {branches}.
+            The requested version ('{version}') is not found. You should be fine since
+            backward compatibility is maintained. If you encounter a problem, contact LeRobot maintainers on
+            Discord ('https://discord.com/invite/s3KuuzsPFb') or open an issue on github.""",
+        )
+        if "main" not in branches:
+            raise ValueError(f"Version 'main' not found on {repo_id}")
+        return "main"
     else:
-        path = hf_hub_download(repo_id, "meta_data/stats.safetensors", repo_type="dataset", revision=version)
-
-    stats = load_file(path)
-    return unflatten_dict(stats)
+        return version
 
 
-def load_info(repo_id, version, root) -> dict:
-    """info contains useful information regarding the dataset that are not stored elsewhere
+def get_hf_features_from_features(features: dict) -> datasets.Features:
+    hf_features = {}
+    for key, ft in features.items():
+        if ft["dtype"] == "video":
+            continue
+        elif ft["dtype"] == "image":
+            hf_features[key] = datasets.Image()
+        elif ft["shape"] == (1,):
+            hf_features[key] = datasets.Value(dtype=ft["dtype"])
+        else:
+            assert len(ft["shape"]) == 1
+            hf_features[key] = datasets.Sequence(
+                length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
+            )
 
-    Example:
-    ```python
-    print("frame per second used to collect the video", info["fps"])
-    ```
-    """
-    if root is not None:
-        path = Path(root) / repo_id / "meta_data" / "info.json"
-    else:
-        path = hf_hub_download(repo_id, "meta_data/info.json", repo_type="dataset", revision=version)
-
-    with open(path) as f:
-        info = json.load(f)
-    return info
+    return datasets.Features(hf_features)
 
 
-def load_previous_and_future_frames(
-    item: dict[str, torch.Tensor],
+def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
+    camera_ft = {}
+    if robot.cameras:
+        camera_ft = {
+            key: {"dtype": "video" if use_videos else "image", **ft}
+            for key, ft in robot.camera_features.items()
+        }
+    return {**robot.motor_features, **camera_ft, **DEFAULT_FEATURES}
+
+
+def create_empty_dataset_info(
+    codebase_version: str,
+    fps: int,
+    robot_type: str,
+    features: dict,
+    use_videos: bool,
+) -> dict:
+    return {
+        "codebase_version": codebase_version,
+        "robot_type": robot_type,
+        "total_episodes": 0,
+        "total_frames": 0,
+        "total_tasks": 0,
+        "total_videos": 0,
+        "total_chunks": 0,
+        "chunks_size": DEFAULT_CHUNK_SIZE,
+        "fps": fps,
+        "splits": {},
+        "data_path": DEFAULT_PARQUET_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
+        "features": features,
+    }
+
+
+def get_episode_data_index(
+    episode_dicts: list[dict], episodes: list[int] | None = None
+) -> dict[str, torch.Tensor]:
+    episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in enumerate(episode_dicts)}
+    if episodes is not None:
+        episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
+
+    cumulative_lenghts = list(accumulate(episode_lengths.values()))
+    return {
+        "from": torch.LongTensor([0] + cumulative_lenghts[:-1]),
+        "to": torch.LongTensor(cumulative_lenghts),
+    }
+
+
+def calculate_total_episode(
+    hf_dataset: datasets.Dataset, raise_if_not_contiguous: bool = True
+) -> dict[str, torch.Tensor]:
+    episode_indices = sorted(hf_dataset.unique("episode_index"))
+    total_episodes = len(episode_indices)
+    if raise_if_not_contiguous and episode_indices != list(range(total_episodes)):
+        raise ValueError("episode_index values are not sorted and contiguous.")
+    return total_episodes
+
+
+def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> dict[str, torch.Tensor]:
+    episode_lengths = []
+    table = hf_dataset.data.table
+    total_episodes = calculate_total_episode(hf_dataset)
+    for ep_idx in range(total_episodes):
+        ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
+        episode_lengths.insert(ep_idx, len(ep_table))
+
+    cumulative_lenghts = list(accumulate(episode_lengths))
+    return {
+        "from": torch.LongTensor([0] + cumulative_lenghts[:-1]),
+        "to": torch.LongTensor(cumulative_lenghts),
+    }
+
+
+def check_timestamps_sync(
     hf_dataset: datasets.Dataset,
     episode_data_index: dict[str, torch.Tensor],
-    delta_timestamps: dict[str, list[float]],
-    tol: float,
-) -> dict[torch.Tensor]:
+    fps: int,
+    tolerance_s: float,
+    raise_value_error: bool = True,
+) -> bool:
     """
-    Given a current item in the dataset containing a timestamp (e.g. 0.6 seconds), and a list of time differences of
-    some modalities (e.g. delta_timestamps={"observation.image": [-0.8, -0.2, 0, 0.2]}), this function computes for each
-    given modality a list of query timestamps (e.g. [-0.2, 0.4, 0.6, 0.8]) and loads the closest frames in the dataset.
-
-    Importantly, when no frame can be found around a query timestamp within a specified tolerance window, this function
-    raises an AssertionError. When a timestamp is queried before the first available timestamp of the episode or after
-    the last available timestamp, the violation of the tolerance doesnt raise an AssertionError, and the function
-    populates a boolean array indicating which frames are outside of the episode range. For instance, this boolean array
-    is useful during batched training to not supervise actions associated to timestamps coming after the end of the
-    episode, or to pad the observations in a specific way. Note that by default the observation frames before the start
-    of the episode are the same as the first frame of the episode.
-
-    Parameters:
-    - item (dict): A dictionary containing all the data related to a frame. It is the result of `dataset[idx]`. Each key
-      corresponds to a different modality (e.g., "timestamp", "observation.image", "action").
-    - hf_dataset (datasets.Dataset): A dictionary containing the full dataset. Each key corresponds to a different
-      modality (e.g., "timestamp", "observation.image", "action").
-    - episode_data_index (dict): A dictionary containing two keys ("from" and "to") associated to dataset indices.
-      They indicate the start index and end index of each episode in the dataset.
-    - delta_timestamps (dict): A dictionary containing lists of delta timestamps for each possible modality to be
-      retrieved. These deltas are added to the item timestamp to form the query timestamps.
-    - tol (float, optional): The tolerance level used to determine if a data point is close enough to the query
-      timestamp by asserting `tol > difference`. It is suggested to set `tol` to a smaller value than the
-      smallest expected inter-frame period, but large enough to account for jitter.
-
-    Returns:
-    - The same item with the queried frames for each modality specified in delta_timestamps, with an additional key for
-      each modality (e.g. "observation.image_is_pad").
-
-    Raises:
-    - AssertionError: If any of the frames unexpectedly violate the tolerance level. This could indicate synchronization
-      issues with timestamps during data collection.
+    This check is to make sure that each timestamps is separated to the next by 1/fps +/- tolerance to
+    account for possible numerical error.
     """
-    # get indices of the frames associated to the episode, and their timestamps
-    ep_id = item["episode_index"].item()
-    ep_data_id_from = episode_data_index["from"][ep_id].item()
-    ep_data_id_to = episode_data_index["to"][ep_id].item()
-    ep_data_ids = torch.arange(ep_data_id_from, ep_data_id_to, 1)
+    timestamps = torch.stack(hf_dataset["timestamp"])
+    diffs = torch.diff(timestamps)
+    within_tolerance = torch.abs(diffs - 1 / fps) <= tolerance_s
 
-    # load timestamps
-    ep_timestamps = hf_dataset.select_columns("timestamp")[ep_data_id_from:ep_data_id_to]["timestamp"]
-    ep_timestamps = torch.stack(ep_timestamps)
+    # We mask differences between the timestamp at the end of an episode
+    # and the one at the start of the next episode since these are expected
+    # to be outside tolerance.
+    mask = torch.ones(len(diffs), dtype=torch.bool)
+    ignored_diffs = episode_data_index["to"][:-1] - 1
+    mask[ignored_diffs] = False
+    filtered_within_tolerance = within_tolerance[mask]
 
-    # we make the assumption that the timestamps are sorted
-    ep_first_ts = ep_timestamps[0]
-    ep_last_ts = ep_timestamps[-1]
-    current_ts = item["timestamp"].item()
+    if not torch.all(filtered_within_tolerance):
+        # Track original indices before masking
+        original_indices = torch.arange(len(diffs))
+        filtered_indices = original_indices[mask]
+        outside_tolerance_filtered_indices = torch.nonzero(~filtered_within_tolerance)  # .squeeze()
+        outside_tolerance_indices = filtered_indices[outside_tolerance_filtered_indices]
+        episode_indices = torch.stack(hf_dataset["episode_index"])
 
-    for key in delta_timestamps:
-        # get timestamps used as query to retrieve data of previous/future frames
-        delta_ts = delta_timestamps[key]
-        query_ts = current_ts + torch.tensor(delta_ts)
+        outside_tolerances = []
+        for idx in outside_tolerance_indices:
+            entry = {
+                "timestamps": [timestamps[idx], timestamps[idx + 1]],
+                "diff": diffs[idx],
+                "episode_index": episode_indices[idx].item(),
+            }
+            outside_tolerances.append(entry)
 
-        # compute distances between each query timestamp and all timestamps of all the frames belonging to the episode
-        dist = torch.cdist(query_ts[:, None], ep_timestamps[:, None], p=1)
-        min_, argmin_ = dist.min(1)
+        if raise_value_error:
+            raise ValueError(
+                f"""One or several timestamps unexpectedly violate the tolerance inside episode range.
+                This might be due to synchronization issues with timestamps during data collection.
+                \n{pformat(outside_tolerances)}"""
+            )
+        return False
 
-        # TODO(rcadene): synchronize timestamps + interpolation if needed
-
-        is_pad = min_ > tol
-
-        # check violated query timestamps are all outside the episode range
-        assert ((query_ts[is_pad] < ep_first_ts) | (ep_last_ts < query_ts[is_pad])).all(), (
-            f"One or several timestamps unexpectedly violate the tolerance ({min_} > {tol=}) inside episode range."
-            "This might be due to synchronization issues with timestamps during data collection."
-        )
-
-        # get dataset indices corresponding to frames to be loaded
-        data_ids = ep_data_ids[argmin_]
-
-        # load frames modality
-        item[key] = hf_dataset.select_columns(key)[data_ids][key]
-        item[key] = torch.stack(item[key])
-        item[f"{key}_is_pad"] = is_pad
-
-    return item
+    return True
 
 
-def get_stats_einops_patterns(hf_dataset):
-    """These einops patterns will be used to aggregate batches and compute statistics.
-
-    Note: We assume the images of `hf_dataset` are in channel first format
+def check_delta_timestamps(
+    delta_timestamps: dict[str, list[float]], fps: int, tolerance_s: float, raise_value_error: bool = True
+) -> bool:
+    """This will check if all the values in delta_timestamps are multiples of 1/fps +/- tolerance.
+    This is to ensure that these delta_timestamps added to any timestamp from a dataset will themselves be
+    actual timestamps from the dataset.
     """
+    outside_tolerance = {}
+    for key, delta_ts in delta_timestamps.items():
+        within_tolerance = [abs(ts * fps - round(ts * fps)) / fps <= tolerance_s for ts in delta_ts]
+        if not all(within_tolerance):
+            outside_tolerance[key] = [
+                ts for ts, is_within in zip(delta_ts, within_tolerance, strict=True) if not is_within
+            ]
 
-    dataloader = torch.utils.data.DataLoader(
-        hf_dataset,
-        num_workers=0,
-        batch_size=2,
-        shuffle=False,
-    )
-    batch = next(iter(dataloader))
+    if len(outside_tolerance) > 0:
+        if raise_value_error:
+            raise ValueError(
+                f"""
+                The following delta_timestamps are found outside of tolerance range.
+                Please make sure they are multiples of 1/{fps} +/- tolerance and adjust
+                their values accordingly.
+                \n{pformat(outside_tolerance)}
+                """
+            )
+        return False
 
-    stats_patterns = {}
-    for key, feats_type in hf_dataset.features.items():
-        # sanity check that tensors are not float64
-        assert batch[key].dtype != torch.float64
-
-        if isinstance(feats_type, Image):
-            # sanity check that images are channel first
-            _, c, h, w = batch[key].shape
-            assert c < h and c < w, f"expect channel first images, but instead {batch[key].shape}"
-
-            # sanity check that images are float32 in range [0,1]
-            assert batch[key].dtype == torch.float32, f"expect torch.float32, but instead {batch[key].dtype=}"
-            assert batch[key].max() <= 1, f"expect pixels lower than 1, but instead {batch[key].max()=}"
-            assert batch[key].min() >= 0, f"expect pixels greater than 1, but instead {batch[key].min()=}"
-
-            stats_patterns[key] = "b c h w -> c 1 1"
-        elif batch[key].ndim == 2:
-            stats_patterns[key] = "b c -> c "
-        elif batch[key].ndim == 1:
-            stats_patterns[key] = "b -> 1"
-        else:
-            raise ValueError(f"{key}, {feats_type}, {batch[key].shape}")
-
-    return stats_patterns
+    return True
 
 
-def compute_stats(hf_dataset, batch_size=32, max_num_samples=None):
-    if max_num_samples is None:
-        max_num_samples = len(hf_dataset)
+def get_delta_indices(delta_timestamps: dict[str, list[float]], fps: int) -> dict[str, list[int]]:
+    delta_indices = {}
+    for key, delta_ts in delta_timestamps.items():
+        delta_indices[key] = (torch.tensor(delta_ts) * fps).long().tolist()
 
-    stats_patterns = get_stats_einops_patterns(hf_dataset)
-
-    # mean and std will be computed incrementally while max and min will track the running value.
-    mean, std, max, min = {}, {}, {}, {}
-    for key in stats_patterns:
-        mean[key] = torch.tensor(0.0).float()
-        std[key] = torch.tensor(0.0).float()
-        max[key] = torch.tensor(-float("inf")).float()
-        min[key] = torch.tensor(float("inf")).float()
-
-    def create_seeded_dataloader(hf_dataset, batch_size, seed):
-        generator = torch.Generator()
-        generator.manual_seed(seed)
-        dataloader = torch.utils.data.DataLoader(
-            hf_dataset,
-            num_workers=4,
-            batch_size=batch_size,
-            shuffle=True,
-            drop_last=False,
-            generator=generator,
-        )
-        return dataloader
-
-    # Note: Due to be refactored soon. The point of storing `first_batch` is to make sure we don't get
-    # surprises when rerunning the sampler.
-    first_batch = None
-    running_item_count = 0  # for online mean computation
-    dataloader = create_seeded_dataloader(hf_dataset, batch_size, seed=1337)
-    for i, batch in enumerate(
-        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute mean, min, max")
-    ):
-        this_batch_size = len(batch["index"])
-        running_item_count += this_batch_size
-        if first_batch is None:
-            first_batch = deepcopy(batch)
-        for key, pattern in stats_patterns.items():
-            batch[key] = batch[key].float()
-            # Numerically stable update step for mean computation.
-            batch_mean = einops.reduce(batch[key], pattern, "mean")
-            # Hint: to update the mean we need x̄ₙ = (Nₙ₋₁x̄ₙ₋₁ + Bₙxₙ) / Nₙ, where the subscript represents
-            # the update step, N is the running item count, B is this batch size, x̄ is the running mean,
-            # and x is the current batch mean. Some rearrangement is then required to avoid risking
-            # numerical overflow. Another hint: Nₙ₋₁ = Nₙ - Bₙ. Rearrangement yields
-            # x̄ₙ = x̄ₙ₋₁ + Bₙ * (xₙ - x̄ₙ₋₁) / Nₙ
-            mean[key] = mean[key] + this_batch_size * (batch_mean - mean[key]) / running_item_count
-            max[key] = torch.maximum(max[key], einops.reduce(batch[key], pattern, "max"))
-            min[key] = torch.minimum(min[key], einops.reduce(batch[key], pattern, "min"))
-
-        if i == ceil(max_num_samples / batch_size) - 1:
-            break
-
-    first_batch_ = None
-    running_item_count = 0  # for online std computation
-    dataloader = create_seeded_dataloader(hf_dataset, batch_size, seed=1337)
-    for i, batch in enumerate(
-        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute std")
-    ):
-        this_batch_size = len(batch["index"])
-        running_item_count += this_batch_size
-        # Sanity check to make sure the batches are still in the same order as before.
-        if first_batch_ is None:
-            first_batch_ = deepcopy(batch)
-            for key in stats_patterns:
-                assert torch.equal(first_batch_[key], first_batch[key])
-        for key, pattern in stats_patterns.items():
-            batch[key] = batch[key].float()
-            # Numerically stable update step for mean computation (where the mean is over squared
-            # residuals).See notes in the mean computation loop above.
-            batch_std = einops.reduce((batch[key] - mean[key]) ** 2, pattern, "mean")
-            std[key] = std[key] + this_batch_size * (batch_std - std[key]) / running_item_count
-
-        if i == ceil(max_num_samples / batch_size) - 1:
-            break
-
-    for key in stats_patterns:
-        std[key] = torch.sqrt(std[key])
-
-    stats = {}
-    for key in stats_patterns:
-        stats[key] = {
-            "mean": mean[key],
-            "std": std[key],
-            "max": max[key],
-            "min": min[key],
-        }
-    return stats
+    return delta_indices
 
 
 def cycle(iterable):
@@ -356,3 +450,55 @@ def cycle(iterable):
             yield next(iterator)
         except StopIteration:
             iterator = iter(iterable)
+
+
+def create_branch(repo_id, *, branch: str, repo_type: str | None = None) -> None:
+    """Create a branch on a existing Hugging Face repo. Delete the branch if it already
+    exists before creating it.
+    """
+    api = HfApi()
+
+    branches = api.list_repo_refs(repo_id, repo_type=repo_type).branches
+    refs = [branch.ref for branch in branches]
+    ref = f"refs/heads/{branch}"
+    if ref in refs:
+        api.delete_branch(repo_id, repo_type=repo_type, branch=branch)
+
+    api.create_branch(repo_id, repo_type=repo_type, branch=branch)
+
+
+def create_lerobot_dataset_card(
+    tags: list | None = None,
+    dataset_info: dict | None = None,
+    **kwargs,
+) -> DatasetCard:
+    """
+    Keyword arguments will be used to replace values in ./lerobot/common/datasets/card_template.md.
+    Note: If specified, license must be one of https://huggingface.co/docs/hub/repositories-licenses.
+    """
+    card_tags = ["LeRobot"]
+    card_template_path = importlib.resources.path("lerobot.common.datasets", "card_template.md")
+
+    if tags:
+        card_tags += tags
+    if dataset_info:
+        dataset_structure = "[meta/info.json](meta/info.json):\n"
+        dataset_structure += f"```json\n{json.dumps(dataset_info, indent=4)}\n```\n"
+        kwargs = {**kwargs, "dataset_structure": dataset_structure}
+    card_data = DatasetCardData(
+        license=kwargs.get("license"),
+        tags=card_tags,
+        task_categories=["robotics"],
+        configs=[
+            {
+                "config_name": "default",
+                "data_files": "data/*/*.parquet",
+            }
+        ],
+    )
+
+    return DatasetCard.from_template(
+        card_data=card_data,
+        template_path=str(card_template_path),
+        **kwargs,
+    )

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -0,0 +1,882 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script is for internal use to convert all datasets under the 'lerobot' hub user account to v2.
+
+Note: Since the original Aloha datasets don't use shadow motors, you need to comment those out in
+lerobot/configs/robot/aloha.yaml before running this script.
+"""
+
+import traceback
+from pathlib import Path
+from textwrap import dedent
+
+from lerobot import available_datasets
+from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset, parse_robot_config
+
+LOCAL_DIR = Path("data/")
+
+ALOHA_CONFIG = Path("lerobot/configs/robot/aloha.yaml")
+ALOHA_MOBILE_INFO = {
+    "robot_config": parse_robot_config(ALOHA_CONFIG),
+    "license": "mit",
+    "url": "https://mobile-aloha.github.io/",
+    "paper": "https://arxiv.org/abs/2401.02117",
+    "citation_bibtex": dedent(r"""
+        @inproceedings{fu2024mobile,
+            author    = {Fu, Zipeng and Zhao, Tony Z. and Finn, Chelsea},
+            title     = {Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation},
+            booktitle = {arXiv},
+            year      = {2024},
+        }""").lstrip(),
+}
+ALOHA_STATIC_INFO = {
+    "robot_config": parse_robot_config(ALOHA_CONFIG),
+    "license": "mit",
+    "url": "https://tonyzhaozh.github.io/aloha/",
+    "paper": "https://arxiv.org/abs/2304.13705",
+    "citation_bibtex": dedent(r"""
+        @article{Zhao2023LearningFB,
+            title={Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware},
+            author={Tony Zhao and Vikash Kumar and Sergey Levine and Chelsea Finn},
+            journal={RSS},
+            year={2023},
+            volume={abs/2304.13705},
+            url={https://arxiv.org/abs/2304.13705}
+        }""").lstrip(),
+}
+PUSHT_INFO = {
+    "license": "mit",
+    "url": "https://diffusion-policy.cs.columbia.edu/",
+    "paper": "https://arxiv.org/abs/2303.04137v5",
+    "citation_bibtex": dedent(r"""
+        @article{chi2024diffusionpolicy,
+            author = {Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song},
+            title ={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+            journal = {The International Journal of Robotics Research},
+            year = {2024},
+        }""").lstrip(),
+}
+XARM_INFO = {
+    "license": "mit",
+    "url": "https://www.nicklashansen.com/td-mpc/",
+    "paper": "https://arxiv.org/abs/2203.04955",
+    "citation_bibtex": dedent(r"""
+        @inproceedings{Hansen2022tdmpc,
+            title={Temporal Difference Learning for Model Predictive Control},
+            author={Nicklas Hansen and Xiaolong Wang and Hao Su},
+            booktitle={ICML},
+            year={2022}
+        }
+    """),
+}
+UNITREEH_INFO = {
+    "license": "apache-2.0",
+}
+
+DATASETS = {
+    "aloha_mobile_cabinet": {
+        "single_task": "Open the top cabinet, store the pot inside it then close the cabinet.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_chair": {
+        "single_task": "Push the chairs in front of the desk to place them against it.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_elevator": {
+        "single_task": "Take the elevator to the 1st floor.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_shrimp": {
+        "single_task": "Sauté the raw shrimp on both sides, then serve it in the bowl.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_wash_pan": {
+        "single_task": "Pick up the pan, rinse it in the sink and then place it in the drying rack.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_mobile_wipe_wine": {
+        "single_task": "Pick up the wet cloth on the faucet and use it to clean the spilled wine on the table and underneath the glass.",
+        **ALOHA_MOBILE_INFO,
+    },
+    "aloha_static_battery": {
+        "single_task": "Place the battery into the slot of the remote controller.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_candy": {"single_task": "Pick up the candy and unwrap it.", **ALOHA_STATIC_INFO},
+    "aloha_static_coffee": {
+        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray, then push the 'Hot Water' and 'Travel Mug' buttons.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_coffee_new": {
+        "single_task": "Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_cups_open": {
+        "single_task": "Pick up the plastic cup and open its lid.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_fork_pick_up": {
+        "single_task": "Pick up the fork and place it on the plate.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_pingpong_test": {
+        "single_task": "Transfer one of the two balls in the right glass into the left glass, then transfer it back to the right glass.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_pro_pencil": {
+        "single_task": "Pick up the pencil with the right arm, hand it over to the left arm then place it back onto the table.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_screw_driver": {
+        "single_task": "Pick up the screwdriver with the right arm, hand it over to the left arm then place it into the cup.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_tape": {
+        "single_task": "Cut a small piece of tape from the tape dispenser then place it on the cardboard box's edge.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_thread_velcro": {
+        "single_task": "Pick up the velcro cable tie with the left arm, then insert the end of the velcro tie into the other end's loop with the right arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_towel": {
+        "single_task": "Pick up a piece of paper towel and place it on the spilled liquid.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_vinh_cup": {
+        "single_task": "Pick up the platic cup with the right arm, then pop its lid open with the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_vinh_cup_left": {
+        "single_task": "Pick up the platic cup with the left arm, then pop its lid open with the right arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_static_ziploc_slide": {"single_task": "Slide open the ziploc bag.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_scripted_image": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_insertion_human": {"single_task": "Insert the peg into the socket.", **ALOHA_STATIC_INFO},
+    "aloha_sim_insertion_human_image": {
+        "single_task": "Insert the peg into the socket.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_scripted": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_scripted_image": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_human": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "aloha_sim_transfer_cube_human_image": {
+        "single_task": "Pick up the cube with the right arm and transfer it to the left arm.",
+        **ALOHA_STATIC_INFO,
+    },
+    "pusht": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "pusht_image": {"single_task": "Push the T-shaped block onto the T-shaped target.", **PUSHT_INFO},
+    "unitreeh1_fold_clothes": {"single_task": "Fold the sweatshirt.", **UNITREEH_INFO},
+    "unitreeh1_rearrange_objects": {"single_task": "Put the object into the bin.", **UNITREEH_INFO},
+    "unitreeh1_two_robot_greeting": {
+        "single_task": "Greet the other robot with a high five.",
+        **UNITREEH_INFO,
+    },
+    "unitreeh1_warehouse": {
+        "single_task": "Grab the spray paint on the shelf and place it in the bin on top of the robot dog.",
+        **UNITREEH_INFO,
+    },
+    "xarm_lift_medium": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_lift_medium_replay_image": {"single_task": "Pick up the cube and lift it.", **XARM_INFO},
+    "xarm_push_medium": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "xarm_push_medium_replay_image": {"single_task": "Push the cube onto the target.", **XARM_INFO},
+    "umi_cup_in_the_wild": {
+        "single_task": "Put the cup on the plate.",
+        "license": "apache-2.0",
+    },
+    "asu_table_top": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://link.springer.com/article/10.1007/s10514-023-10129-1",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{zhou2023modularity,
+                title={Modularity through Attention: Efficient Training and Transfer of Language-Conditioned Policies for Robot Manipulation},
+                author={Zhou, Yifan and Sonawani, Shubham and Phielipp, Mariano and Stepputtis, Simon and Amor, Heni},
+                booktitle={Conference on Robot Learning},
+                pages={1684--1695},
+                year={2023},
+                organization={PMLR}
+            }
+            @article{zhou2023learning,
+                title={Learning modular language-conditioned robot policies through attention},
+                author={Zhou, Yifan and Sonawani, Shubham and Phielipp, Mariano and Ben Amor, Heni and Stepputtis, Simon},
+                journal={Autonomous Robots},
+                pages={1--21},
+                year={2023},
+                publisher={Springer}
+            }""").lstrip(),
+    },
+    "austin_buds_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/BUDS-website/",
+        "paper": "https://arxiv.org/abs/2109.13841",
+        "citation_bibtex": dedent(r"""
+            @article{zhu2022bottom,
+                title={Bottom-Up Skill Discovery From Unsegmented Demonstrations for Long-Horizon Robot Manipulation},
+                author={Zhu, Yifeng and Stone, Peter and Zhu, Yuke},
+                journal={IEEE Robotics and Automation Letters},
+                volume={7},
+                number={2},
+                pages={4126--4133},
+                year={2022},
+                publisher={IEEE}
+            }""").lstrip(),
+    },
+    "austin_sailor_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/sailor/",
+        "paper": "https://arxiv.org/abs/2210.11435",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{nasiriany2022sailor,
+                title={Learning and Retrieval from Prior Data for Skill-based Imitation Learning},
+                author={Soroush Nasiriany and Tian Gao and Ajay Mandlekar and Yuke Zhu},
+                booktitle={Conference on Robot Learning (CoRL)},
+                year={2022}
+            }""").lstrip(),
+    },
+    "austin_sirius_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/sirius/",
+        "paper": "https://arxiv.org/abs/2211.08416",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{liu2022robot,
+                title = {Robot Learning on the Job: Human-in-the-Loop Autonomy and Learning During Deployment},
+                author = {Huihan Liu and Soroush Nasiriany and Lance Zhang and Zhiyao Bao and Yuke Zhu},
+                booktitle = {Robotics: Science and Systems (RSS)},
+                year = {2023}
+            }""").lstrip(),
+    },
+    "berkeley_autolab_ur5": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://sites.google.com/view/berkeley-ur5/home",
+        "citation_bibtex": dedent(r"""
+            @misc{BerkeleyUR5Website,
+                title = {Berkeley {UR5} Demonstration Dataset},
+                author = {Lawrence Yunliang Chen and Simeon Adebola and Ken Goldberg},
+                howpublished = {https://sites.google.com/view/berkeley-ur5/home},
+            }""").lstrip(),
+    },
+    "berkeley_cable_routing": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://sites.google.com/view/cablerouting/home",
+        "paper": "https://arxiv.org/abs/2307.08927",
+        "citation_bibtex": dedent(r"""
+            @article{luo2023multistage,
+                author    = {Jianlan Luo and Charles Xu and Xinyang Geng and Gilbert Feng and Kuan Fang and Liam Tan and Stefan Schaal and Sergey Levine},
+                title     = {Multi-Stage Cable Routing through Hierarchical Imitation Learning},
+                journal   = {arXiv pre-print},
+                year      = {2023},
+                url       = {https://arxiv.org/abs/2307.08927},
+            }""").lstrip(),
+    },
+    "berkeley_fanuc_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/berkeley.edu/fanuc-manipulation",
+        "citation_bibtex": dedent(r"""
+            @article{fanuc_manipulation2023,
+                title={Fanuc Manipulation: A Dataset for Learning-based Manipulation with FANUC Mate 200iD Robot},
+                author={Zhu, Xinghao and Tian, Ran and Xu, Chenfeng and Ding, Mingyu and Zhan, Wei and Tomizuka, Masayoshi},
+                year={2023},
+            }""").lstrip(),
+    },
+    "berkeley_gnm_cory_hall": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/1709.10489",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{kahn2018self,
+                title={Self-supervised deep reinforcement learning with generalized computation graphs for robot navigation},
+                author={Kahn, Gregory and Villaflor, Adam and Ding, Bosen and Abbeel, Pieter and Levine, Sergey},
+                booktitle={2018 IEEE international conference on robotics and automation (ICRA)},
+                pages={5129--5136},
+                year={2018},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "berkeley_gnm_recon": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/recon-robot",
+        "paper": "https://arxiv.org/abs/2104.05859",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{shah2021rapid,
+                title={Rapid Exploration for Open-World Navigation with Latent Goal Models},
+                author={Dhruv Shah and Benjamin Eysenbach and Nicholas Rhinehart and Sergey Levine},
+                booktitle={5th Annual Conference on Robot Learning },
+                year={2021},
+                url={https://openreview.net/forum?id=d_SWJhyKfVw}
+            }""").lstrip(),
+    },
+    "berkeley_gnm_sac_son": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/SACSoN-review",
+        "paper": "https://arxiv.org/abs/2306.01874",
+        "citation_bibtex": dedent(r"""
+            @article{hirose2023sacson,
+                title={SACSoN: Scalable Autonomous Data Collection for Social Navigation},
+                author={Hirose, Noriaki and Shah, Dhruv and Sridhar, Ajay and Levine, Sergey},
+                journal={arXiv preprint arXiv:2306.01874},
+                year={2023}
+            }""").lstrip(),
+    },
+    "berkeley_mvp": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/2203.06173",
+        "citation_bibtex": dedent(r"""
+            @InProceedings{Radosavovic2022,
+                title = {Real-World Robot Learning with Masked Visual Pre-training},
+                author = {Ilija Radosavovic and Tete Xiao and Stephen James and Pieter Abbeel and Jitendra Malik and Trevor Darrell},
+                booktitle = {CoRL},
+                year = {2022}
+            }""").lstrip(),
+    },
+    "berkeley_rpt": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://arxiv.org/abs/2306.10007",
+        "citation_bibtex": dedent(r"""
+            @article{Radosavovic2023,
+                title={Robot Learning with Sensorimotor Pre-training},
+                author={Ilija Radosavovic and Baifeng Shi and Letian Fu and Ken Goldberg and Trevor Darrell and Jitendra Malik},
+                year={2023},
+                journal={arXiv:2306.10007}
+            }""").lstrip(),
+    },
+    "cmu_franka_exploration_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://human-world-model.github.io/",
+        "paper": "https://arxiv.org/abs/2308.10901",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{mendonca2023structured,
+                title={Structured World Models from Human Videos},
+                author={Mendonca, Russell  and Bahl, Shikhar and Pathak, Deepak},
+                journal={RSS},
+                year={2023}
+            }""").lstrip(),
+    },
+    "cmu_play_fusion": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://play-fusion.github.io/",
+        "paper": "https://arxiv.org/abs/2312.04549",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{chen2023playfusion,
+                title={PlayFusion: Skill Acquisition via Diffusion from Language-Annotated Play},
+                author={Chen, Lili and Bahl, Shikhar and Pathak, Deepak},
+                booktitle={CoRL},
+                year={2023}
+            }""").lstrip(),
+    },
+    "cmu_stretch": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://robo-affordances.github.io/",
+        "paper": "https://arxiv.org/abs/2304.08488",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{bahl2023affordances,
+                title={Affordances from Human Videos as a Versatile Representation for Robotics},
+                author={Bahl, Shikhar and Mendonca, Russell and Chen, Lili and Jain, Unnat and Pathak, Deepak},
+                booktitle={CVPR},
+                year={2023}
+            }
+                @article{mendonca2023structured,
+                title={Structured World Models from Human Videos},
+                author={Mendonca, Russell and Bahl, Shikhar and Pathak, Deepak},
+                journal={CoRL},
+                year={2023}
+            }""").lstrip(),
+    },
+    "columbia_cairlab_pusht_real": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://diffusion-policy.cs.columbia.edu/",
+        "paper": "https://arxiv.org/abs/2303.04137v5",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{chi2023diffusionpolicy,
+                title={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
+                author={Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin and Song, Shuran},
+                booktitle={Proceedings of Robotics: Science and Systems (RSS)},
+                year={2023}
+            }""").lstrip(),
+    },
+    "conq_hose_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/conq-hose-manipulation-dataset/home",
+        "citation_bibtex": dedent(r"""
+            @misc{ConqHoseManipData,
+                author={Peter Mitrano and Dmitry Berenson},
+                title={Conq Hose Manipulation Dataset, v1.15.0},
+                year={2024},
+                howpublished={https://sites.google.com/view/conq-hose-manipulation-dataset}
+            }""").lstrip(),
+    },
+    "dlr_edan_shared_control": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://ieeexplore.ieee.org/document/9341156",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{vogel_edan_2020,
+                title = {EDAN - an EMG-Controlled Daily Assistant to Help People with Physical Disabilities},
+                language = {en},
+                booktitle = {2020 {IEEE}/{RSJ} {International} {Conference} on {Intelligent} {Robots} and {Systems} ({IROS})},
+                author = {Vogel, Jörn and Hagengruber, Annette and Iskandar, Maged and Quere, Gabriel and Leipscher, Ulrike and Bustamante, Samuel and Dietrich, Alexander and Hoeppner, Hannes and Leidner, Daniel and Albu-Schäffer, Alin},
+                year = {2020}
+            }
+            @inproceedings{quere_shared_2020,
+                address = {Paris, France},
+                title = {Shared {Control} {Templates} for {Assistive} {Robotics}},
+                language = {en},
+                booktitle = {2020 {IEEE} {International} {Conference} on {Robotics} and {Automation} ({ICRA})},
+                author = {Quere, Gabriel and Hagengruber, Annette and Iskandar, Maged and Bustamante, Samuel and Leidner, Daniel and Stulp, Freek and Vogel, Joern},
+                year = {2020},
+                pages = {7},
+            }""").lstrip(),
+    },
+    "dlr_sara_grid_clamp": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://www.researchsquare.com/article/rs-3289569/v1",
+        "citation_bibtex": dedent(r"""
+            @article{padalkar2023guided,
+                title={A guided reinforcement learning approach using shared control templates for learning manipulation skills in the real world},
+                author={Padalkar, Abhishek and Quere, Gabriel and Raffin, Antonin and Silv{\'e}rio, Jo{\~a}o and Stulp, Freek},
+                journal={Research square preprint rs-3289569/v1},
+                year={2023}
+            }""").lstrip(),
+    },
+    "dlr_sara_pour": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "paper": "https://elib.dlr.de/193739/1/padalkar2023rlsct.pdf",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{padalkar2023guiding,
+                title={Guiding Reinforcement Learning with Shared Control Templates},
+                author={Padalkar, Abhishek and Quere, Gabriel and Steinmetz, Franz and Raffin, Antonin and Nieuwenhuisen, Matthias and Silv{\'e}rio, Jo{\~a}o and Stulp, Freek},
+                booktitle={40th IEEE International Conference on Robotics and Automation, ICRA 2023},
+                year={2023},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "droid_100": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://droid-dataset.github.io/",
+        "paper": "https://arxiv.org/abs/2403.12945",
+        "citation_bibtex": dedent(r"""
+            @article{khazatsky2024droid,
+                title   = {DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset},
+                author  = {Alexander Khazatsky and Karl Pertsch and Suraj Nair and Ashwin Balakrishna and Sudeep Dasari and Siddharth Karamcheti and Soroush Nasiriany and Mohan Kumar Srirama and Lawrence Yunliang Chen and Kirsty Ellis and Peter David Fagan and Joey Hejna and Masha Itkina and Marion Lepert and Yecheng Jason Ma and Patrick Tree Miller and Jimmy Wu and Suneel Belkhale and Shivin Dass and Huy Ha and Arhan Jain and Abraham Lee and Youngwoon Lee and Marius Memmel and Sungjae Park and Ilija Radosavovic and Kaiyuan Wang and Albert Zhan and Kevin Black and Cheng Chi and Kyle Beltran Hatch and Shan Lin and Jingpei Lu and Jean Mercat and Abdul Rehman and Pannag R Sanketi and Archit Sharma and Cody Simpson and Quan Vuong and Homer Rich Walke and Blake Wulfe and Ted Xiao and Jonathan Heewon Yang and Arefeh Yavary and Tony Z. Zhao and Christopher Agia and Rohan Baijal and Mateo Guaman Castro and Daphne Chen and Qiuyu Chen and Trinity Chung and Jaimyn Drake and Ethan Paul Foster and Jensen Gao and David Antonio Herrera and Minho Heo and Kyle Hsu and Jiaheng Hu and Donovon Jackson and Charlotte Le and Yunshuang Li and Kevin Lin and Roy Lin and Zehan Ma and Abhiram Maddukuri and Suvir Mirchandani and Daniel Morton and Tony Nguyen and Abigail O'Neill and Rosario Scalise and Derick Seale and Victor Son and Stephen Tian and Emi Tran and Andrew E. Wang and Yilin Wu and Annie Xie and Jingyun Yang and Patrick Yin and Yunchu Zhang and Osbert Bastani and Glen Berseth and Jeannette Bohg and Ken Goldberg and Abhinav Gupta and Abhishek Gupta and Dinesh Jayaraman and Joseph J Lim and Jitendra Malik and Roberto Martín-Martín and Subramanian Ramamoorthy and Dorsa Sadigh and Shuran Song and Jiajun Wu and Michael C. Yip and Yuke Zhu and Thomas Kollar and Sergey Levine and Chelsea Finn},
+                year    = {2024},
+            }""").lstrip(),
+    },
+    "fmb": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://functional-manipulation-benchmark.github.io/",
+        "paper": "https://arxiv.org/abs/2401.08553",
+        "citation_bibtex": dedent(r"""
+            @article{luo2024fmb,
+                title={FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning},
+                author={Luo, Jianlan and Xu, Charles and Liu, Fangchen and Tan, Liam and Lin, Zipeng and Wu, Jeffrey and Abbeel, Pieter and Levine, Sergey},
+                journal={arXiv preprint arXiv:2401.08553},
+                year={2024}
+            }""").lstrip(),
+    },
+    "iamlab_cmu_pickup_insert": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://openreview.net/forum?id=WuBv9-IGDUA",
+        "paper": "https://arxiv.org/abs/2401.14502",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{saxena2023multiresolution,
+                title={Multi-Resolution Sensing for Real-Time Control with Vision-Language Models},
+                author={Saumya Saxena and Mohit Sharma and Oliver Kroemer},
+                booktitle={7th Annual Conference on Robot Learning},
+                year={2023},
+                url={https://openreview.net/forum?id=WuBv9-IGDUA}
+            }""").lstrip(),
+    },
+    "imperialcollege_sawyer_wrist_cam": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+    },
+    "jaco_play": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://github.com/clvrai/clvr_jaco_play_dataset",
+        "citation_bibtex": dedent(r"""
+            @software{dass2023jacoplay,
+                author = {Dass, Shivin and Yapeter, Jullian and Zhang, Jesse and Zhang, Jiahui
+                            and Pertsch, Karl and Nikolaidis, Stefanos and Lim, Joseph J.},
+                title = {CLVR Jaco Play Dataset},
+                url = {https://github.com/clvrai/clvr_jaco_play_dataset},
+                version = {1.0.0},
+                year = {2023}
+            }""").lstrip(),
+    },
+    "kaist_nonprehensile": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://github.com/JaeHyung-Kim/rlds_dataset_builder",
+        "citation_bibtex": dedent(r"""
+            @article{kimpre,
+                title={Pre-and post-contact policy decomposition for non-prehensile manipulation with zero-shot sim-to-real transfer},
+                author={Kim, Minchan and Han, Junhyek and Kim, Jaehyung and Kim, Beomjoon},
+                booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+                year={2023},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "nyu_door_opening_surprising_effectiveness": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://jyopari.github.io/VINN/",
+        "paper": "https://arxiv.org/abs/2112.01511",
+        "citation_bibtex": dedent(r"""
+            @misc{pari2021surprising,
+                title={The Surprising Effectiveness of Representation Learning for Visual Imitation},
+                author={Jyothish Pari and Nur Muhammad Shafiullah and Sridhar Pandian Arunachalam and Lerrel Pinto},
+                year={2021},
+                eprint={2112.01511},
+                archivePrefix={arXiv},
+                primaryClass={cs.RO}
+            }""").lstrip(),
+    },
+    "nyu_franka_play_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://play-to-policy.github.io/",
+        "paper": "https://arxiv.org/abs/2210.10047",
+        "citation_bibtex": dedent(r"""
+            @article{cui2022play,
+                title   = {From Play to Policy: Conditional Behavior Generation from Uncurated Robot Data},
+                author  = {Cui, Zichen Jeff and Wang, Yibin and Shafiullah, Nur Muhammad Mahi and Pinto, Lerrel},
+                journal = {arXiv preprint arXiv:2210.10047},
+                year    = {2022}
+            }""").lstrip(),
+    },
+    "nyu_rot_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://rot-robot.github.io/",
+        "paper": "https://arxiv.org/abs/2206.15469",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{haldar2023watch,
+                title={Watch and match: Supercharging imitation with regularized optimal transport},
+                author={Haldar, Siddhant and Mathur, Vaibhav and Yarats, Denis and Pinto, Lerrel},
+                booktitle={Conference on Robot Learning},
+                pages={32--43},
+                year={2023},
+                organization={PMLR}
+            }""").lstrip(),
+    },
+    "roboturk": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://roboturk.stanford.edu/dataset_real.html",
+        "paper": "PAPER",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{mandlekar2019scaling,
+                title={Scaling robot supervision to hundreds of hours with roboturk: Robotic manipulation dataset through human reasoning and dexterity},
+                author={Mandlekar, Ajay and Booher, Jonathan and Spero, Max and Tung, Albert and Gupta, Anchit and Zhu, Yuke and Garg, Animesh and Savarese, Silvio and Fei-Fei, Li},
+                booktitle={2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
+                pages={1048--1055},
+                year={2019},
+                organization={IEEE}
+            }""").lstrip(),
+    },
+    "stanford_hydra_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/hydra-il-2023",
+        "paper": "https://arxiv.org/abs/2306.17237",
+        "citation_bibtex": dedent(r"""
+            @article{belkhale2023hydra,
+                title={HYDRA: Hybrid Robot Actions for Imitation Learning},
+                author={Belkhale, Suneel and Cui, Yuchen and Sadigh, Dorsa},
+                journal={arxiv},
+                year={2023}
+            }""").lstrip(),
+    },
+    "stanford_kuka_multimodal_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://sites.google.com/view/visionandtouch",
+        "paper": "https://arxiv.org/abs/1810.10191",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{lee2019icra,
+                title={Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks},
+                author={Lee, Michelle A and Zhu, Yuke and Srinivasan, Krishnan and Shah, Parth and Savarese, Silvio and Fei-Fei, Li and  Garg, Animesh and Bohg, Jeannette},
+                booktitle={2019 IEEE International Conference on Robotics and Automation (ICRA)},
+                year={2019},
+                url={https://arxiv.org/abs/1810.10191}
+            }""").lstrip(),
+    },
+    "stanford_robocook": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://hshi74.github.io/robocook/",
+        "paper": "https://arxiv.org/abs/2306.14447",
+        "citation_bibtex": dedent(r"""
+            @article{shi2023robocook,
+                title={RoboCook: Long-Horizon Elasto-Plastic Object Manipulation with Diverse Tools},
+                author={Shi, Haochen and Xu, Huazhe and Clarke, Samuel and Li, Yunzhu and Wu, Jiajun},
+                journal={arXiv preprint arXiv:2306.14447},
+                year={2023}
+            }""").lstrip(),
+    },
+    "taco_play": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "url": "https://www.kaggle.com/datasets/oiermees/taco-robot",
+        "paper": "https://arxiv.org/abs/2209.08959, https://arxiv.org/abs/2210.01911",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{rosete2022tacorl,
+                author = {Erick Rosete-Beas and Oier Mees and Gabriel Kalweit and Joschka Boedecker and Wolfram Burgard},
+                title = {Latent Plans for Task Agnostic Offline Reinforcement Learning},
+                journal = {Proceedings of the 6th Conference on Robot Learning (CoRL)},
+                year = {2022}
+            }
+            @inproceedings{mees23hulc2,
+                title={Grounding  Language  with  Visual  Affordances  over  Unstructured  Data},
+                author={Oier Mees and Jessica Borja-Diaz and Wolfram Burgard},
+                booktitle = {Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
+                year={2023},
+                address = {London, UK}
+            }""").lstrip(),
+    },
+    "tokyo_u_lsmo": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "URL",
+        "paper": "https://arxiv.org/abs/2107.05842",
+        "citation_bibtex": dedent(r"""
+            @Article{Osa22,
+                author  = {Takayuki Osa},
+                journal = {The International Journal of Robotics Research},
+                title   = {Motion Planning by Learning the Solution Manifold in Trajectory Optimization},
+                year    = {2022},
+                number  = {3},
+                pages   = {291--311},
+                volume  = {41},
+            }""").lstrip(),
+    },
+    "toto": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://toto-benchmark.org/",
+        "paper": "https://arxiv.org/abs/2306.00942",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{zhou2023train,
+                author={Zhou, Gaoyue and Dean, Victoria and Srirama, Mohan Kumar and Rajeswaran, Aravind and Pari, Jyothish and Hatch, Kyle and Jain, Aryan and Yu, Tianhe and Abbeel, Pieter and Pinto, Lerrel and Finn, Chelsea and Gupta, Abhinav},
+                booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
+                title={Train Offline, Test Online: A Real Robot Learning Benchmark},
+                year={2023},
+            }""").lstrip(),
+    },
+    "ucsd_kitchen_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @ARTICLE{ucsd_kitchens,
+                author = {Ge Yan, Kris Wu, and Xiaolong Wang},
+                title = {{ucsd kitchens Dataset}},
+                year = {2023},
+                month = {August}
+            }""").lstrip(),
+    },
+    "ucsd_pick_and_place_dataset": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://owmcorl.github.io/#",
+        "paper": "https://arxiv.org/abs/2310.16029",
+        "citation_bibtex": dedent(r"""
+            @preprint{Feng2023Finetuning,
+                title={Finetuning Offline World Models in the Real World},
+                author={Yunhai Feng, Nicklas Hansen, Ziyan Xiong, Chandramouli Rajagopalan, Xiaolong Wang},
+                year={2023}
+            }""").lstrip(),
+    },
+    "uiuc_d3field": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://robopil.github.io/d3fields/",
+        "paper": "https://arxiv.org/abs/2309.16118",
+        "citation_bibtex": dedent(r"""
+            @article{wang2023d3field,
+                title={D^3Field: Dynamic 3D Descriptor Fields for Generalizable Robotic Manipulation},
+                author={Wang, Yixuan and Li, Zhuoran and Zhang, Mingtong and Driggs-Campbell, Katherine and Wu, Jiajun and Fei-Fei, Li and Li, Yunzhu},
+                journal={arXiv preprint arXiv:},
+                year={2023},
+            }""").lstrip(),
+    },
+    "usc_cloth_sim": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://uscresl.github.io/dmfd/",
+        "paper": "https://arxiv.org/abs/2207.10148",
+        "citation_bibtex": dedent(r"""
+            @article{salhotra2022dmfd,
+                author={Salhotra, Gautam and Liu, I-Chun Arthur and Dominguez-Kuhne, Marcus and Sukhatme, Gaurav S.},
+                journal={IEEE Robotics and Automation Letters},
+                title={Learning Deformable Object Manipulation From Expert Demonstrations},
+                year={2022},
+                volume={7},
+                number={4},
+                pages={8775-8782},
+                doi={10.1109/LRA.2022.3187843}
+            }""").lstrip(),
+    },
+    "utaustin_mutex": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/MUTEX/",
+        "paper": "https://arxiv.org/abs/2309.14320",
+        "citation_bibtex": dedent(r"""
+            @inproceedings{shah2023mutex,
+                title={{MUTEX}: Learning Unified Policies from Multimodal Task Specifications},
+                author={Rutav Shah and Roberto Mart{\'\i}n-Mart{\'\i}n and Yuke Zhu},
+                booktitle={7th Annual Conference on Robot Learning},
+                year={2023},
+                url={https://openreview.net/forum?id=PwqiqaaEzJ}
+            }""").lstrip(),
+    },
+    "utokyo_pr2_opening_fridge": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @misc{oh2023pr2utokyodatasets,
+                author={Jihoon Oh and Naoaki Kanazawa and Kento Kawaharazuka},
+                title={X-Embodiment U-Tokyo PR2 Datasets},
+                year={2023},
+                url={https://github.com/ojh6404/rlds_dataset_builder},
+            }""").lstrip(),
+    },
+    "utokyo_pr2_tabletop_manipulation": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "citation_bibtex": dedent(r"""
+            @misc{oh2023pr2utokyodatasets,
+                author={Jihoon Oh and Naoaki Kanazawa and Kento Kawaharazuka},
+                title={X-Embodiment U-Tokyo PR2 Datasets},
+                year={2023},
+                url={https://github.com/ojh6404/rlds_dataset_builder},
+            }""").lstrip(),
+    },
+    "utokyo_saytap": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://saytap.github.io/",
+        "paper": "https://arxiv.org/abs/2306.07580",
+        "citation_bibtex": dedent(r"""
+            @article{saytap2023,
+                author = {Yujin Tang and Wenhao Yu and Jie Tan and Heiga Zen and Aleksandra Faust and
+                Tatsuya Harada},
+                title  = {SayTap: Language to Quadrupedal Locomotion},
+                eprint = {arXiv:2306.07580},
+                url    = {https://saytap.github.io},
+                note   = {https://saytap.github.io},
+                year   = {2023}
+            }""").lstrip(),
+    },
+    "utokyo_xarm_bimanual": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "citation_bibtex": dedent(r"""
+            @misc{matsushima2023weblab,
+                title={Weblab xArm Dataset},
+                author={Tatsuya Matsushima and Hiroki Furuta and Yusuke Iwasawa and Yutaka Matsuo},
+                year={2023},
+            }""").lstrip(),
+    },
+    "utokyo_xarm_pick_and_place": {
+        "tasks_col": "language_instruction",
+        "license": "cc-by-4.0",
+        "citation_bibtex": dedent(r"""
+            @misc{matsushima2023weblab,
+                title={Weblab xArm Dataset},
+                author={Tatsuya Matsushima and Hiroki Furuta and Yusuke Iwasawa and Yutaka Matsuo},
+                year={2023},
+            }""").lstrip(),
+    },
+    "viola": {
+        "tasks_col": "language_instruction",
+        "license": "mit",
+        "url": "https://ut-austin-rpl.github.io/VIOLA/",
+        "paper": "https://arxiv.org/abs/2210.11339",
+        "citation_bibtex": dedent(r"""
+            @article{zhu2022viola,
+                title={VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Priors},
+                author={Zhu, Yifeng and Joshi, Abhishek and Stone, Peter and Zhu, Yuke},
+                journal={6th Annual Conference on Robot Learning (CoRL)},
+                year={2022}
+            }""").lstrip(),
+    },
+}
+
+
+def batch_convert():
+    status = {}
+    logfile = LOCAL_DIR / "conversion_log.txt"
+    assert set(DATASETS) == {id_.split("/")[1] for id_ in available_datasets}
+    for num, (name, kwargs) in enumerate(DATASETS.items()):
+        repo_id = f"lerobot/{name}"
+        print(f"\nConverting {repo_id} ({num}/{len(DATASETS)})")
+        print("---------------------------------------------------------")
+        try:
+            convert_dataset(repo_id, LOCAL_DIR, **kwargs)
+            status = f"{repo_id}: success."
+            with open(logfile, "a") as file:
+                file.write(status + "\n")
+        except Exception:
+            status = f"{repo_id}: failed\n    {traceback.format_exc()}"
+            with open(logfile, "a") as file:
+                file.write(status + "\n")
+            continue
+
+
+if __name__ == "__main__":
+    batch_convert()

lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py

@@ -0,0 +1,665 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 1.6 to
+2.0. You will be required to provide the 'tasks', which is a short but accurate description in plain English
+for each of the task performed in the dataset. This will allow to easily train models with task-conditionning.
+
+We support 3 different scenarios for these tasks (see instructions below):
+    1. Single task dataset: all episodes of your dataset have the same single task.
+    2. Single task episodes: the episodes of your dataset each contain a single task but they can differ from
+      one episode to the next.
+    3. Multi task episodes: episodes of your dataset may each contain several different tasks.
+
+
+Can you can also provide a robot config .yaml file (not mandatory) to this script via the option
+'--robot-config' so that it writes information about the robot (robot type, motors names) this dataset was
+recorded with. For now, only Aloha/Koch type robots are supported with this option.
+
+
+# 1. Single task dataset
+If your dataset contains a single task, you can simply provide it directly via the CLI with the
+'--single-task' option.
+
+Examples:
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id lerobot/aloha_sim_insertion_human_image \
+    --single-task "Insert the peg into the socket." \
+    --robot-config lerobot/configs/robot/aloha.yaml \
+    --local-dir data
+```
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id aliberts/koch_tutorial \
+    --single-task "Pick the Lego block and drop it in the box on the right." \
+    --robot-config lerobot/configs/robot/koch.yaml \
+    --local-dir data
+```
+
+
+# 2. Single task episodes
+If your dataset is a multi-task dataset, you have two options to provide the tasks to this script:
+
+- If your dataset already contains a language instruction column in its parquet file, you can simply provide
+  this column's name with the '--tasks-col' arg.
+
+    Example:
+
+    ```bash
+    python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+        --repo-id lerobot/stanford_kuka_multimodal_dataset \
+        --tasks-col "language_instruction" \
+        --local-dir data
+    ```
+
+- If your dataset doesn't contain a language instruction, you should provide the path to a .json file with the
+  '--tasks-path' arg. This file should have the following structure where keys correspond to each
+  episode_index in the dataset, and values are the language instruction for that episode.
+
+    Example:
+
+    ```json
+    {
+        "0": "Do something",
+        "1": "Do something else",
+        "2": "Do something",
+        "3": "Go there",
+        ...
+    }
+    ```
+
+# 3. Multi task episodes
+If you have multiple tasks per episodes, your dataset should contain a language instruction column in its
+parquet file, and you must provide this column's name with the '--tasks-col' arg.
+
+Example:
+
+```bash
+python lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py \
+    --repo-id lerobot/stanford_kuka_multimodal_dataset \
+    --tasks-col "language_instruction" \
+    --local-dir data
+```
+"""
+
+import argparse
+import contextlib
+import filecmp
+import json
+import logging
+import math
+import shutil
+import subprocess
+import tempfile
+from pathlib import Path
+
+import datasets
+import pyarrow.compute as pc
+import pyarrow.parquet as pq
+import torch
+from datasets import Dataset
+from huggingface_hub import HfApi
+from huggingface_hub.errors import EntryNotFoundError, HfHubHTTPError
+from safetensors.torch import load_file
+
+from lerobot.common.datasets.utils import (
+    DEFAULT_CHUNK_SIZE,
+    DEFAULT_PARQUET_PATH,
+    DEFAULT_VIDEO_PATH,
+    EPISODES_PATH,
+    INFO_PATH,
+    STATS_PATH,
+    TASKS_PATH,
+    create_branch,
+    create_lerobot_dataset_card,
+    flatten_dict,
+    get_hub_safe_version,
+    load_json,
+    unflatten_dict,
+    write_json,
+    write_jsonlines,
+)
+from lerobot.common.datasets.video_utils import (
+    VideoFrame,  # noqa: F401
+    get_image_pixel_channels,
+    get_video_info,
+)
+from lerobot.common.utils.utils import init_hydra_config
+
+V16 = "v1.6"
+V20 = "v2.0"
+
+GITATTRIBUTES_REF = "aliberts/gitattributes_reference"
+V1_VIDEO_FILE = "{video_key}_episode_{episode_index:06d}.mp4"
+V1_INFO_PATH = "meta_data/info.json"
+V1_STATS_PATH = "meta_data/stats.safetensors"
+
+
+def parse_robot_config(config_path: Path, config_overrides: list[str] | None = None) -> tuple[str, dict]:
+    robot_cfg = init_hydra_config(config_path, config_overrides)
+    if robot_cfg["robot_type"] in ["aloha", "koch"]:
+        state_names = [
+            f"{arm}_{motor}" if len(robot_cfg["follower_arms"]) > 1 else motor
+            for arm in robot_cfg["follower_arms"]
+            for motor in robot_cfg["follower_arms"][arm]["motors"]
+        ]
+        action_names = [
+            # f"{arm}_{motor}" for arm in ["left", "right"] for motor in robot_cfg["leader_arms"][arm]["motors"]
+            f"{arm}_{motor}" if len(robot_cfg["leader_arms"]) > 1 else motor
+            for arm in robot_cfg["leader_arms"]
+            for motor in robot_cfg["leader_arms"][arm]["motors"]
+        ]
+    # elif robot_cfg["robot_type"] == "stretch3": TODO
+    else:
+        raise NotImplementedError(
+            "Please provide robot_config={'robot_type': ..., 'names': ...} directly to convert_dataset()."
+        )
+
+    return {
+        "robot_type": robot_cfg["robot_type"],
+        "names": {
+            "observation.state": state_names,
+            "observation.effort": state_names,
+            "action": action_names,
+        },
+    }
+
+
+def convert_stats_to_json(v1_dir: Path, v2_dir: Path) -> None:
+    safetensor_path = v1_dir / V1_STATS_PATH
+    stats = load_file(safetensor_path)
+    serialized_stats = {key: value.tolist() for key, value in stats.items()}
+    serialized_stats = unflatten_dict(serialized_stats)
+
+    json_path = v2_dir / STATS_PATH
+    json_path.parent.mkdir(exist_ok=True, parents=True)
+    with open(json_path, "w") as f:
+        json.dump(serialized_stats, f, indent=4)
+
+    # Sanity check
+    with open(json_path) as f:
+        stats_json = json.load(f)
+
+    stats_json = flatten_dict(stats_json)
+    stats_json = {key: torch.tensor(value) for key, value in stats_json.items()}
+    for key in stats:
+        torch.testing.assert_close(stats_json[key], stats[key])
+
+
+def get_features_from_hf_dataset(dataset: Dataset, robot_config: dict | None = None) -> dict[str, list]:
+    features = {}
+    for key, ft in dataset.features.items():
+        if isinstance(ft, datasets.Value):
+            dtype = ft.dtype
+            shape = (1,)
+            names = None
+        if isinstance(ft, datasets.Sequence):
+            assert isinstance(ft.feature, datasets.Value)
+            dtype = ft.feature.dtype
+            shape = (ft.length,)
+            motor_names = (
+                robot_config["names"][key] if robot_config else [f"motor_{i}" for i in range(ft.length)]
+            )
+            assert len(motor_names) == shape[0]
+            names = {"motors": motor_names}
+        elif isinstance(ft, datasets.Image):
+            dtype = "image"
+            image = dataset[0][key]  # Assuming first row
+            channels = get_image_pixel_channels(image)
+            shape = (image.height, image.width, channels)
+            names = ["height", "width", "channel"]
+        elif ft._type == "VideoFrame":
+            dtype = "video"
+            shape = None  # Add shape later
+            names = ["height", "width", "channel"]
+
+        features[key] = {
+            "dtype": dtype,
+            "shape": shape,
+            "names": names,
+        }
+
+    return features
+
+
+def add_task_index_by_episodes(dataset: Dataset, tasks_by_episodes: dict) -> tuple[Dataset, list[str]]:
+    df = dataset.to_pandas()
+    tasks = list(set(tasks_by_episodes.values()))
+    tasks_to_task_index = {task: task_idx for task_idx, task in enumerate(tasks)}
+    episodes_to_task_index = {ep_idx: tasks_to_task_index[task] for ep_idx, task in tasks_by_episodes.items()}
+    df["task_index"] = df["episode_index"].map(episodes_to_task_index).astype(int)
+
+    features = dataset.features
+    features["task_index"] = datasets.Value(dtype="int64")
+    dataset = Dataset.from_pandas(df, features=features, split="train")
+    return dataset, tasks
+
+
+def add_task_index_from_tasks_col(
+    dataset: Dataset, tasks_col: str
+) -> tuple[Dataset, dict[str, list[str]], list[str]]:
+    df = dataset.to_pandas()
+
+    # HACK: This is to clean some of the instructions in our version of Open X datasets
+    prefix_to_clean = "tf.Tensor(b'"
+    suffix_to_clean = "', shape=(), dtype=string)"
+    df[tasks_col] = df[tasks_col].str.removeprefix(prefix_to_clean).str.removesuffix(suffix_to_clean)
+
+    # Create task_index col
+    tasks_by_episode = df.groupby("episode_index")[tasks_col].unique().apply(lambda x: x.tolist()).to_dict()
+    tasks = df[tasks_col].unique().tolist()
+    tasks_to_task_index = {task: idx for idx, task in enumerate(tasks)}
+    df["task_index"] = df[tasks_col].map(tasks_to_task_index).astype(int)
+
+    # Build the dataset back from df
+    features = dataset.features
+    features["task_index"] = datasets.Value(dtype="int64")
+    dataset = Dataset.from_pandas(df, features=features, split="train")
+    dataset = dataset.remove_columns(tasks_col)
+
+    return dataset, tasks, tasks_by_episode
+
+
+def split_parquet_by_episodes(
+    dataset: Dataset,
+    total_episodes: int,
+    total_chunks: int,
+    output_dir: Path,
+) -> list:
+    table = dataset.data.table
+    episode_lengths = []
+    for ep_chunk in range(total_chunks):
+        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
+        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
+        chunk_dir = "/".join(DEFAULT_PARQUET_PATH.split("/")[:-1]).format(episode_chunk=ep_chunk)
+        (output_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
+        for ep_idx in range(ep_chunk_start, ep_chunk_end):
+            ep_table = table.filter(pc.equal(table["episode_index"], ep_idx))
+            episode_lengths.insert(ep_idx, len(ep_table))
+            output_file = output_dir / DEFAULT_PARQUET_PATH.format(
+                episode_chunk=ep_chunk, episode_index=ep_idx
+            )
+            pq.write_table(ep_table, output_file)
+
+    return episode_lengths
+
+
+def move_videos(
+    repo_id: str,
+    video_keys: list[str],
+    total_episodes: int,
+    total_chunks: int,
+    work_dir: Path,
+    clean_gittatributes: Path,
+    branch: str = "main",
+) -> None:
+    """
+    HACK: Since HfApi() doesn't provide a way to move files directly in a repo, this function will run git
+    commands to fetch git lfs video files references to move them into subdirectories without having to
+    actually download them.
+    """
+    _lfs_clone(repo_id, work_dir, branch)
+
+    videos_moved = False
+    video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*.mp4")]
+    if len(video_files) == 0:
+        video_files = [str(f.relative_to(work_dir)) for f in work_dir.glob("videos*/*/*/*.mp4")]
+        videos_moved = True  # Videos have already been moved
+
+    assert len(video_files) == total_episodes * len(video_keys)
+
+    lfs_untracked_videos = _get_lfs_untracked_videos(work_dir, video_files)
+
+    current_gittatributes = work_dir / ".gitattributes"
+    if not filecmp.cmp(current_gittatributes, clean_gittatributes, shallow=False):
+        fix_gitattributes(work_dir, current_gittatributes, clean_gittatributes)
+
+    if lfs_untracked_videos:
+        fix_lfs_video_files_tracking(work_dir, video_files)
+
+    if videos_moved:
+        return
+
+    video_dirs = sorted(work_dir.glob("videos*/"))
+    for ep_chunk in range(total_chunks):
+        ep_chunk_start = DEFAULT_CHUNK_SIZE * ep_chunk
+        ep_chunk_end = min(DEFAULT_CHUNK_SIZE * (ep_chunk + 1), total_episodes)
+        for vid_key in video_keys:
+            chunk_dir = "/".join(DEFAULT_VIDEO_PATH.split("/")[:-1]).format(
+                episode_chunk=ep_chunk, video_key=vid_key
+            )
+            (work_dir / chunk_dir).mkdir(parents=True, exist_ok=True)
+
+            for ep_idx in range(ep_chunk_start, ep_chunk_end):
+                target_path = DEFAULT_VIDEO_PATH.format(
+                    episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_idx
+                )
+                video_file = V1_VIDEO_FILE.format(video_key=vid_key, episode_index=ep_idx)
+                if len(video_dirs) == 1:
+                    video_path = video_dirs[0] / video_file
+                else:
+                    for dir in video_dirs:
+                        if (dir / video_file).is_file():
+                            video_path = dir / video_file
+                            break
+
+                video_path.rename(work_dir / target_path)
+
+    commit_message = "Move video files into chunk subdirectories"
+    subprocess.run(["git", "add", "."], cwd=work_dir, check=True)
+    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def fix_lfs_video_files_tracking(work_dir: Path, lfs_untracked_videos: list[str]) -> None:
+    """
+    HACK: This function fixes the tracking by git lfs which was not properly set on some repos. In that case,
+    there's no other option than to download the actual files and reupload them with lfs tracking.
+    """
+    for i in range(0, len(lfs_untracked_videos), 100):
+        files = lfs_untracked_videos[i : i + 100]
+        try:
+            subprocess.run(["git", "rm", "--cached", *files], cwd=work_dir, capture_output=True, check=True)
+        except subprocess.CalledProcessError as e:
+            print("git rm --cached ERROR:")
+            print(e.stderr)
+        subprocess.run(["git", "add", *files], cwd=work_dir, check=True)
+
+    commit_message = "Track video files with git lfs"
+    subprocess.run(["git", "commit", "-m", commit_message], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def fix_gitattributes(work_dir: Path, current_gittatributes: Path, clean_gittatributes: Path) -> None:
+    shutil.copyfile(clean_gittatributes, current_gittatributes)
+    subprocess.run(["git", "add", ".gitattributes"], cwd=work_dir, check=True)
+    subprocess.run(["git", "commit", "-m", "Fix .gitattributes"], cwd=work_dir, check=True)
+    subprocess.run(["git", "push"], cwd=work_dir, check=True)
+
+
+def _lfs_clone(repo_id: str, work_dir: Path, branch: str) -> None:
+    subprocess.run(["git", "lfs", "install"], cwd=work_dir, check=True)
+    repo_url = f"https://huggingface.co/datasets/{repo_id}"
+    env = {"GIT_LFS_SKIP_SMUDGE": "1"}  # Prevent downloading LFS files
+    subprocess.run(
+        ["git", "clone", "--branch", branch, "--single-branch", "--depth", "1", repo_url, str(work_dir)],
+        check=True,
+        env=env,
+    )
+
+
+def _get_lfs_untracked_videos(work_dir: Path, video_files: list[str]) -> list[str]:
+    lfs_tracked_files = subprocess.run(
+        ["git", "lfs", "ls-files", "-n"], cwd=work_dir, capture_output=True, text=True, check=True
+    )
+    lfs_tracked_files = set(lfs_tracked_files.stdout.splitlines())
+    return [f for f in video_files if f not in lfs_tracked_files]
+
+
+def get_videos_info(repo_id: str, local_dir: Path, video_keys: list[str], branch: str) -> dict:
+    # Assumes first episode
+    video_files = [
+        DEFAULT_VIDEO_PATH.format(episode_chunk=0, video_key=vid_key, episode_index=0)
+        for vid_key in video_keys
+    ]
+    hub_api = HfApi()
+    hub_api.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", local_dir=local_dir, revision=branch, allow_patterns=video_files
+    )
+    videos_info_dict = {}
+    for vid_key, vid_path in zip(video_keys, video_files, strict=True):
+        videos_info_dict[vid_key] = get_video_info(local_dir / vid_path)
+
+    return videos_info_dict
+
+
+def convert_dataset(
+    repo_id: str,
+    local_dir: Path,
+    single_task: str | None = None,
+    tasks_path: Path | None = None,
+    tasks_col: Path | None = None,
+    robot_config: dict | None = None,
+    test_branch: str | None = None,
+    **card_kwargs,
+):
+    v1 = get_hub_safe_version(repo_id, V16)
+    v1x_dir = local_dir / V16 / repo_id
+    v20_dir = local_dir / V20 / repo_id
+    v1x_dir.mkdir(parents=True, exist_ok=True)
+    v20_dir.mkdir(parents=True, exist_ok=True)
+
+    hub_api = HfApi()
+    hub_api.snapshot_download(
+        repo_id=repo_id, repo_type="dataset", revision=v1, local_dir=v1x_dir, ignore_patterns="videos*/"
+    )
+    branch = "main"
+    if test_branch:
+        branch = test_branch
+        create_branch(repo_id=repo_id, branch=test_branch, repo_type="dataset")
+
+    metadata_v1 = load_json(v1x_dir / V1_INFO_PATH)
+    dataset = datasets.load_dataset("parquet", data_dir=v1x_dir / "data", split="train")
+    features = get_features_from_hf_dataset(dataset, robot_config)
+    video_keys = [key for key, ft in features.items() if ft["dtype"] == "video"]
+
+    if single_task and "language_instruction" in dataset.column_names:
+        logging.warning(
+            "'single_task' provided but 'language_instruction' tasks_col found. Using 'language_instruction'.",
+        )
+        single_task = None
+        tasks_col = "language_instruction"
+
+    # Episodes & chunks
+    episode_indices = sorted(dataset.unique("episode_index"))
+    total_episodes = len(episode_indices)
+    assert episode_indices == list(range(total_episodes))
+    total_videos = total_episodes * len(video_keys)
+    total_chunks = total_episodes // DEFAULT_CHUNK_SIZE
+    if total_episodes % DEFAULT_CHUNK_SIZE != 0:
+        total_chunks += 1
+
+    # Tasks
+    if single_task:
+        tasks_by_episodes = {ep_idx: single_task for ep_idx in episode_indices}
+        dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
+        tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
+    elif tasks_path:
+        tasks_by_episodes = load_json(tasks_path)
+        tasks_by_episodes = {int(ep_idx): task for ep_idx, task in tasks_by_episodes.items()}
+        dataset, tasks = add_task_index_by_episodes(dataset, tasks_by_episodes)
+        tasks_by_episodes = {ep_idx: [task] for ep_idx, task in tasks_by_episodes.items()}
+    elif tasks_col:
+        dataset, tasks, tasks_by_episodes = add_task_index_from_tasks_col(dataset, tasks_col)
+    else:
+        raise ValueError
+
+    assert set(tasks) == {task for ep_tasks in tasks_by_episodes.values() for task in ep_tasks}
+    tasks = [{"task_index": task_idx, "task": task} for task_idx, task in enumerate(tasks)]
+    write_jsonlines(tasks, v20_dir / TASKS_PATH)
+    features["task_index"] = {
+        "dtype": "int64",
+        "shape": (1,),
+        "names": None,
+    }
+
+    # Videos
+    if video_keys:
+        assert metadata_v1.get("video", False)
+        dataset = dataset.remove_columns(video_keys)
+        clean_gitattr = Path(
+            hub_api.hf_hub_download(
+                repo_id=GITATTRIBUTES_REF, repo_type="dataset", local_dir=local_dir, filename=".gitattributes"
+            )
+        ).absolute()
+        with tempfile.TemporaryDirectory() as tmp_video_dir:
+            move_videos(
+                repo_id, video_keys, total_episodes, total_chunks, Path(tmp_video_dir), clean_gitattr, branch
+            )
+        videos_info = get_videos_info(repo_id, v1x_dir, video_keys=video_keys, branch=branch)
+        for key in video_keys:
+            features[key]["shape"] = (
+                videos_info[key].pop("video.height"),
+                videos_info[key].pop("video.width"),
+                videos_info[key].pop("video.channels"),
+            )
+            features[key]["video_info"] = videos_info[key]
+            assert math.isclose(videos_info[key]["video.fps"], metadata_v1["fps"], rel_tol=1e-3)
+            if "encoding" in metadata_v1:
+                assert videos_info[key]["video.pix_fmt"] == metadata_v1["encoding"]["pix_fmt"]
+    else:
+        assert metadata_v1.get("video", 0) == 0
+        videos_info = None
+
+    # Split data into 1 parquet file by episode
+    episode_lengths = split_parquet_by_episodes(dataset, total_episodes, total_chunks, v20_dir)
+
+    if robot_config is not None:
+        robot_type = robot_config["robot_type"]
+        repo_tags = [robot_type]
+    else:
+        robot_type = "unknown"
+        repo_tags = None
+
+    # Episodes
+    episodes = [
+        {"episode_index": ep_idx, "tasks": tasks_by_episodes[ep_idx], "length": episode_lengths[ep_idx]}
+        for ep_idx in episode_indices
+    ]
+    write_jsonlines(episodes, v20_dir / EPISODES_PATH)
+
+    # Assemble metadata v2.0
+    metadata_v2_0 = {
+        "codebase_version": V20,
+        "robot_type": robot_type,
+        "total_episodes": total_episodes,
+        "total_frames": len(dataset),
+        "total_tasks": len(tasks),
+        "total_videos": total_videos,
+        "total_chunks": total_chunks,
+        "chunks_size": DEFAULT_CHUNK_SIZE,
+        "fps": metadata_v1["fps"],
+        "splits": {"train": f"0:{total_episodes}"},
+        "data_path": DEFAULT_PARQUET_PATH,
+        "video_path": DEFAULT_VIDEO_PATH if video_keys else None,
+        "features": features,
+    }
+    write_json(metadata_v2_0, v20_dir / INFO_PATH)
+    convert_stats_to_json(v1x_dir, v20_dir)
+    card = create_lerobot_dataset_card(tags=repo_tags, dataset_info=metadata_v2_0, **card_kwargs)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta_data", repo_type="dataset", revision=branch)
+
+    with contextlib.suppress(EntryNotFoundError, HfHubHTTPError):
+        hub_api.delete_folder(repo_id=repo_id, path_in_repo="meta", repo_type="dataset", revision=branch)
+
+    hub_api.upload_folder(
+        repo_id=repo_id,
+        path_in_repo="data",
+        folder_path=v20_dir / "data",
+        repo_type="dataset",
+        revision=branch,
+    )
+    hub_api.upload_folder(
+        repo_id=repo_id,
+        path_in_repo="meta",
+        folder_path=v20_dir / "meta",
+        repo_type="dataset",
+        revision=branch,
+    )
+
+    card.push_to_hub(repo_id=repo_id, repo_type="dataset", revision=branch)
+
+    if not test_branch:
+        create_branch(repo_id=repo_id, branch=V20, repo_type="dataset")
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    task_args = parser.add_mutually_exclusive_group(required=True)
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Repository identifier on Hugging Face: a community or a user name `/` the name of the dataset (e.g. `lerobot/pusht`, `cadene/aloha_sim_insertion_human`).",
+    )
+    task_args.add_argument(
+        "--single-task",
+        type=str,
+        help="A short but accurate description of the single task performed in the dataset.",
+    )
+    task_args.add_argument(
+        "--tasks-col",
+        type=str,
+        help="The name of the column containing language instructions",
+    )
+    task_args.add_argument(
+        "--tasks-path",
+        type=Path,
+        help="The path to a .json file containing one language instruction for each episode_index",
+    )
+    parser.add_argument(
+        "--robot-config",
+        type=Path,
+        default=None,
+        help="Path to the robot's config yaml the dataset during conversion.",
+    )
+    parser.add_argument(
+        "--robot-overrides",
+        type=str,
+        nargs="*",
+        help="Any key=value arguments to override the robot config values (use dots for.nested=overrides)",
+    )
+    parser.add_argument(
+        "--local-dir",
+        type=Path,
+        default=None,
+        help="Local directory to store the dataset during conversion. Defaults to /tmp/lerobot_dataset_v2",
+    )
+    parser.add_argument(
+        "--license",
+        type=str,
+        default="apache-2.0",
+        help="Repo license. Must be one of https://huggingface.co/docs/hub/repositories-licenses. Defaults to mit.",
+    )
+    parser.add_argument(
+        "--test-branch",
+        type=str,
+        default=None,
+        help="Repo branch to test your conversion first (e.g. 'v2.0.test')",
+    )
+
+    args = parser.parse_args()
+    if not args.local_dir:
+        args.local_dir = Path("/tmp/lerobot_dataset_v2")
+
+    robot_config = parse_robot_config(args.robot_config, args.robot_overrides) if args.robot_config else None
+    del args.robot_config, args.robot_overrides
+
+    convert_dataset(**vars(args), robot_config=robot_config)
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/video_utils.py

@@ -0,0 +1,315 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import json
+import logging
+import subprocess
+import warnings
+from collections import OrderedDict
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, ClassVar
+
+import pyarrow as pa
+import torch
+import torchvision
+from datasets.features.features import register_feature
+from PIL import Image
+
+
+def decode_video_frames_torchvision(
+    video_path: Path | str,
+    timestamps: list[float],
+    tolerance_s: float,
+    backend: str = "pyav",
+    log_loaded_timestamps: bool = False,
+) -> torch.Tensor:
+    """Loads frames associated to the requested timestamps of a video
+
+    The backend can be either "pyav" (default) or "video_reader".
+    "video_reader" requires installing torchvision from source, see:
+    https://github.com/pytorch/vision/blob/main/torchvision/csrc/io/decoder/gpu/README.rst
+    (note that you need to compile against ffmpeg<4.3)
+
+    While both use cpu, "video_reader" is supposedly faster than "pyav" but requires additional setup.
+    For more info on video decoding, see `benchmark/video/README.md`
+
+    See torchvision doc for more info on these two backends:
+    https://pytorch.org/vision/0.18/index.html?highlight=backend#torchvision.set_video_backend
+
+    Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
+    the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
+    that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
+    and all subsequent frames until reaching the requested frame. The number of key frames in a video
+    can be adjusted during encoding to take into account decoding time and video size in bytes.
+    """
+    video_path = str(video_path)
+
+    # set backend
+    keyframes_only = False
+    torchvision.set_video_backend(backend)
+    if backend == "pyav":
+        keyframes_only = True  # pyav doesnt support accuracte seek
+
+    # set a video stream reader
+    # TODO(rcadene): also load audio stream at the same time
+    reader = torchvision.io.VideoReader(video_path, "video")
+
+    # set the first and last requested timestamps
+    # Note: previous timestamps are usually loaded, since we need to access the previous key frame
+    first_ts = timestamps[0]
+    last_ts = timestamps[-1]
+
+    # access closest key frame of the first requested frame
+    # Note: closest key frame timestamp is usally smaller than `first_ts` (e.g. key frame can be the first frame of the video)
+    # for details on what `seek` is doing see: https://pyav.basswood-io.com/docs/stable/api/container.html?highlight=inputcontainer#av.container.InputContainer.seek
+    reader.seek(first_ts, keyframes_only=keyframes_only)
+
+    # load all frames until last requested frame
+    loaded_frames = []
+    loaded_ts = []
+    for frame in reader:
+        current_ts = frame["pts"]
+        if log_loaded_timestamps:
+            logging.info(f"frame loaded at timestamp={current_ts:.4f}")
+        loaded_frames.append(frame["data"])
+        loaded_ts.append(current_ts)
+        if current_ts >= last_ts:
+            break
+
+    if backend == "pyav":
+        reader.container.close()
+
+    reader = None
+
+    query_ts = torch.tensor(timestamps)
+    loaded_ts = torch.tensor(loaded_ts)
+
+    # compute distances between each query timestamp and timestamps of all loaded frames
+    dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
+    min_, argmin_ = dist.min(1)
+
+    is_within_tol = min_ < tolerance_s
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+        f"\nqueried timestamps: {query_ts}"
+        f"\nloaded timestamps: {loaded_ts}"
+        f"\nvideo: {video_path}"
+        f"\nbackend: {backend}"
+    )
+
+    # get closest frames to the query timestamps
+    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
+    closest_ts = loaded_ts[argmin_]
+
+    if log_loaded_timestamps:
+        logging.info(f"{closest_ts=}")
+
+    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
+    closest_frames = closest_frames.type(torch.float32) / 255
+
+    assert len(timestamps) == len(closest_frames)
+    return closest_frames
+
+
+def encode_video_frames(
+    imgs_dir: Path | str,
+    video_path: Path | str,
+    fps: int,
+    vcodec: str = "libsvtav1",
+    pix_fmt: str = "yuv420p",
+    g: int | None = 2,
+    crf: int | None = 30,
+    fast_decode: int = 0,
+    log_level: str | None = "error",
+    overwrite: bool = False,
+) -> None:
+    """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
+    video_path = Path(video_path)
+    video_path.parent.mkdir(parents=True, exist_ok=True)
+
+    ffmpeg_args = OrderedDict(
+        [
+            ("-f", "image2"),
+            ("-r", str(fps)),
+            ("-i", str(imgs_dir / "frame_%06d.png")),
+            ("-vcodec", vcodec),
+            ("-pix_fmt", pix_fmt),
+        ]
+    )
+
+    if g is not None:
+        ffmpeg_args["-g"] = str(g)
+
+    if crf is not None:
+        ffmpeg_args["-crf"] = str(crf)
+
+    if fast_decode:
+        key = "-svtav1-params" if vcodec == "libsvtav1" else "-tune"
+        value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
+        ffmpeg_args[key] = value
+
+    if log_level is not None:
+        ffmpeg_args["-loglevel"] = str(log_level)
+
+    ffmpeg_args = [item for pair in ffmpeg_args.items() for item in pair]
+    if overwrite:
+        ffmpeg_args.append("-y")
+
+    ffmpeg_cmd = ["ffmpeg"] + ffmpeg_args + [str(video_path)]
+    # redirect stdin to subprocess.DEVNULL to prevent reading random keyboard inputs from terminal
+    subprocess.run(ffmpeg_cmd, check=True, stdin=subprocess.DEVNULL)
+
+    if not video_path.exists():
+        raise OSError(
+            f"Video encoding did not work. File not found: {video_path}. "
+            f"Try running the command manually to debug: `{''.join(ffmpeg_cmd)}`"
+        )
+
+
+@dataclass
+class VideoFrame:
+    # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
+    """
+    Provides a type for a dataset containing video frames.
+
+    Example:
+
+    ```python
+    data_dict = [{"image": {"path": "videos/episode_0.mp4", "timestamp": 0.3}}]
+    features = {"image": VideoFrame()}
+    Dataset.from_dict(data_dict, features=Features(features))
+    ```
+    """
+
+    pa_type: ClassVar[Any] = pa.struct({"path": pa.string(), "timestamp": pa.float32()})
+    _type: str = field(default="VideoFrame", init=False, repr=False)
+
+    def __call__(self):
+        return self.pa_type
+
+
+with warnings.catch_warnings():
+    warnings.filterwarnings(
+        "ignore",
+        "'register_feature' is experimental and might be subject to breaking changes in the future.",
+        category=UserWarning,
+    )
+    # to make VideoFrame available in HuggingFace `datasets`
+    register_feature(VideoFrame, "VideoFrame")
+
+
+def get_audio_info(video_path: Path | str) -> dict:
+    ffprobe_audio_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "a:0",
+        "-show_entries",
+        "stream=channels,codec_name,bit_rate,sample_rate,bit_depth,channel_layout,duration",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_audio_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+
+    info = json.loads(result.stdout)
+    audio_stream_info = info["streams"][0] if info.get("streams") else None
+    if audio_stream_info is None:
+        return {"has_audio": False}
+
+    # Return the information, defaulting to None if no audio stream is present
+    return {
+        "has_audio": True,
+        "audio.channels": audio_stream_info.get("channels", None),
+        "audio.codec": audio_stream_info.get("codec_name", None),
+        "audio.bit_rate": int(audio_stream_info["bit_rate"]) if audio_stream_info.get("bit_rate") else None,
+        "audio.sample_rate": int(audio_stream_info["sample_rate"])
+        if audio_stream_info.get("sample_rate")
+        else None,
+        "audio.bit_depth": audio_stream_info.get("bit_depth", None),
+        "audio.channel_layout": audio_stream_info.get("channel_layout", None),
+    }
+
+
+def get_video_info(video_path: Path | str) -> dict:
+    ffprobe_video_cmd = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-select_streams",
+        "v:0",
+        "-show_entries",
+        "stream=r_frame_rate,width,height,codec_name,nb_frames,duration,pix_fmt",
+        "-of",
+        "json",
+        str(video_path),
+    ]
+    result = subprocess.run(ffprobe_video_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if result.returncode != 0:
+        raise RuntimeError(f"Error running ffprobe: {result.stderr}")
+
+    info = json.loads(result.stdout)
+    video_stream_info = info["streams"][0]
+
+    # Calculate fps from r_frame_rate
+    r_frame_rate = video_stream_info["r_frame_rate"]
+    num, denom = map(int, r_frame_rate.split("/"))
+    fps = num / denom
+
+    pixel_channels = get_video_pixel_channels(video_stream_info["pix_fmt"])
+
+    video_info = {
+        "video.fps": fps,
+        "video.height": video_stream_info["height"],
+        "video.width": video_stream_info["width"],
+        "video.channels": pixel_channels,
+        "video.codec": video_stream_info["codec_name"],
+        "video.pix_fmt": video_stream_info["pix_fmt"],
+        "video.is_depth_map": False,
+        **get_audio_info(video_path),
+    }
+
+    return video_info
+
+
+def get_video_pixel_channels(pix_fmt: str) -> int:
+    if "gray" in pix_fmt or "depth" in pix_fmt or "monochrome" in pix_fmt:
+        return 1
+    elif "rgba" in pix_fmt or "yuva" in pix_fmt:
+        return 4
+    elif "rgb" in pix_fmt or "yuv" in pix_fmt:
+        return 3
+    else:
+        raise ValueError("Unknown format")
+
+
+def get_image_pixel_channels(image: Image):
+    if image.mode == "L":
+        return 1  # Grayscale
+    elif image.mode == "LA":
+        return 2  # Grayscale + Alpha
+    elif image.mode == "RGB":
+        return 3  # RGB
+    elif image.mode == "RGBA":
+        return 4  # RGBA
+    else:
+        raise ValueError("Unknown format")

lerobot/common/envs/factory.py

@@ -1,20 +1,34 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 import importlib
 
 import gymnasium as gym
+from omegaconf import DictConfig
 
 
-def make_env(cfg, num_parallel_envs=0) -> gym.Env | gym.vector.SyncVectorEnv:
+def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv | None:
+    """Makes a gym vector environment according to the evaluation config.
+
+    n_envs can be used to override eval.batch_size in the configuration. Must be at least 1.
     """
-    Note: When `num_parallel_envs > 0`, this function returns a `SyncVectorEnv` which takes batched action as input and
-    returns batched observation, reward, terminated, truncated of `num_parallel_envs` items.
-    """
-    kwargs = {
-        "obs_type": "pixels_agent_pos",
-        "render_mode": "rgb_array",
-        "max_episode_steps": cfg.env.episode_length,
-        "visualization_width": 384,
-        "visualization_height": 384,
-    }
+    if n_envs is not None and n_envs < 1:
+        raise ValueError("`n_envs must be at least 1")
+
+    if cfg.env.name == "real_world":
+        return
 
     package_name = f"gym_{cfg.env.name}"
 
@@ -27,17 +41,18 @@ def make_env(cfg, num_parallel_envs=0) -> gym.Env | gym.vector.SyncVectorEnv:
         raise e
 
     gym_handle = f"{package_name}/{cfg.env.task}"
+    gym_kwgs = dict(cfg.env.get("gym", {}))
 
-    if num_parallel_envs == 0:
-        # non-batched version of the env that returns an observation of shape (c)
-        env = gym.make(gym_handle, disable_env_checker=True, **kwargs)
-    else:
-        # batched version of the env that returns an observation of shape (b, c)
-        env = gym.vector.SyncVectorEnv(
-            [
-                lambda: gym.make(gym_handle, disable_env_checker=True, **kwargs)
-                for _ in range(num_parallel_envs)
-            ]
-        )
+    if cfg.env.get("episode_length"):
+        gym_kwgs["max_episode_steps"] = cfg.env.episode_length
+
+    # batched version of the env that returns an observation of shape (b, c)
+    env_cls = gym.vector.AsyncVectorEnv if cfg.eval.use_async_envs else gym.vector.SyncVectorEnv
+    env = env_cls(
+        [
+            lambda: gym.make(gym_handle, disable_env_checker=True, **gym_kwgs)
+            for _ in range(n_envs if n_envs is not None else cfg.eval.batch_size)
+        ]
+    )
 
     return env

lerobot/common/envs/utils.py

@@ -1,42 +1,62 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 import einops
+import numpy as np
 import torch
+from torch import Tensor
 
 
-def preprocess_observation(observation):
+def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
+    """Convert environment observation to LeRobot format observation.
+    Args:
+        observation: Dictionary of observation batches from a Gym vector environment.
+    Returns:
+        Dictionary of observation batches with keys renamed to LeRobot format and values as tensors.
+    """
     # map to expected inputs for the policy
-    obs = {}
+    return_observations = {}
+    if "pixels" in observations:
+        if isinstance(observations["pixels"], dict):
+            imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
+        else:
+            imgs = {"observation.image": observations["pixels"]}
 
-    if isinstance(observation["pixels"], dict):
-        imgs = {f"observation.images.{key}": img for key, img in observation["pixels"].items()}
-    else:
-        imgs = {"observation.image": observation["pixels"]}
+        for imgkey, img in imgs.items():
+            img = torch.from_numpy(img)
 
-    for imgkey, img in imgs.items():
-        img = torch.from_numpy(img)
+            # sanity check that images are channel last
+            _, h, w, c = img.shape
+            assert c < h and c < w, f"expect channel last images, but instead got {img.shape=}"
 
-        # sanity check that images are channel last
-        _, h, w, c = img.shape
-        assert c < h and c < w, f"expect channel first images, but instead {img.shape}"
+            # sanity check that images are uint8
+            assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
 
-        # sanity check that images are uint8
-        assert img.dtype == torch.uint8, f"expect torch.uint8, but instead {img.dtype=}"
+            # convert to channel first of type float32 in range [0,1]
+            img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
+            img = img.type(torch.float32)
+            img /= 255
 
-        # convert to channel first of type float32 in range [0,1]
-        img = einops.rearrange(img, "b h w c -> b c h w").contiguous()
-        img = img.type(torch.float32)
-        img /= 255
+            return_observations[imgkey] = img
 
-        obs[imgkey] = img
+    if "environment_state" in observations:
+        return_observations["observation.environment_state"] = torch.from_numpy(
+            observations["environment_state"]
+        ).float()
 
-    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing requirement for "agent_pos"
-    obs["observation.state"] = torch.from_numpy(observation["agent_pos"]).float()
-
-    return obs
-
-
-def postprocess_action(action):
-    action = action.to("cpu").numpy()
-    assert (
-        action.ndim == 2
-    ), "we assume dimensions are respectively the number of parallel envs, action dimensions"
-    return action
+    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
+    # requirement for "agent_pos"
+    return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
+    return return_observations