WIP after Francesco discussion

switch back to set_transform instead of set_format

Add video_files_size_in_mb

pre-commit run --all-files

.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"
+    ]
+}

.gitattributes

@@ -11,11 +11,10 @@
 # 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.
+
 *.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
-tests/artifacts/cameras/*.png filter=lfs diff=lfs merge=lfs -text
-*.bag filter=lfs diff=lfs merge=lfs -text

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

@@ -40,24 +40,24 @@ jobs:
           git lfs install
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@b5ca514318bd6ebac0fb2aedd5d36ec1b5c232a2 # v3.10.0
+        uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
 
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
 
       - name: Login to DockerHub
-        uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
 
       - name: Build and Push CPU
-        uses: docker/build-push-action@ca052bb54ab0790a636c9b5f226502c73d547a25 # v5.4.0
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/lerobot-cpu/Dockerfile
@@ -78,24 +78,24 @@ jobs:
           git lfs install
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@b5ca514318bd6ebac0fb2aedd5d36ec1b5c232a2 # v3.10.0
+        uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
 
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           lfs: true
           persist-credentials: false
 
       - name: Login to DockerHub
-        uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
 
       - name: Build and Push GPU
-        uses: docker/build-push-action@ca052bb54ab0790a636c9b5f226502c73d547a25 # v5.4.0
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/lerobot-gpu/Dockerfile
@@ -110,23 +110,23 @@ jobs:
       group: aws-general-8-plus
     steps:
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@b5ca514318bd6ebac0fb2aedd5d36ec1b5c232a2 # v3.10.0
+        uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
 
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
       - name: Login to DockerHub
-        uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
+        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@ca052bb54ab0790a636c9b5f226502c73d547a25 # v5.4.0
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/lerobot-gpu-dev/Dockerfile

.github/workflows/build_documentation.yml

@@ -1,23 +0,0 @@
-name: Build documentation
-
-on:
-  workflow_dispatch:
-  push:
-    paths:
-      - "docs/**"
-    branches:
-    - main
-    - doc-builder*
-    - v*-release
-
-
-jobs:
-  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
-    uses: huggingface/doc-builder/.github/workflows/build_main_documentation.yml@main
-    with:
-      commit_sha: ${{ github.sha }}
-      package: lerobot
-      additional_args: --not_python_module
-    secrets:
-      token: ${{ secrets.HUGGINGFACE_PUSH }}
-      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -1,19 +0,0 @@
-name: Build PR Documentation
-
-on:
-  pull_request:
-    paths:
-      - "docs/**"
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
-    uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
-    with:
-      commit_sha: ${{ github.event.pull_request.head.sha }}
-      pr_number: ${{ github.event.number }}
-      package: lerobot
-      additional_args: --not_python_module

.github/workflows/nightly-tests.yml

@@ -33,7 +33,7 @@ jobs:
     runs-on:
       group: aws-general-8-plus
     container:
-      image: huggingface/lerobot-cpu:latest  # zizmor: ignore[unpinned-images]
+      image: huggingface/lerobot-cpu:latest
       options: --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_USERNAME }}
@@ -60,7 +60,7 @@ jobs:
       CUDA_VISIBLE_DEVICES: "0"
       TEST_TYPE: "single_gpu"
     container:
-      image: huggingface/lerobot-gpu:latest  # zizmor: ignore[unpinned-images]
+      image: huggingface/lerobot-gpu:latest
       options: --gpus all --shm-size "16gb"
       credentials:
         username: ${{ secrets.DOCKERHUB_USERNAME }}

.github/workflows/quality.yml

@@ -33,12 +33,12 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout Repository
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
       - name: Set up Python
-        uses: actions/setup-python@7f4fc3e22c37d6ff65e88745f38bd3157c663f7c # v4.9.1
+        uses: actions/setup-python@v4
         with:
           python-version: ${{ env.PYTHON_VERSION }}
 
@@ -64,9 +64,9 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout Repository
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
       - name: typos-action
-        uses: crate-ci/typos@db35ee91e80fbb447f33b0e5fbddb24d2a1a884f # v1.29.10
+        uses: crate-ci/typos@v1.29.10

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

@@ -35,7 +35,7 @@ jobs:
       matrix: ${{ steps.set-matrix.outputs.matrix }}
     steps:
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
@@ -64,17 +64,17 @@ jobs:
         docker-file: ${{ fromJson(needs.get_changed_files.outputs.matrix) }}
     steps:
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@b5ca514318bd6ebac0fb2aedd5d36ec1b5c232a2 # v3.10.0
+        uses: docker/setup-buildx-action@v3
         with:
           cache-binary: false
 
       - name: Check out code
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        uses: actions/checkout@v4
         with:
           persist-credentials: false
 
       - name: Build Docker image
-        uses: docker/build-push-action@ca052bb54ab0790a636c9b5f226502c73d547a25 # v5.4.0
+        uses: docker/build-push-action@v5
         with:
           file: ${{ matrix.docker-file }}
           context: .

.github/workflows/test.yml

@@ -50,7 +50,7 @@ jobs:
     env:
       MUJOCO_GL: egl
     steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - uses: actions/checkout@v4
         with:
           lfs: true  # Ensure LFS files are pulled
           persist-credentials: false
@@ -62,7 +62,7 @@ jobs:
           sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
 
       - name: Install uv and python
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5.4.2
+        uses: astral-sh/setup-uv@v5
         with:
           enable-cache: true
           version: ${{ env.UV_VERSION }}
@@ -85,7 +85,7 @@ jobs:
     env:
       MUJOCO_GL: egl
     steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - uses: actions/checkout@v4
         with:
           lfs: true  # Ensure LFS files are pulled
           persist-credentials: false
@@ -94,7 +94,7 @@ jobs:
         run: sudo apt-get update && sudo apt-get install -y ffmpeg
 
       - name: Install uv and python
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5.4.2
+        uses: astral-sh/setup-uv@v5
         with:
           enable-cache: true
           version: ${{ env.UV_VERSION }}
@@ -117,7 +117,7 @@ jobs:
     env:
       MUJOCO_GL: egl
     steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      - uses: actions/checkout@v4
         with:
           lfs: true  # Ensure LFS files are pulled
           persist-credentials: false
@@ -129,7 +129,7 @@ jobs:
           sudo apt-get install -y libegl1-mesa-dev ffmpeg portaudio19-dev
 
       - name: Install uv and python
-        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5.4.2
+        uses: astral-sh/setup-uv@v5
         with:
           enable-cache: true
           version: ${{ env.UV_VERSION }}

.github/workflows/trufflehog.yml

@@ -24,12 +24,12 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - name: Checkout code
-      uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      uses: actions/checkout@v4
       with:
         fetch-depth: 0
         persist-credentials: false
 
     - name: Secret Scanning
-      uses: trufflesecurity/trufflehog@90694bf9af66e7536abc5824e7a87246dbf933cb # v3.88.35
+      uses: trufflesecurity/trufflehog@main
       with:
         extra_args: --only-verified

.github/workflows/upload_pr_documentation.yml

@@ -1,16 +0,0 @@
-name: Upload PR Documentation
-
-on: # zizmor: ignore[dangerous-triggers] We follow the same pattern as in Transformers
-  workflow_run:
-    workflows: [ "Build PR Documentation" ]
-    types:
-    - completed
-
-jobs:
-  build:  # zizmor: ignore[excessive-permissions] We follow the same pattern as in Transformers
-    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@main
-    with:
-      package_name: lerobot
-    secrets:
-      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
-      comment_bot_token: ${{ secrets.COMMENT_BOT_TOKEN }}

.gitignore

@@ -12,9 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# Dev scripts
-.dev
-
 # Logging
 logs
 tmp
@@ -29,7 +26,6 @@ outputs
 
 # VS Code
 .vscode
-.devcontainer
 
 # HPC
 nautilus/*.yaml
@@ -95,8 +91,10 @@ coverage.xml
 .hypothesis/
 .pytest_cache/
 
-# Ignore .cache
+# Ignore .cache except calibration
 .cache/*
+!.cache/calibration/
+!.cache/calibration/**
 
 # Translations
 *.mo

.pre-commit-config.yaml

@@ -37,17 +37,18 @@ repos:
       - id: trailing-whitespace
 
   - repo: https://github.com/adhtruong/mirrors-typos
-    rev: v1.32.0
+    rev: v1.31.1
     hooks:
       - id: typos
         args: [--force-exclude]
 
   - repo: https://github.com/asottile/pyupgrade
-    rev: v3.20.0
+    rev: v3.19.1
     hooks:
     -   id: pyupgrade
+
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.11.11
+    rev: v0.11.5
     hooks:
       - id: ruff
         args: [--fix]
@@ -56,12 +57,12 @@ repos:
 
   ##### Security #####
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.26.0
+    rev: v8.24.3
     hooks:
       - id: gitleaks
 
   - repo: https://github.com/woodruffw/zizmor-pre-commit
-    rev: v1.8.0
+    rev: v1.5.2
     hooks:
       - id: zizmor
 

CONTRIBUTING.md

@@ -269,6 +269,9 @@ Follow these steps to start contributing:
    the PR as a draft PR. These are useful to avoid duplicated work, and to differentiate
    it from PRs ready to be merged;
 4. Make sure existing tests pass;
+<!-- 5. Add high-coverage tests. No quality testing = no merge.
+
+See an example of a good PR here: https://github.com/huggingface/lerobot/pull/ -->
 
 ### Tests
 

README.md

@@ -23,35 +23,21 @@
 </div>
 
 <h2 align="center">
-    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/12_use_so101.md">
-        Build Your Own SO-101 Robot!</a></p>
+    <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
+        Build Your Own SO-100 Robot!</a></p>
 </h2>
 
 <div align="center">
-  <div style="display: flex; gap: 1rem; justify-content: center; align-items: center;" >
-    <img
-      src="media/so101/so101.webp?raw=true"
-      alt="SO-101 follower arm"
-      title="SO-101 follower arm"
-      style="width: 40%;"
-    />
-    <img
-      src="media/so101/so101-leader.webp?raw=true"
-      alt="SO-101 leader arm"
-      title="SO-101 leader arm"
-      style="width: 40%;"
-    />
-  </div>
+  <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><strong>Meet the updated SO100, the SO-101 – Just €114 per arm!</strong></p>
+  <p><strong>Meet the SO-100 – Just $110 per arm!</strong></p>
   <p>Train it in minutes with a few simple moves on your laptop.</p>
   <p>Then sit back and watch your creation act autonomously! 🤯</p>
 
-  <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/12_use_so101.md">
-      See the full SO-101 tutorial here.</a></p>
+  <p><a href="https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md">
+      Get the full SO-100 tutorial here.</a></p>
 
-  <p>Want to take it to the next level? Make your SO-101 mobile by building LeKiwi!</p>
+  <p>Want to take it to the next level? Make your SO-100 mobile by building LeKiwi!</p>
   <p>Check out the <a href="https://github.com/huggingface/lerobot/blob/main/examples/11_use_lekiwi.md">LeKiwi tutorial</a> and bring your robot to life on wheels.</p>
 
   <img src="media/lekiwi/kiwi.webp?raw=true" alt="LeKiwi mobile robot" title="LeKiwi mobile robot" width="50%">
@@ -65,6 +51,7 @@
 
 ---
 
+
 🤗 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.
@@ -211,6 +198,7 @@ Under the hood, the `LeRobotDataset` format makes use of several ways to seriali
 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:
 
 ```
+TODO: IMPROVE
 dataset attributes:
   ├ hf_dataset: a Hugging Face dataset (backed by Arrow/parquet). Typical features example:
   │  ├ observation.images.cam_high (VideoFrame):
@@ -221,9 +209,9 @@ dataset attributes:
   │  ├ 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 an episode ; True for the last frame in each 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
+  ├ meta: 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
@@ -270,7 +258,7 @@ See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-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.
+Check out [example 3](./examples/3_train_policy.py) that illustrate 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.
 
 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 `--wandb.enable=true`.
 
@@ -321,7 +309,7 @@ Once you have trained a policy you may upload it to the Hugging Face hub using a
 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.
-- `train_config.json`: A consolidated configuration containing all parameters used for training. The policy configuration should match `config.json` exactly. This is useful for anyone who wants to evaluate your policy or for reproducibility.
+- `train_config.json`: A consolidated configuration containing all parameter userd for training. The policy configuration should match `config.json` exactly. Thisis useful for anyone who wants to evaluate your policy or for reproducibility.
 
 To upload these to the hub, run the following:
 ```bash
@@ -360,7 +348,7 @@ with profile(
 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 Palma, Steven and Kooijmans, Pepijn and Aractingi, Michel and Shukor, Mustafa and Aubakirova, Dana and Russi, Martino and Capuano, Francesco and Pascale, Caroline and Choghari, Jade and Moss, Jess and Wolf, Thomas},
+    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}
@@ -408,19 +396,6 @@ Additionally, if you are using any of the particular policy architecture, pretra
   year={2024}
 }
 ```
-
-
-- [HIL-SERL](https://hil-serl.github.io/)
-```bibtex
-@Article{luo2024hilserl,
-title={Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning},
-author={Jianlan Luo and Charles Xu and Jeffrey Wu and Sergey Levine},
-year={2024},
-eprint={2410.21845},
-archivePrefix={arXiv},
-primaryClass={cs.RO}
-}
-```
 ## Star History
 
 [![Star History Chart](https://api.star-history.com/svg?repos=huggingface/lerobot&type=Timeline)](https://star-history.com/#huggingface/lerobot&Timeline)

benchmarks/video/run_video_benchmark.py

@@ -108,7 +108,8 @@ def save_decoded_frames(
 
 
 def save_first_episode(imgs_dir: Path, dataset: LeRobotDataset) -> None:
-    ep_num_images = dataset.episode_data_index["to"][0].item()
+    episode_index = 0
+    ep_num_images = dataset.meta.episodes["length"][episode_index]
     if imgs_dir.exists() and len(list(imgs_dir.glob("frame_*.png"))) == ep_num_images:
         return
 
@@ -265,7 +266,8 @@ def benchmark_encoding_decoding(
             overwrite=True,
         )
 
-    ep_num_images = dataset.episode_data_index["to"][0].item()
+    episode_index = 0
+    ep_num_images = dataset.meta.episodes["length"][episode_index]
     width, height = tuple(dataset[0][dataset.meta.camera_keys[0]].shape[-2:])
     num_pixels = width * height
     video_size_bytes = video_path.stat().st_size
@@ -416,7 +418,7 @@ if __name__ == "__main__":
         "--vcodec",
         type=str,
         nargs="*",
-        default=["libx264", "hevc", "libsvtav1"],
+        default=["libx264", "libx265", "libsvtav1"],
         help="Video codecs to be tested",
     )
     parser.add_argument(
@@ -446,7 +448,7 @@ if __name__ == "__main__":
     #     nargs="*",
     #     default=[0, 1],
     #     help="Use the fastdecode tuning option. 0 disables it. "
-    #         "For libx264 and libx265/hevc, only 1 is possible. "
+    #         "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(

docs/README.md

@@ -1,137 +0,0 @@
-<!---
-Copyright 2020 The HuggingFace 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.
--->
-
-# Generating the documentation
-
-To generate the documentation, you first have to build it. Several packages are necessary to build the doc,
-you can install them with the following command, at the root of the code repository:
-
-```bash
-pip install -e ".[docs]"
-```
-
-You will also need `nodejs`. Please refer to their [installation page](https://nodejs.org/en/download)
-
----
-**NOTE**
-
-You only need to generate the documentation to inspect it locally (if you're planning changes and want to
-check how they look before committing for instance). You don't have to `git commit` the built documentation.
-
----
-
-## Building the documentation
-
-Once you have setup the `doc-builder` and additional packages, you can generate the documentation by
-typing the following command:
-
-```bash
-doc-builder build lerobot docs/source/ --build_dir ~/tmp/test-build
-```
-
-You can adapt the `--build_dir` to set any temporary folder that you prefer. This command will create it and generate
-the MDX files that will be rendered as the documentation on the main website. You can inspect them in your favorite
-Markdown editor.
-
-## Previewing the documentation
-
-To preview the docs, first install the `watchdog` module with:
-
-```bash
-pip install watchdog
-```
-
-Then run the following command:
-
-```bash
-doc-builder preview lerobot docs/source/
-```
-
-The docs will be viewable at [http://localhost:3000](http://localhost:3000). You can also preview the docs once you have opened a PR. You will see a bot add a comment to a link where the documentation with your changes lives.
-
----
-**NOTE**
-
-The `preview` command only works with existing doc files. When you add a completely new file, you need to update `_toctree.yml` & restart `preview` command (`ctrl-c` to stop it & call `doc-builder preview ...` again).
-
----
-
-## Adding a new element to the navigation bar
-
-Accepted files are Markdown (.md).
-
-Create a file with its extension and put it in the source directory. You can then link it to the toc-tree by putting
-the filename without the extension in the [`_toctree.yml`](https://github.com/huggingface/lerobot/blob/main/docs/source/_toctree.yml) file.
-
-## Renaming section headers and moving sections
-
-It helps to keep the old links working when renaming the section header and/or moving sections from one document to another. This is because the old links are likely to be used in Issues, Forums, and Social media and it'd make for a much more superior user experience if users reading those months later could still easily navigate to the originally intended information.
-
-Therefore, we simply keep a little map of moved sections at the end of the document where the original section was. The key is to preserve the original anchor.
-
-So if you renamed a section from: "Section A" to "Section B", then you can add at the end of the file:
-
-```
-Sections that were moved:
-
-[ <a href="#section-b">Section A</a><a id="section-a"></a> ]
-```
-and of course, if you moved it to another file, then:
-
-```
-Sections that were moved:
-
-[ <a href="../new-file#section-b">Section A</a><a id="section-a"></a> ]
-```
-
-Use the relative style to link to the new file so that the versioned docs continue to work.
-
-For an example of a rich moved sections set please see the very end of [the transformers Trainer doc](https://github.com/huggingface/transformers/blob/main/docs/source/en/main_classes/trainer.md).
-
-### Adding a new tutorial
-
-Adding a new tutorial or section is done in two steps:
-
-- Add a new file under `./source`. This file can either be ReStructuredText (.rst) or Markdown (.md).
-- Link that file in `./source/_toctree.yml` on the correct toc-tree.
-
-Make sure to put your new file under the proper section. If you have a doubt, feel free to ask in a Github Issue or PR.
-
-### Writing source documentation
-
-Values that should be put in `code` should either be surrounded by backticks: \`like so\`. Note that argument names
-and objects like True, None or any strings should usually be put in `code`.
-
-#### Writing a multi-line code block
-
-Multi-line code blocks can be useful for displaying examples. They are done between two lines of three backticks as usual in Markdown:
-
-
-````
-```
-# first line of code
-# second line
-# etc
-```
-````
-
-#### Adding an image
-
-Due to the rapidly growing repository, it is important to make sure that no files that would significantly weigh down the repository are added. This includes images, videos, and other non-text files. We prefer to leverage a hf.co hosted `dataset` like
-the ones hosted on [`hf-internal-testing`](https://huggingface.co/hf-internal-testing) in which to place these files and reference
-them by URL. We recommend putting them in the following dataset: [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images).
-If an external contribution, feel free to add the images to your PR and ask a Hugging Face member to migrate your images
-to this dataset.

docs/source/_toctree.yml

@@ -1,28 +0,0 @@
-- sections:
-  - local: index
-    title: LeRobot
-  - local: installation
-    title: Installation
-  title: Get started
-- sections:
-  - local: getting_started_real_world_robot
-    title: Getting Started with Real-World Robots
-  - local: cameras
-    title: Cameras
-  - local: hilserl
-    title: Getting Started with Reinforcement Learning
-  title: "Tutorials"
-- sections:
-  - local: so101
-    title: SO-101
-  - local: so100
-    title: SO-100
-  - local: koch
-    title: Koch v1.1
-  - local: lekiwi
-    title: LeKiwi
-  title: "Robots"
-- sections:
-  - local: contributing
-    title: Contribute to LeRobot
-  title: "Contribute"

docs/source/cameras.mdx

@@ -1,173 +0,0 @@
-# Cameras
-
-LeRobot offers multiple options for video capture, including phone cameras, built-in laptop cameras, external webcams, and Intel RealSense cameras. To efficiently record frames from most cameras, you can use either the `OpenCVCamera` or `RealSenseCamera` class. For additional compatibility details on the `OpenCVCamera` class, refer to the [Video I/O with OpenCV Overview](https://docs.opencv.org/4.x/d0/da7/videoio_overview.html).
-
-### Finding your camera
-
-To instantiate a camera, you need a camera identifier. This identifier might change if you reboot your computer or re-plug your camera, a behavior mostly dependant on your operating system.
-
-To find the camera indices of the cameras plugged into your system, run the following script:
-```bash
-python lerobot/find_cameras.py opencv # or realsense for Intel Realsense cameras
-```
-
-The output will look something like this if you have two cameras connected:
-```
---- Detected Cameras ---
-Camera #0:
-  Name: OpenCV Camera @ 0
-  Type: OpenCV
-  Id: 0
-  Backend api: AVFOUNDATION
-  Default stream profile:
-    Format: 16.0
-    Width: 1920
-    Height: 1080
-    Fps: 15.0
---------------------
-(more cameras ...)
-```
-
-> [!WARNING]
-> When using Intel RealSense cameras in `macOS`, you could get this [error](https://github.com/IntelRealSense/librealsense/issues/12307): `Error finding RealSense cameras: failed to set power state`, this can be solved by running the same command with `sudo` permissions. Note that using RealSense cameras in `macOS` is unstable.
-
-
-## Use Cameras
-
-Below are two examples, demonstrating how to work with the API.
-
-- **Asynchronous frame capture** using an OpenCV-based camera
-- **Color and depth capture** using an Intel RealSense camera
-
-
-<hfoptions id="shell_restart">
-<hfoption id="Open CV Camera">
-
-```python
-from lerobot.common.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.common.cameras.opencv.camera_opencv import OpenCVCamera
-from lerobot.common.cameras.configs import ColorMode, Cv2Rotation
-
-# Construct an `OpenCVCameraConfig` with your desired FPS, resolution, color mode, and rotation.
-config = OpenCVCameraConfig(
-    index_or_path=0,
-    fps=15,
-    width=1920,
-    height=1080,
-    color_mode=ColorMode.RGB,
-    rotation=Cv2Rotation.NO_ROTATION
-)
-
-# Instantiate and connect an `OpenCVCamera`, performing a warm-up read (default).
-camera = OpenCVCamera(config)
-camera.connect()
-
-# Read frames asynchronously in a loop via `async_read(timeout_ms)`
-try:
-    for i in range(10):
-        frame = camera.async_read(timeout_ms=200)
-        print(f"Async frame {i} shape:", frame.shape)
-finally:
-    camera.disconnect()
-```
-
-</hfoption>
-<hfoption id="Intel Realsense Camera">
-
-```python
-from lerobot.common.cameras.intel.configuration_realsense import RealSenseCameraConfig
-from lerobot.common.cameras.intel.camera_realsense import RealSenseCamera
-from lerobot.common.cameras.configs import ColorMode, Cv2Rotation
-
-# Create a `RealSenseCameraConfig` specifying your camera’s serial number and enabling depth.
-config = RealSenseCameraConfig(
-    serial_number="233522074606",
-    fps=15,
-    width=640,
-    height=480,
-    color_mode=ColorMode.RGB,
-    use_depth=True,
-    rotation=Cv2Rotation.NO_ROTATION
-)
-
-# Instantiate and connect a `RealSenseCamera` with warm-up read (default).
-camera = RealSenseCamera(config)
-camera.connect()
-
-# Capture a color frame via `read()` and a depth map via `read_depth()`.
-try:
-    color_frame = camera.read()
-    depth_map = camera.read_depth()
-    print("Color frame shape:", color_frame.shape)
-    print("Depth map shape:", depth_map.shape)
-finally:
-    camera.disconnect()
-```
-</hfoption>
-</hfoptions>
-
-
-## Use your phone
-<hfoptions id="use phone">
-<hfoption id="Mac">
-
-To use your iPhone as a camera on macOS, enable the Continuity Camera feature:
-- Ensure your Mac is running macOS 13 or later, and your iPhone is on iOS 16 or later.
-- Sign in both devices with the same Apple ID.
-- Connect your devices with a USB cable or turn on Wi-Fi and Bluetooth for a wireless connection.
-
-For more details, visit [Apple support](https://support.apple.com/en-gb/guide/mac-help/mchl77879b8a/mac).
-
-Your iPhone should be detected automatically when running the camera setup script in the next section.
-
-</hfoption>
-<hfoption id="Linux">
-
-If you want to use your phone as a camera on Linux, follow these steps to set up a virtual camera
-
-1. *Install `v4l2loopback-dkms` and `v4l-utils`*. Those packages are required to create virtual camera devices (`v4l2loopback`) and verify their settings with the `v4l2-ctl` utility from `v4l-utils`. Install them using:
-```python
-sudo apt install v4l2loopback-dkms v4l-utils
-```
-2. *Install [DroidCam](https://droidcam.app) on your phone*. This app is available for both iOS and Android.
-3. *Install [OBS Studio](https://obsproject.com)*. This software will help you manage the camera feed. Install it using [Flatpak](https://flatpak.org):
-```python
-flatpak install flathub com.obsproject.Studio
-```
-4. *Install the DroidCam OBS plugin*. This plugin integrates DroidCam with OBS Studio. Install it with:
-```python
-flatpak install flathub com.obsproject.Studio.Plugin.DroidCam
-```
-5. *Start OBS Studio*. Launch with:
-```python
-flatpak run com.obsproject.Studio
-```
-6. *Add your phone as a source*. Follow the instructions [here](https://droidcam.app/obs/usage). Be sure to set the resolution to `640x480`.
-7. *Adjust resolution settings*. In OBS Studio, go to `File > Settings > Video`. Change the `Base(Canvas) Resolution` and the `Output(Scaled) Resolution` to `640x480` by manually typing it in.
-8. *Start virtual camera*. In OBS Studio, follow the instructions [here](https://obsproject.com/kb/virtual-camera-guide).
-9. *Verify the virtual camera setup*. Use `v4l2-ctl` to list the devices:
-```python
-v4l2-ctl --list-devices
-```
-You should see an entry like:
-```
-VirtualCam (platform:v4l2loopback-000):
-/dev/video1
-```
-10. *Check the camera resolution*. Use `v4l2-ctl` to ensure that the virtual camera output resolution is `640x480`. Change `/dev/video1` to the port of your virtual camera from the output of `v4l2-ctl --list-devices`.
-```python
-v4l2-ctl -d /dev/video1 --get-fmt-video
-```
-You should see an entry like:
-```
->>> Format Video Capture:
->>>	Width/Height      : 640/480
->>>	Pixel Format      : 'YUYV' (YUYV 4:2:2)
-```
-
-Troubleshooting: If the resolution is not correct you will have to delete the Virtual Camera port and try again as it cannot be changed.
-
-If everything is set up correctly, you can proceed with the rest of the tutorial.
-
-</hfoption>
-</hfoptions>

docs/source/contributing.md

@@ -1 +0,0 @@
-../../CONTRIBUTING.md

docs/source/getting_started_real_world_robot.mdx

@@ -1,321 +0,0 @@
-# Getting Started with Real-World Robots
-
-This tutorial will explain how to train a neural network to control a real robot autonomously.
-
-**You'll learn:**
-1. How to record and visualize your dataset.
-2. How to train a policy using your data and prepare it for evaluation.
-3. How to evaluate your policy and visualize the results.
-
-By following these steps, you'll be able to replicate tasks, such as picking up a Lego block and placing it in a bin with a high success rate, as shown in the video below.
-
-<details>
-<summary><strong>Video: pickup lego block task</strong></summary>
-
-<div class="video-container">
-  <video controls width="600">
-    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot_task.mp4" type="video/mp4" />
-  </video>
-</div>
-
-</details>
-
-This tutorial isn’t tied to a specific robot: we walk you through the commands and API snippets you can adapt for any supported platform.
-
-During data collection, you’ll use a “teloperation” device, such as a leader arm or keyboard to teleoperate the robot and record its motion trajectories.
-
-Once you’ve gathered enough trajectories, you’ll train a neural network to imitate these trajectories and deploy the trained model so your robot can perform the task autonomously.
-
-If you run into any issues at any point, jump into our [Discord community](https://discord.com/invite/s3KuuzsPFb) for support.
-
-## Set up and Calibrate
-
-If you haven't yet set up and calibrated your robot and teleop device, please do so by following the robot-specific tutorial.
-
-## Teleoperate
-
-In this example, we’ll demonstrate how to teleoperate the SO101 robot. For each command, we also provide a corresponding API example.
-
-<hfoptions id="teleoperate_so101">
-<hfoption id="Command">
-```bash
-python -m lerobot.teleoperate \
-    --robot.type=so101_follower \
-    --robot.port=/dev/tty.usbmodem58760431541 \
-    --robot.id=my_red_robot_arm \
-    --teleop.type=so101_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=my_blue_leader_arm
-```
-</hfoption>
-<hfoption id="API example">
-```python
-from lerobot.common.teleoperators.so101_leader import SO101LeaderConfig, SO101Leader
-from lerobot.common.robots.so101_follower import SO101FollowerConfig, SO101Follower
-
-robot_config = SO101FollowerConfig(
-    port="/dev/tty.usbmodem58760431541",
-    id="my_red_robot_arm",
-)
-
-teleop_config = SO101LeaderConfig(
-    port="/dev/tty.usbmodem58760431551",
-    id="my_blue_leader_arm",
-)
-
-robot = SO101Follower(robot_config)
-teleop_device = SO101Leader(teleop_config)
-robot.connect()
-teleop_device.connect()
-
-while True:
-    action = teleop_device.get_action()
-    robot.send_action(action)
-```
-</hfoption>
-</hfoptions>
-
-The teleoperate command will automatically:
-1. Identify any missing calibrations and initiate the calibration procedure.
-2. Connect the robot and teleop device and start teleoperation.
-
-## Cameras
-
-To add cameras to your setup, follow this [Guide](./cameras#setup-cameras).
-
-## Teleoperate with cameras
-
-With `rerun`, you can teleoperate again while simultaneously visualizing the camera feeds and joint positions. In this example, we’re using the Koch arm.
-
-<hfoptions id="teleoperate_koch_camera">
-<hfoption id="Command">
-```bash
-python -m lerobot.teleoperate \
-    --robot.type=koch_follower \
-    --robot.port=/dev/tty.usbmodem58760431541 \
-    --robot.id=my_koch_robot \
-    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
-    --teleop.type=koch_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=my_koch_teleop \
-    --display_data=true
-```
-</hfoption>
-<hfoption id="API example">
-```python
-from lerobot.common.cameras.opencv.configuration_opencv import OpenCVCameraConfig
-from lerobot.common.teleoperators.koch_leader import KochLeaderConfig, KochLeader
-from lerobot.common.robots.koch_follower import KochFollowerConfig, KochFollower
-
-camera_config = {
-    "front": OpenCVCameraConfig(index_or_path=0, width=1920, height=1080, fps=30)
-}
-
-robot_config = KochFollowerConfig(
-    port="/dev/tty.usbmodem585A0076841",
-    id="my_red_robot_arm",
-    cameras=camera_config
-)
-
-teleop_config = KochLeaderConfig(
-    port="/dev/tty.usbmodem58760431551",
-    id="my_blue_leader_arm",
-)
-
-robot = KochFollower(robot_config)
-teleop_device = KochLeader(teleop_config)
-robot.connect()
-teleop_device.connect()
-
-while True:
-    observation = robot.get_observation()
-    action = teleop_device.get_action()
-    robot.send_action(action)
-```
-</hfoption>
-</hfoptions>
-
-## Record a dataset
-
-Once you're familiar with teleoperation, you can record your first dataset.
-
-We use the Hugging Face hub features for uploading your dataset. If you haven't previously used the Hub, make sure you can login via the cli using a write-access token, this token can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens).
-
-Add your token to the CLI by running this command:
-```bash
-huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
-```
-
-Then store your Hugging Face repository name in a variable:
-```bash
-HF_USER=$(huggingface-cli whoami | head -n 1)
-echo $HF_USER
-```
-
-Now you can record a dataset. To record 2 episodes and upload your dataset to the hub, execute this command tailored to the SO101.
-```bash
-python -m lerobot.record \
-    --robot.type=so101_follower \
-    --robot.port=/dev/tty.usbmodem585A0076841 \
-    --robot.id=my_red_robot_arm \
-    --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
-    --teleop.type=so101_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=my_blue_leader_arm \
-    --display_data=true \
-    --dataset.repo_id=aliberts/record-test \
-    --dataset.num_episodes=2 \
-    --dataset.single_task="Grab the black cube"
-```
-
-#### Dataset upload
-Locally, your dataset is stored in this folder: `~/.cache/huggingface/lerobot/{repo-id}`. At the end of data recording, your dataset will be uploaded on your Hugging Face page (e.g. https://huggingface.co/datasets/cadene/so101_test) that you can obtain by running:
-```bash
-echo https://huggingface.co/datasets/${HF_USER}/so101_test
-```
-Your dataset will be automatically tagged with `LeRobot` for the community to find it easily, and you can also add custom tags (in this case `tutorial` for example).
-
-You can look for other LeRobot datasets on the hub by searching for `LeRobot` [tags](https://huggingface.co/datasets?other=LeRobot).
-
-#### Record function
-
-The `record` function provides a suite of tools for capturing and managing data during robot operation:
-
-##### 1. Data Storage
-- Data is stored using the `LeRobotDataset` format and is stored on disk during recording.
-- By default, the dataset is pushed to your Hugging Face page after recording.
-  - To disable uploading, use `--dataset.push_to_hub=False`.
-
-##### 2. Checkpointing and Resuming
-- Checkpoints are automatically created during recording.
-- If an issue occurs, you can resume by re-running the same command with `--control.resume=true`.
-- To start recording from scratch, **manually delete** the dataset directory.
-
-##### 3. Recording Parameters
-Set the flow of data recording using command-line arguments:
-- `--dataset.episode_time_s=60`
-  Duration of each data recording episode (default: **60 seconds**).
-- `--dataset.reset_time_s=60`
-  Duration for resetting the environment after each episode (default: **60 seconds**).
-- `--dataset.num_episodes=50`
-  Total number of episodes to record (default: **50**).
-
-##### 4. Keyboard Controls During Recording
-Control the data recording flow using keyboard shortcuts:
-- Press **Right Arrow (`→`)**: Early stop the current episode or reset time and move to the next.
-- Press **Left Arrow (`←`)**: Cancel the current episode and re-record it.
-- Press **Escape (`ESC`)**: Immediately stop the session, encode videos, and upload the dataset.
-
-#### Tips for gathering data
-
-Once you're comfortable with data recording, you can create a larger dataset for training. A good starting task is grasping an object at different locations and placing it in a bin. We suggest recording at least 50 episodes, with 10 episodes per location. Keep the cameras fixed and maintain consistent grasping behavior throughout the recordings. Also make sure the object you are manipulating is visible on the camera's. A good rule of thumb is you should be able to do the task yourself by only looking at the camera images.
-
-In the following sections, you’ll train your neural network. After achieving reliable grasping performance, you can start introducing more variations during data collection, such as additional grasp locations, different grasping techniques, and altering camera positions.
-
-Avoid adding too much variation too quickly, as it may hinder your results.
-
-If you want to dive deeper into this important topic, you can check out the [blog post](https://huggingface.co/blog/lerobot-datasets#what-makes-a-good-dataset) we wrote on what makes a good dataset.
-
-
-#### Troubleshooting:
-- On Linux, if the left and right arrow keys and escape key don't have any effect during data recording, make sure you've set the `$DISPLAY` environment variable. See [pynput limitations](https://pynput.readthedocs.io/en/latest/limitations.html#linux).
-
-## Visualize a dataset
-
-If you uploaded your dataset to the hub with `--control.push_to_hub=true`, 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}/so101_test
-```
-
-If you didn't upload with `--control.push_to_hub=false`, you can visualize it locally with (via a window in the browser `http://127.0.0.1:9090` with the visualization tool):
-```bash
-python lerobot/scripts/visualize_dataset_html.py \
-  --repo-id ${HF_USER}/so101_test \
-  --local-files-only 1
-```
-
-This will launch a local web server that looks like this:
-<div style="text-align:center;">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/visualize_dataset_html.webp?raw=true" alt="Koch v1.1 leader and follower arms" title="Koch v1.1 leader and follower arms" width="100%"></img>
-</div>
-
-## Replay an episode
-
-A useful feature is the `replay` function, which allows you to replay any episode that you've recorded or episodes from any dataset out there. This function helps you test the repeatability of your robot's actions and assess transferability across robots of the same model.
-
-You can replay the first episode on your robot with:
-```bash
-python -m lerobot.replay \
-    --robot.type=so101_follower \
-    --robot.port=/dev/tty.usbmodem58760431541 \
-    --robot.id=black \
-    --dataset.repo_id=aliberts/record-test \
-    --dataset.episode=2
-```
-
-Your robot should replicate movements similar to those you recorded. For example, check out [this video](https://x.com/RemiCadene/status/1793654950905680090) where we use `replay` on a Aloha robot from [Trossen Robotics](https://www.trossenrobotics.com).
-
-## Train a policy
-
-To train a policy to control your robot, use the [`python lerobot/scripts/train.py`](../lerobot/scripts/train.py) script. A few arguments are required. Here is an example command:
-```bash
-python lerobot/scripts/train.py \
-  --dataset.repo_id=${HF_USER}/so101_test \
-  --policy.type=act \
-  --output_dir=outputs/train/act_so101_test \
-  --job_name=act_so101_test \
-  --policy.device=cuda \
-  --wandb.enable=true
-```
-
-Let's explain the command:
-1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so101_test`.
-2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
-4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
-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_so101_test/checkpoints`.
-
-To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so101_test` policy:
-```bash
-python lerobot/scripts/train.py \
-  --config_path=outputs/train/act_so101_test/checkpoints/last/pretrained_model/train_config.json \
-  --resume=true
-```
-
-#### Upload policy checkpoints
-
-Once training is done, upload the latest checkpoint with:
-```bash
-huggingface-cli upload ${HF_USER}/act_so101_test \
-  outputs/train/act_so101_test/checkpoints/last/pretrained_model
-```
-
-You can also upload intermediate checkpoints with:
-```bash
-CKPT=010000
-huggingface-cli upload ${HF_USER}/act_so101_test${CKPT} \
-  outputs/train/act_so101_test/checkpoints/${CKPT}/pretrained_model
-```
-
-## Evaluate your policy
-
-You can use the `record` script from [`lerobot/record.py`](https://github.com/huggingface/lerobot/blob/main/lerobot/record.py) but with a policy checkpoint as input. For instance, run this command to record 10 evaluation episodes:
-```bash
-python -m lerobot.record  \
-  --robot.type=so100_follower \
-  --robot.port=/dev/ttyACM1 \
-  --robot.cameras="{ up: {type: opencv, index_or_path: /dev/video10, width: 640, height: 480, fps: 30}, side: {type: intelrealsense, serial_number_or_name: 233522074606, width: 640, height: 480, fps: 30}}" \
-  --robot.id=blue_follower_arm \
-  --teleop.type=so100_leader \
-  --teleop.port=/dev/ttyACM0 \
-  --teleop.id=red_leader_arm \
-  --display_data=false \
-  --dataset.repo_id=$HF_USER/eval_lego_${EPOCHREALTIME/[^0-9]/} \
-  --dataset.single_task="Put lego brick into the transparent box" \
-  --policy.path=${HF_USER}/act_johns_arm
-```
-
-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 `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_so101_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so101_test`).
-2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so101_test`).

docs/source/hilserl.mdx

@@ -1,512 +0,0 @@
-# HilSerl Real Robot Training Workflow Guide
-
-Human-in-the-Loop Sample-Efficient Reinforcement Learning (HIL-SERL) with LeRobot workflow for taking a policy from “zero” to real-world robot mastery in just a couple of hours.
-It combines three ingredients:
-	1.	**Offline demonstrations & reward classifier:** a handful of human-teleop episodes plus a vision-based success detector give the policy a shaped starting point.
-	2.	**On-robot actor / learner loop with human interventions:** a distributed SAC/RLPD learner updates the policy while an actor explores on the physical robot; the human can jump in at any time to correct dangerous or unproductive behaviour.
-	3.	**Safety & efficiency tools:** joint/EE bounds, impedance control, crop-ROI preprocessing and WandB monitoring keep the data useful and the hardware safe.
-
-Together these elements let HIL-SERL reach near-perfect task success and faster cycle times than imitation-only baselines.
-
-<p align="center">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/hilserl-main-figure.png" alt="HIL-SERL workflow" title="HIL-SERL workflow" width="100%"></img>
-</p>
-
-<p align="center"><i>HIL-SERL workflow, Luo et al. 2024</i></p>
-
-This guide provides step-by-step instructions for training a robot policy using LeRobot's HilSerl implementation to train on a real robot.
-
-
-# 1. Real Robot Training Workflow
-
-## 1.1 Understanding Configuration
-
-The training process begins with proper configuration for the HILSerl environment. The configuration class of interest is `HILSerlRobotEnvConfig` in `lerobot/common/envs/configs.py`. Which is defined as:
-
-```python
-class HILSerlRobotEnvConfig(EnvConfig):
-    robot: Optional[RobotConfig] = None    # Main robot agent (defined in `lerobot/common/robots`)
-    teleop: Optional[TeleoperatorConfig] = None    # Teleoperator agent, e.g., gamepad or leader arm, (defined in `lerobot/common/teleoperators`)
-    wrapper: Optional[EnvTransformConfig] = None    # Environment wrapper settings; check `lerobot/scripts/server/gym_manipulator.py`
-    fps: int = 10    # Control frequency
-    name: str = "real_robot"    # Environment name
-    mode: str = None    # "record", "replay", or None (for training)
-    repo_id: Optional[str] = None    # LeRobot dataset repository ID
-    dataset_root: Optional[str] = None    # Local dataset root (optional)
-    task: str = ""    # Task identifier
-    num_episodes: int = 10    # Number of episodes for recording
-    episode: int = 0    # episode index for replay
-    device: str = "cuda"    # Compute device
-    push_to_hub: bool = True    # Whether to push the recorded datasets to Hub
-    pretrained_policy_name_or_path: Optional[str] = None    # For policy loading
-    reward_classifier_pretrained_path: Optional[str] = None    # For reward model
-```
-
-
-## 1.2 Finding Robot Workspace Bounds
-
-Before collecting demonstrations, you need to determine the appropriate operational bounds for your robot.
-
-This helps simplifying the problem of learning on the real robot by limiting the robot's operational space to a specific region that solves the task and avoids unnecessary or unsafe exploration.
-
-### 1.2.1 Using find_joint_limits.py
-
-This script helps you find the safe operational bounds for your robot's end-effector. Given that you have a follower and leader arm, you can use the script to find the bounds for the follower arm that will be applied during training.
-Bounding the action space will reduce the redundant exploration of the agent and guarantees safety.
-
-```bash
-python -m lerobot.scripts.find_joint_limits \
-    --robot.type=so100_follower \
-    --robot.port=/dev/tty.usbmodem58760431541 \
-    --robot.id=black \
-    --teleop.type=so100_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=blue
-```
-
-### 1.2.2 Workflow
-
-1. Run the script and move the robot through the space that solves the task
-2. The script will record the minimum and maximum end-effector positions and the joint angles and prints them to the console, for example:
-   ```
-   Max ee position [0.24170487 0.201285   0.10273342]
-   Min ee position [0.16631757 -0.08237468  0.03364977]
-   Max joint positions [-20.0, -20.0, -20.0, -20.0, -20.0, -20.0]
-   Min joint positions [50.0, 50.0, 50.0, 50.0, 50.0, 50.0]
-   ```
-3. Use these values in the configuration of you teleoperation device (TeleoperatorConfig) under the `end_effector_bounds` field
-
-### 1.2.3 Example Configuration
-
-```json
-"end_effector_bounds": {
-    "max": [0.24, 0.20, 0.10],
-    "min": [0.16, -0.08, 0.03]
-}
-```
-
-## 1.3 Collecting Demonstrations
-
-With the bounds defined, you can safely collect demonstrations for training. Training RL with off-policy algorithm allows us to use offline datasets collected in order to improve the efficiency of the learning process.
-
-### 1.3.1 Setting Up Record Mode
-
-Create a configuration file for recording demonstrations (or edit an existing one like `env_config_so100.json`):
-
-1. Set `mode` to `"record"`
-2. Specify a unique `repo_id` for your dataset (e.g., "username/task_name")
-3. Set `num_episodes` to the number of demonstrations you want to collect
-4. Set `crop_params_dict` to `null` initially (we'll determine crops later)
-5. Configure `robot`, `cameras`, and other hardware settings
-
-Example configuration section:
-```json
-"mode": "record",
-"repo_id": "username/pick_lift_cube",
-"dataset_root": null,
-"task": "pick_and_lift",
-"num_episodes": 15,
-"episode": 0,
-"push_to_hub": true
-```
-
-### 1.3.2 Gamepad Controls
-
-<p align="center">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/gamepad_guide.jpg?raw=true" alt="Figure shows the control mappings on a Logitech gamepad." title="Gamepad Control Mapping" width="100%"></img>
-</p>
-<p align="center"><i>Gamepad button mapping for robot control and episode management</i></p>
-
-
-### 1.3.3 Recording Demonstrations
-
-Start the recording process:
-
-```bash
-python lerobot/scripts/rl/gym_manipulator.py --config_path lerobot/configs/env_config_so100.json
-```
-
-During recording:
-1. The robot will reset to the initial position defined in the configuration file `fixed_reset_position`
-2. Use the gamepad to control the robot by setting `"control_mode"="gamepad"` in the configuration file
-3. Complete the task successfully
-4. The episode ends with a reward of 1 when you press the "success" button
-5. If the time limit is reached, or the fail button is pressed, the episode ends with a reward of 0
-6. You can rerecord an episode by pressing the "rerecord" button
-7. The process automatically continues to the next episode
-8. After recording all episodes, the dataset is pushed to the Hugging Face Hub (optional) and saved locally
-
-
-
-## 1.4 Processing the Dataset
-
-After collecting demonstrations, process them to determine optimal camera crops.
-Reinforcement learning is sensitive to background distractions, so it is important to crop the images to the relevant workspace area.
-Note: If you already know the crop parameters, you can skip this step and just set the `crop_params_dict` in the configuration file during recording.
-
-### 1.4.1 Determining Crop Parameters
-
-Use the `crop_dataset_roi.py` script to interactively select regions of interest in your camera images:
-
-```bash
-python lerobot/scripts/rl/crop_dataset_roi.py --repo-id username/pick_lift_cube
-```
-
-1. For each camera view, the script will display the first frame
-2. Draw a rectangle around the relevant workspace area
-3. Press 'c' to confirm the selection
-4. Repeat for all camera views
-5. The script outputs cropping parameters and creates a new cropped dataset
-
-Example output:
-```
-Selected Rectangular Regions of Interest (top, left, height, width):
-observation.images.side: [180, 207, 180, 200]
-observation.images.front: [180, 250, 120, 150]
-```
-
-<p align="center">
-<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/crop_dataset.gif" width="600"/>
-</p>
-
-<p align="center"><i>Interactive cropping tool for selecting regions of interest</i></p>
-
-
-### 1.4.2 Updating Configuration
-
-Add these crop parameters to your training configuration:
-
-```json
-"crop_params_dict": {
-    "observation.images.side": [180, 207, 180, 200],
-    "observation.images.front": [180, 250, 120, 150]
-},
-"resize_size": [128, 128]
-```
-
-## 1.5 Training with Actor-Learner
-
-The LeRobot system uses a distributed actor-learner architecture for training. You will need to start two processes: a learner and an actor.
-
-### 1.5.1 Configuration Setup
-
-Create a training configuration file (See example `train_config_hilserl_so100.json`). The training config is based on the main `TrainPipelineConfig` class in `lerobot/configs/train.py`.
-
-1. Set `mode` to `null` (for training mode)
-2. Configure the policy settings (`type`, `device`, etc.)
-3. Set `dataset` to your cropped dataset
-4. Configure environment settings with crop parameters
-5. Check the other parameters related to SAC.
-6. Verify that the `policy` config is correct with the right `input_features` and `output_features` for your task.
-
-### 1.5.2 Starting the Learner
-
-First, start the learner server process:
-
-```bash
-python lerobot/scripts/rl/learner.py --config_path lerobot/configs/train_config_hilserl_so100.json
-```
-
-The learner:
-- Initializes the policy network
-- Prepares replay buffers
-- Opens a gRPC server to communicate with actors
-- Processes transitions and updates the policy
-
-### 1.5.3 Starting the Actor
-
-In a separate terminal, start the actor process with the same configuration:
-
-```bash
-python lerobot/scripts/rl/actor.py --config_path lerobot/configs/train_config_hilserl_so100.json
-```
-
-The actor:
-- Connects to the learner via gRPC
-- Initializes the environment
-- Execute rollouts of the policy to collect experience
-- Sends transitions to the learner
-- Receives updated policy parameters
-
-### 1.5.4 Training Flow
-
-The training proceeds automatically:
-
-1. The actor executes the policy in the environment
-2. Transitions are collected and sent to the learner
-3. The learner updates the policy based on these transitions
-4. Updated policy parameters are sent back to the actor
-5. The process continues until the specified step limit is reached
-
-### 1.5.5 Human in the Loop
-
-- The key to learning efficiently is to have a human interventions to provide corrective feedback and completing the task to aide the policy learning and exploration.
-- To perform human interventions, you can press the upper right trigger button on the gamepad. This will pause the policy actions and allow you to take over.
-- A successful experiment is one where the human has to intervene at the start but then reduces the amount of interventions as the policy improves. You can monitor the intervention rate in the `wandb` dashboard.
-
-<p align="center">
-  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/hil_effect.png?raw=true" alt="Figure shows the control mappings on a Logitech gamepad." title="Gamepad Control Mapping" width="100%"></img>
-</p>
-
-<p align="center"><i>Example showing how human interventions help guide policy learning over time</i></p>
-
-- The figure shows the plot of the episodic reward over interaction step. The figure shows the effect of human interventions on the policy learning.
-- The orange curve is an experiment without any human interventions. While the pink and blue curves are experiments with human interventions.
-- We can observe that the number of steps where the policy starts achieving the maximum reward is cut by a quarter when human interventions are present.
-
-#### Guide to Human Interventions
-The strategy to follow is to intervene heavily at the start of training and then reduce the amount of interventions as the training progresses. Some tips and hints:
-- Interevene for almost the length of the entire episode at the first few episodes.
-- When the policy is less chaotic, gradually reduce the intervention time during one episode and let the policy explore for a longer time.
-- Once the policy start guiding the robot towards achieving the task, even if its not perfect, you can limit your interventions to simple quick actions like a grasping command, or grasp and lift command.
-
-## 1.6 Monitoring and Debugging
-
-If you have `wandb.enable` set to `true` in your configuration, you can monitor training progress in real-time through the [Weights & Biases](https://wandb.ai/site/) dashboard.
-
-# 2. Training a Reward Classifier with LeRobot
-
-This guide explains how to train a reward classifier for human-in-the-loop reinforcement learning implementation of  LeRobot. Reward classifiers learn to predict the reward value given a state which can be used in an RL setup to train a policy.
-
-
-The reward classifier implementation in `modeling_classifier.py` uses a pretrained vision model to process the images. It can output either a single value for binary rewards to predict success/fail cases or multiple values for multi-class settings.
-
-## 2.1 Collecting a Dataset
-Before training, you need to collect a dataset with labeled examples. The `record_dataset` function in `gym_manipulator.py` enables the process of collecting a dataset of observations, actions, and rewards.
-
-To collect a dataset, you need to modeify some parameters in the environment configuration based on HILSerlRobotEnvConfig.
-
-```bash
-python lerobot/scripts/rl/gym_manipulator.py --config_path lerobot/configs/reward_classifier_train_config.json
-```
-
-### 2.1.1 Key Parameters for Data Collection:
-
-- **mode**: set it to "record" to collect a dataset
-- **repo_id**: "hf_username/dataset_name", name of the dataset and repo on the hub
-- **num_episodes**: Number of episodes to record
-- **number_of_steps_after_success**: Number of additional frames to record after a success (reward=1) is detected
-- **fps**: Number of frames per second to record
-- **push_to_hub**: Whether to push the dataset to the hub
-
-The `number_of_steps_after_success` parameter is crucial as it allows you to collect more positive examples. When a success is detected, the system will continue recording for the specified number of steps while maintaining the reward=1 label. Otherwise, there won't be enough states in the dataset labeled to 1 to train a good classifier.
-
-Example configuration section for data collection:
-
-```json
-{
-    "mode": "record",
-    "repo_id": "hf_username/dataset_name",
-    "dataset_root": "data/your_dataset",
-    "num_episodes": 20,
-    "push_to_hub": true,
-    "fps": 10,
-    "number_of_steps_after_success": 15
-}
-```
-
-## 2.2 Reward Classifier Configuration
-
-The reward classifier is configured using `configuration_classifier.py`. Here are the key parameters:
-
-- **model_name**: Base model architecture (e.g., we mainly use "helper2424/resnet10")
-- **model_type**: "cnn" or "transformer"
-- **num_cameras**: Number of camera inputs
-- **num_classes**: Number of output classes (typically 2 for binary success/failure)
-- **hidden_dim**: Size of hidden representation
-- **dropout_rate**: Regularization parameter
-- **learning_rate**: Learning rate for optimizer
-
-Example configuration from `reward_classifier_train_config.json`:
-
-```json
-{
-  "policy": {
-    "type": "reward_classifier",
-    "model_name": "helper2424/resnet10",
-    "model_type": "cnn",
-    "num_cameras": 2,
-    "num_classes": 2,
-    "hidden_dim": 256,
-    "dropout_rate": 0.1,
-    "learning_rate": 1e-4,
-    "device": "cuda",
-    "use_amp": true,
-    "input_features": {
-      "observation.images.front": {
-        "type": "VISUAL",
-        "shape": [3, 128, 128]
-      },
-      "observation.images.side": {
-        "type": "VISUAL",
-        "shape": [3, 128, 128]
-      }
-    }
-  }
-}
-```
-
-## 2.3 Training the Classifier
-
-To train the classifier, use the `train.py` script with your configuration:
-
-```bash
-python lerobot/scripts/train.py --config_path lerobot/configs/reward_classifier_train_config.json
-```
-
-## 2.4 Deploying and Testing the Model
-
-To use your trained reward classifier, configure the `HILSerlRobotEnvConfig` to use your model:
-
-```python
-env_config = HILSerlRobotEnvConfig(
-    reward_classifier_pretrained_path="path_to_your_pretrained_trained_model",
-    # Other environment parameters
-)
-```
-or set the argument in the json config file.
-
-```json
-{
-    "reward_classifier_pretrained_path": "path_to_your_pretrained_model"
-}
-```
-
-Run gym_manipulator.py to test the model.
-```bash
-python lerobot/scripts/rl/gym_manipulator.py --config_path lerobot/configs/env_config.json
-```
-
-The reward classifier will automatically provide rewards based on the visual input from the robot's cameras.
-
-## 2.5 Example Workflow
-
-1. **Create the configuration files**:
-   Create the necessary json configuration files for the reward classifier and the environment. Check the `json_examples` directory for examples.
-
-2. **Collect a dataset**:
-   ```bash
-   python lerobot/scripts/rl/gym_manipulator.py --config_path lerobot/configs/env_config.json
-   ```
-
-3. **Train the classifier**:
-   ```bash
-   python lerobot/scripts/train.py --config_path lerobot/configs/reward_classifier_train_config.json
-   ```
-
-4. **Test the classifier**:
-   ```bash
-   python lerobot/scripts/rl/gym_manipulator.py --config_path lerobot/configs/env_config.json
-   ```
-# 3. Using gym_hil Simulation Environments with LeRobot
-
-This guide explains how to use the `gym_hil` simulation environments as an alternative to real robots when working with the LeRobot framework for Human-In-the-Loop (HIL) reinforcement learning.
-
-`gym_hil` is a package that provides Gymnasium-compatible simulation environments specifically designed for Human-In-the-Loop reinforcement learning. These environments allow you to:
-
-- Train policies in simulation to test the RL stack before training on real robots
-
-- Collect demonstrations in sim using external devices like gamepads or keyboards
-- Perform human interventions during policy learning
-
-Currently, the main environment is a Franka Panda robot simulation based on MuJoCo, with tasks like picking up a cube.
-
-## 3.1 Installation
-
-First, install the `gym_hil` package within the LeRobot environment:
-
-```bash
-pip install gym_hil
-
-# Or in LeRobot
-cd lerobot
-pip install -e .[hilserl]
-```
-
-## 3.2 Configuration
-
-To use `gym_hil` with LeRobot, you need to create a configuration file. An example is provided in `gym_hil_env.json`. Key configuration sections include:
-
-### 3.2.1 Environment Type and Task
-
-```json
-{
-    "type": "hil",
-    "name": "franka_sim",
-    "task": "PandaPickCubeGamepad-v0",
-    "device": "cuda"
-}
-```
-
-Available tasks:
-- `PandaPickCubeBase-v0`: Basic environment
-- `PandaPickCubeGamepad-v0`: With gamepad control
-- `PandaPickCubeKeyboard-v0`: With keyboard control
-
-### 3.2.2 Gym Wrappers Configuration
-
-```json
-"wrapper": {
-    "gripper_penalty": -0.02,
-    "control_time_s": 15.0,
-    "use_gripper": true,
-    "fixed_reset_joint_positions": [0.0, 0.195, 0.0, -2.43, 0.0, 2.62, 0.785],
-    "end_effector_step_sizes": {
-        "x": 0.025,
-        "y": 0.025,
-        "z": 0.025
-    },
-    "control_mode": "gamepad"
-    }
-```
-
-Important parameters:
-- `gripper_penalty`: Penalty for excessive gripper movement
-- `use_gripper`: Whether to enable gripper control
-- `end_effector_step_sizes`: Size of the steps in the x,y,z axes of the end-effector
-- `control_mode`: Set to "gamepad" to use a gamepad controller
-
-## 3.3 Running with HIL RL of LeRobot
-
-### 3.3.1 Basic Usage
-
-To run the environment, set mode to null:
-
-```python
-python lerobot/scripts/rl/gym_manipulator.py --config_path path/to/gym_hil_env.json
-```
-
-### 3.3.2 Recording a Dataset
-
-To collect a dataset, set the mode to `record` whilst defining the repo_id and number of episodes to record:
-
-```python
-python lerobot/scripts/rl/gym_manipulator.py --config_path path/to/gym_hil_env.json
-```
-
-### 3.3.3 Training a Policy
-
-To train a policy, checkout the example json in `train_gym_hil_env.json` and run the actor and learner servers:
-
-```python
-python lerobot/scripts/rl/actor.py --config_path path/to/train_gym_hil_env.json
-```
-
-In a different terminal, run the learner server:
-
-```python
-python lerobot/scripts/rl/learner.py --config_path path/to/train_gym_hil_env.json
-```
-
-The simulation environment provides a safe and repeatable way to develop and test your Human-In-the-Loop reinforcement learning components before deploying to real robots.
-
-Paper citation:
-
-```
-@article{luo2024precise,
-  title={Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning},
-  author={Luo, Jianlan and Xu, Charles and Wu, Jeffrey and Levine, Sergey},
-  journal={arXiv preprint arXiv:2410.21845},
-  year={2024}
-}
-```

docs/source/index.mdx

@@ -1,19 +0,0 @@
-<div class="flex justify-center">
-  <a target="_blank" href="https://huggingface.co/lerobot">
-      <img alt="HuggingFace Expert Acceleration Program" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/lerobot/lerobot-logo-thumbnail.png" style="width: 100%"></img>
-  </a>
-</div>
-
-# LeRobot
-
-**State-of-the-art machine learning for real-world robotics**
-
-🤗 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 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.
-
-🤗 LeRobot hosts pretrained models and datasets on the LeRobot HuggingFace page.
-
-Join the LeRobot community on [Discord](https://discord.gg/s3KuuzsPFb)

docs/source/installation.mdx

@@ -1,70 +0,0 @@
-# Installation
-
-## Install LeRobot
-
-Currently only available from source.
-
-Download our source code:
-```bash
-git clone https://github.com/huggingface/lerobot.git
-cd lerobot
-```
-
-Create a virtual environment with Python 3.10, using [`Miniconda`](https://docs.anaconda.com/miniconda/install/#quick-command-line-install)
-```bash
-conda create -y -n lerobot python=3.10
-```
-
-Then activate your conda environment, you have to do this each time you open a shell to use lerobot:
-```bash
-conda activate lerobot
-```
-
-When using `miniconda`, install `ffmpeg` in your environment:
-```bash
-conda install ffmpeg -c conda-forge
-```
-
-> [!TIP]
-> This usually installs `ffmpeg 7.X` for your platform compiled with the `libsvtav1` encoder. If `libsvtav1` is not supported (check supported encoders with `ffmpeg -encoders`), you can:
->  - _[On any platform]_ Explicitly install `ffmpeg 7.X` using:
->  ```bash
->  conda install ffmpeg=7.1.1 -c conda-forge
->  ```
->  - _[On Linux only]_ If you want to bring your own ffmpeg: Install [ffmpeg build dependencies](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#GettheDependencies) and [compile ffmpeg from source with libsvtav1](https://trac.ffmpeg.org/wiki/CompilationGuide/Ubuntu#libsvtav1), and make sure you use the corresponding ffmpeg binary to your install with `which ffmpeg`.
-
-Install 🤗 LeRobot:
-```bash
-pip install -e .
-```
-
-### Troubleshooting
-If you encounter build errors, you may need to install additional dependencies: `cmake`, `build-essential`, and `ffmpeg libs`.
-To install these for linux run:
-```bash
-sudo apt-get install cmake build-essential python-dev pkg-config libavformat-dev libavcodec-dev libavdevice-dev libavutil-dev libswscale-dev libswresample-dev libavfilter-dev pkg-config
-```
-For other systems, see: [Compiling PyAV](https://pyav.org/docs/develop/overview/installation.html#bring-your-own-ffmpeg)
-
-## Optional dependencies
-
-LeRobot provides optional extras for specific functionalities. Multiple extras can be combined (e.g., `.[aloha,feetech]`). For all available extras, refer to `pyproject.toml`.
-
-### Simulations
-Install environment packages: `aloha` ([gym-aloha](https://github.com/huggingface/gym-aloha)), `xarm` ([gym-xarm](https://github.com/huggingface/gym-xarm)), or `pusht` ([gym-pusht](https://github.com/huggingface/gym-pusht))
-Example:
-```bash
-pip install -e ".[aloha]" # or "[pusht]" for example
-```
-
-### Motor Control
-For Koch v1.1 install the Dynamixel SDK, for SO100/SO101/Moss install the Feetech SDK.
-```bash
-pip install -e ".[feetech]" # or "[dynamixel]" for example
-```
-
-### Experiment Tracking
-To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
-```bash
-wandb login
-```

docs/source/koch.mdx

@@ -1 +0,0 @@
-../../lerobot/common/robots/koch_follower/koch.mdx

docs/source/lekiwi.mdx

@@ -1 +0,0 @@
-../../lerobot/common/robots/lekiwi/lekiwi.mdx

docs/source/so100.mdx

@@ -1 +0,0 @@
-../../lerobot/common/robots/so100_follower/so100.mdx

docs/source/so101.mdx

@@ -1 +0,0 @@
-../../lerobot/common/robots/so101_follower/so101.mdx

examples/10_use_so100.md

@@ -0,0 +1,624 @@
+# Using the [SO-100](https://github.com/TheRobotStudio/SO-ARM100) with LeRobot
+
+## Table of Contents
+
+  - [A. Source the parts](#a-source-the-parts)
+  - [B. Install LeRobot](#b-install-lerobot)
+  - [C. Configure the Motors](#c-configure-the-motors)
+  - [D. Step-by-Step Assembly Instructions](#d-step-by-step-assembly-instructions)
+  - [E. Calibrate](#e-calibrate)
+  - [F. Teleoperate](#f-teleoperate)
+  - [G. Record a dataset](#g-record-a-dataset)
+  - [H. Visualize a dataset](#h-visualize-a-dataset)
+  - [I. Replay an episode](#i-replay-an-episode)
+  - [J. Train a policy](#j-train-a-policy)
+  - [K. Evaluate your policy](#k-evaluate-your-policy)
+  - [L. More Information](#l-more-information)
+
+## A. Source the parts
+
+Follow this [README](https://github.com/TheRobotStudio/SO-ARM100). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts,
+and advice if it's your first time printing or if you don't own a 3D printer.
+
+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.
+
+## B. Install LeRobot
+
+> [!TIP]
+> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
+
+On your computer:
+
+#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
+
+#### 2. Restart shell
+Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
+
+#### 3. Create and activate a fresh conda environment for lerobot
+
+<details>
+<summary><strong>Video install instructions</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
+
+</details>
+
+```bash
+conda create -y -n lerobot python=3.10
+```
+
+Then activate your conda environment (do this each time you open a shell to use lerobot!):
+```bash
+conda activate lerobot
+```
+
+#### 4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+#### 5. Install ffmpeg in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+#### 6. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms :robot:.
+Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
+
+## C. Configure the motors
+
+> [!NOTE]
+> Throughout this tutorial you will find videos on how to do the steps, the full video tutorial can be found here: [assembly video](https://www.youtube.com/watch?v=FioA2oeFZ5I).
+
+### 1. Find the USB ports associated to each arm
+
+Designate one bus servo adapter and 6 motors for your leader arm, and similarly the other bus servo adapter and 6 motors for the follower arm. It's convenient to label them and write on each motor if it's for the follower `F` or for the leader `L` and it's ID from 1 to 6 (F1...F6 and L1...L6).
+
+#### a. Run the script to find port
+
+<details>
+<summary><strong>Video finding port</strong></summary>
+  <video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
+  <video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
+</details>
+
+To find the port for each bus servo adapter, run the utility script:
+```bash
+python lerobot/scripts/find_motors_bus_port.py
+```
+
+#### b. Example outputs
+
+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 MotorsBus and press Enter when done.
+
+[...Disconnect leader arm and press Enter...]
+
+The port of this MotorsBus 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 MotorsBus and press Enter when done.
+
+[...Disconnect follower arm and press Enter...]
+
+The port of this MotorsBus is /dev/tty.usbmodem575E0032081
+Reconnect the usb cable.
+```
+
+#### c. 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
+```
+
+#### d. Update config file
+
+IMPORTANTLY: Now that you have your ports, update the **port** default values of [`SO100RobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
+```python
+@RobotConfig.register_subclass("so100")
+@dataclass
+class So100RobotConfig(ManipulatorRobotConfig):
+    calibration_dir: str = ".cache/calibration/so100"
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
+
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+```
+
+### 2. Assembling the Base
+Let's begin with assembling the follower arm base
+
+#### a. Set IDs for all 12 motors
+
+<details>
+<summary><strong>Video configuring motor</strong></summary>
+  <video src="https://github.com/user-attachments/assets/ef9b3317-2e11-4858-b9d3-f0a02fb48ecf"></video>
+  <video src="https://github.com/user-attachments/assets/f36b5ed5-c803-4ebe-8947-b39278776a0d"></video>
+</details>
+
+Plug your first motor F1 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. Replace the text after --port to the corresponding follower control board port and run this command in cmd:
+```bash
+python lerobot/scripts/configure_motor.py \
+  --port /dev/tty.usbmodem58760432961 \
+  --brand feetech \
+  --model sts3215 \
+  --baudrate 1000000 \
+  --ID 1
+```
+
+> [!NOTE]
+> These motors are currently limited. 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.
+
+
+#### b. Remove the gears of the 6 leader motors
+
+<details>
+<summary><strong>Video removing gears</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/0c95b88c-5b85-413d-ba19-aee2f864f2a7"></video>
+
+</details>
+
+
+Follow the video for removing gears. 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.
+
+## D. Step-by-Step Assembly Instructions
+
+**Step 1: Clean Parts**
+- Remove all support material from the 3D-printed parts.
+---
+
+### Additional Guidance
+
+<details>
+<summary><strong>Video assembling arms</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/488a39de-0189-4461-9de3-05b015f90cca"></video>
+
+</details>
+
+**Note:**
+This video provides visual guidance for assembling the arms, but it doesn't specify when or how to do the wiring. Inserting the cables beforehand is much easier than doing it afterward. The first arm may take a bit more than 1 hour to assemble, but once you get used to it, you can assemble the second arm in under 1 hour.
+
+---
+
+### First Motor
+
+**Step 2: Insert Wires**
+- Insert two wires into the first motor.
+
+  <img src="../media/tutorial/img1.jpg" style="height:300px;">
+
+**Step 3: Install in Base**
+- Place the first motor into the base.
+
+  <img src="../media/tutorial/img2.jpg" style="height:300px;">
+
+**Step 4: Secure Motor**
+- Fasten the motor with 4 screws. Two from the bottom and two from top.
+
+**Step 5: Attach Motor Holder**
+- Slide over the first motor holder and fasten it using two screws (one on each side).
+
+  <img src="../media/tutorial/img4.jpg" style="height:300px;">
+
+**Step 6: Attach Motor Horns**
+- Install both motor horns, securing the top horn with a screw. Try not to move the motor position when attaching the motor horn, especially for the leader arms, where we removed the gears.
+
+  <img src="../media/tutorial/img5.jpg" style="height:300px;">
+<details>
+  <summary><strong>Video adding motor horn</strong></summary>
+  <video src="https://github.com/user-attachments/assets/ef3391a4-ad05-4100-b2bd-1699bf86c969"></video>
+</details>
+
+**Step 7: Attach Shoulder Part**
+- Route one wire to the back of the robot and the other to the left or in photo towards you (see photo).
+- Attach the shoulder part.
+
+  <img src="../media/tutorial/img6.jpg" style="height:300px;">
+
+**Step 8: Secure Shoulder**
+- Tighten the shoulder part with 4 screws on top and 4 on the bottom
+*(access bottom holes by turning the shoulder).*
+
+---
+
+### Second Motor Assembly
+
+**Step 9: Install Motor 2**
+- Slide the second motor in from the top and link the wire from motor 1 to motor 2.
+
+  <img src="../media/tutorial/img8.jpg" style="height:300px;">
+
+**Step 10: Attach Shoulder Holder**
+- Add the shoulder motor holder.
+- Ensure the wire from motor 1 to motor 2 goes behind the holder while the other wire is routed upward (see photo).
+- This part can be tight to assemble, you can use a workbench like the image or a similar setup to push the part around the motor.
+
+  <div style="display: flex;">
+    <img src="../media/tutorial/img9.jpg" style="height:250px;">
+    <img src="../media/tutorial/img10.jpg" style="height:250px;">
+    <img src="../media/tutorial/img12.jpg" style="height:250px;">
+  </div>
+
+**Step 11: Secure Motor 2**
+- Fasten the second motor with 4 screws.
+
+**Step 12: Attach Motor Horn**
+- Attach both motor horns to motor 2, again use the horn screw.
+
+**Step 13: Attach Base**
+- Install the base attachment using 2 screws.
+
+  <img src="../media/tutorial/img11.jpg" style="height:300px;">
+
+**Step 14: Attach Upper Arm**
+- Attach the upper arm with 4 screws on each side.
+
+  <img src="../media/tutorial/img13.jpg" style="height:300px;">
+
+---
+
+### Third Motor Assembly
+
+**Step 15: Install Motor 3**
+- Route the motor cable from motor 2 through the cable holder to motor 3, then secure motor 3 with 4 screws.
+
+**Step 16: Attach Motor Horn**
+- Attach both motor horns to motor 3 and secure one again with a horn screw.
+
+  <img src="../media/tutorial/img14.jpg" style="height:300px;">
+
+**Step 17: Attach Forearm**
+- Connect the forearm to motor 3 using 4 screws on each side.
+
+  <img src="../media/tutorial/img15.jpg" style="height:300px;">
+
+---
+
+### Fourth Motor Assembly
+
+**Step 18: Install Motor 4**
+- Slide in motor 4, attach the cable from motor 3, and secure the cable in its holder with a screw.
+
+  <div style="display: flex;">
+    <img src="../media/tutorial/img16.jpg" style="height:300px;">
+    <img src="../media/tutorial/img19.jpg" style="height:300px;">
+  </div>
+
+**Step 19: Attach Motor Holder 4**
+- Install the fourth motor holder (a tight fit). Ensure one wire is routed upward and the wire from motor 3 is routed downward (see photo).
+
+  <img src="../media/tutorial/img17.jpg" style="height:300px;">
+
+**Step 20: Secure Motor 4 & Attach Horn**
+- Fasten motor 4 with 4 screws and attach its motor horns, use for one a horn screw.
+
+  <img src="../media/tutorial/img18.jpg" style="height:300px;">
+
+---
+
+### Wrist Assembly
+
+**Step 21: Install Motor 5**
+- Insert motor 5 into the wrist holder and secure it with 2 front screws.
+
+  <img src="../media/tutorial/img20.jpg" style="height:300px;">
+
+**Step 22: Attach Wrist**
+- Connect the wire from motor 4 to motor 5. And already insert the other wire for the gripper.
+- Secure the wrist to motor 4 using 4 screws on both sides.
+
+  <img src="../media/tutorial/img22.jpg" style="height:300px;">
+
+**Step 23: Attach Wrist Horn**
+- Install only one motor horn on the wrist motor and secure it with a horn screw.
+
+  <img src="../media/tutorial/img23.jpg" style="height:300px;">
+
+---
+
+### Follower Configuration
+
+**Step 24: Attach Gripper**
+- Attach the gripper to motor 5.
+
+  <img src="../media/tutorial/img24.jpg" style="height:300px;">
+
+**Step 25: Install Gripper Motor**
+- Insert the gripper motor, connect the motor wire from motor 5 to motor 6, and secure it with 3 screws on each side.
+
+  <img src="../media/tutorial/img25.jpg" style="height:300px;">
+
+**Step 26: Attach Gripper Horn & Claw**
+- Attach the motor horns and again use a horn screw.
+- Install the gripper claw and secure it with 4 screws on both sides.
+
+  <img src="../media/tutorial/img26.jpg" style="height:300px;">
+
+**Step 27: Mount Controller**
+- Attach the motor controller on the back.
+
+  <div style="display: flex;">
+    <img src="../media/tutorial/img27.jpg" style="height:300px;">
+    <img src="../media/tutorial/img28.jpg" style="height:300px;">
+  </div>
+
+*Assembly complete – proceed to Leader arm assembly.*
+
+---
+
+### Leader Configuration
+
+For the leader configuration, perform **Steps 1–23**. Make sure that you removed the motor gears from the motors.
+
+**Step 24: Attach Leader Holder**
+- Mount the leader holder onto the wrist and secure it with a screw.
+
+  <img src="../media/tutorial/img29.jpg" style="height:300px;">
+
+**Step 25: Attach Handle**
+- Attach the handle to motor 5 using 4 screws.
+
+  <img src="../media/tutorial/img30.jpg" style="height:300px;">
+
+**Step 26: Install Gripper Motor**
+- Insert the gripper motor, secure it with 3 screws on each side, attach a motor horn using a horn screw, and connect the motor wire.
+
+  <img src="../media/tutorial/img31.jpg" style="height:300px;">
+
+**Step 27: Attach Trigger**
+- Attach the follower trigger with 4 screws.
+
+  <img src="../media/tutorial/img32.jpg" style="height:300px;">
+
+**Step 28: Mount Controller**
+- Attach the motor controller on the back.
+
+  <div style="display: flex;">
+    <img src="../media/tutorial/img27.jpg" style="height:300px;">
+    <img src="../media/tutorial/img28.jpg" style="height:300px;">
+  </div>
+
+*Assembly complete – proceed to calibration.*
+
+
+## E. 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.
+
+#### a. Manual calibration of follower arm
+
+> [!IMPORTANT]
+> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) 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 \
+  --robot.type=so100 \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_follower"]'
+```
+
+#### b. Manual calibration of leader arm
+Follow step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) 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 \
+  --robot.type=so100 \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_leader"]'
+```
+
+## F. 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 \
+  --robot.type=so100 \
+  --robot.cameras='{}' \
+  --control.type=teleoperate
+```
+
+
+#### a. 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.
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
+
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --control.type=teleoperate
+```
+
+## G. 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 $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/so100_test \
+  --control.tags='["so100","tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
+```
+
+Note: You can resume recording by adding `--control.resume=true`.
+
+## H. Visualize a dataset
+
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, 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 `--control.push_to_hub=false`, you can also visualize it locally with (a window can be opened in the browser `http://127.0.0.1:9090` with the visualization tool):
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/so100_test \
+  --local-files-only 1
+```
+
+## I. Replay an episode
+
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=so100 \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/so100_test \
+  --control.episode=0
+```
+
+## J. 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=${HF_USER}/so100_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_so100_test \
+  --job_name=act_so100_test \
+  --policy.device=cuda \
+  --wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/so100_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
+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`.
+
+To resume training from a checkpoint, below is an example command to resume from `last` checkpoint of the `act_so100_test` policy:
+```bash
+python lerobot/scripts/train.py \
+  --config_path=outputs/train/act_so100_test/checkpoints/last/pretrained_model/train_config.json \
+  --resume=true
+```
+
+## K. 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 \
+  --robot.type=so100 \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/eval_act_so100_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=10 \
+  --control.push_to_hub=true \
+  --control.policy.path=outputs/train/act_so100_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_so100_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_so100_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_so100_test`).
+
+## L. More Information
+
+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.
+
+> [!TIP]
+>  If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb) in the channel [`#so100-arm`](https://discord.com/channels/1216765309076115607/1237741463832363039).

examples/11_use_lekiwi.md

@@ -0,0 +1,597 @@
+# Using the [LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi) Robot with LeRobot
+
+## Table of Contents
+
+  - [A. Source the parts](#a-source-the-parts)
+  - [B. Install software Pi](#b-install-software-on-pi)
+  - [C. Setup LeRobot laptop/pc](#c-install-lerobot-on-laptop)
+  - [D. Assemble the arms](#d-assembly)
+  - [E. Calibrate](#e-calibration)
+  - [F. Teleoperate](#f-teleoperate)
+  - [G. Record a dataset](#g-record-a-dataset)
+  - [H. Visualize a dataset](#h-visualize-a-dataset)
+  - [I. Replay an episode](#i-replay-an-episode)
+  - [J. Train a policy](#j-train-a-policy)
+  - [K. Evaluate your policy](#k-evaluate-your-policy)
+
+> [!TIP]
+>  If you have any questions or need help, please reach out on [Discord](https://discord.com/invite/s3KuuzsPFb) in the channel [`#mobile-so-100-arm`](https://discord.com/channels/1216765309076115607/1318390825528332371).
+
+## A. Source the parts
+
+Follow this [README](https://github.com/SIGRobotics-UIUC/LeKiwi). It contains the bill of materials, with a link to source the parts, as well as the instructions to 3D print the parts, and advice if it's your first time printing or if you don't own a 3D printer.
+
+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.
+
+### Wired version
+If you have the **wired** LeKiwi version you can skip the installation of the Raspberry Pi and setting up SSH. You can also run all commands directly on your PC for both the LeKiwi scripts and the leader arm scripts for teleoperating.
+
+## B. Install software on Pi
+Now we have to setup the remote PC that will run on the LeKiwi Robot. This is normally a Raspberry Pi, but can be any PC that can run on 5V and has enough usb ports (2 or more) for the cameras and motor control board.
+
+### Install OS
+For setting up the Raspberry Pi and its SD-card see: [Setup PI](https://www.raspberrypi.com/documentation/computers/getting-started.html). Here is explained how to download the [Imager](https://www.raspberrypi.com/software/) to install Raspberry Pi OS or Ubuntu.
+
+### Setup SSH
+After setting up your Pi, you should enable and setup [SSH](https://www.raspberrypi.com/news/coding-on-raspberry-pi-remotely-with-visual-studio-code/) (Secure Shell Protocol) so you can login into the Pi from your laptop without requiring a screen, keyboard and mouse in the Pi. A great tutorial on how to do this can be found [here](https://www.raspberrypi.com/documentation/computers/remote-access.html#ssh). Logging into your Pi can be done in your Command Prompt (cmd) or if you use VSCode you can use [this](https://marketplace.visualstudio.com/items?itemName=ms-vscode-remote.remote-ssh) extension.
+
+### Install LeRobot
+
+On your Raspberry Pi:
+
+#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
+
+#### 2. Restart shell
+Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
+
+#### 3. Create and activate a fresh conda environment for lerobot
+
+<details>
+<summary><strong>Video install instructions</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
+
+</details>
+
+```bash
+conda create -y -n lerobot python=3.10
+```
+
+Then activate your conda environment (do this each time you open a shell to use lerobot!):
+```bash
+conda activate lerobot
+```
+
+#### 4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+#### 5. Install ffmpeg in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+#### 6. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+## C. Install LeRobot on laptop
+If you already have install LeRobot on your laptop you can skip this step, otherwise please follow along as we do the same steps we did on the Pi.
+
+> [!TIP]
+> We use the Command Prompt (cmd) quite a lot. If you are not comfortable using the cmd or want to brush up using the command line you can have a look here: [Command line crash course](https://developer.mozilla.org/en-US/docs/Learn_web_development/Getting_started/Environment_setup/Command_line)
+
+On your computer:
+
+#### 1. [Install Miniconda](https://docs.anaconda.com/miniconda/install/#quick-command-line-install):
+
+#### 2. Restart shell
+Copy paste in your shell: `source ~/.bashrc` or for Mac: `source ~/.bash_profile` or `source ~/.zshrc` if you're using zshell
+
+#### 3. Create and activate a fresh conda environment for lerobot
+
+<details>
+<summary><strong>Video install instructions</strong></summary>
+
+<video src="https://github.com/user-attachments/assets/17172d3b-3b64-4b80-9cf1-b2b7c5cbd236"></video>
+
+</details>
+
+```bash
+conda create -y -n lerobot python=3.10
+```
+
+Then activate your conda environment (do this each time you open a shell to use lerobot!):
+```bash
+conda activate lerobot
+```
+
+#### 4. Clone LeRobot:
+```bash
+git clone https://github.com/huggingface/lerobot.git ~/lerobot
+```
+
+#### 5. Install ffmpeg in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+#### 6. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+Great :hugs:! You are now done installing LeRobot and we can begin assembling the SO100 arms and Mobile base :robot:.
+Every time you now want to use LeRobot you can go to the `~/lerobot` folder where we installed LeRobot and run one of the commands.
+
+# D. Assembly
+
+First we will assemble the two SO100 arms. One to attach to the mobile base and one for teleoperation. Then we will assemble the mobile base.
+
+## SO100 Arms
+### Configure motors
+The instructions for configuring the motors can be found [Here](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#c-configure-the-motors) in step C of the SO100 tutorial. Besides the ID's for the arm motors we also need to set the motor ID's for the mobile base. These needs to be in a specific order to work. Below an image of the motor ID's and motor mounting positions for the mobile base. Note that we only use one Motor Control board on LeKiwi. This means the motor ID's for the wheels are 7, 8 and 9.
+
+<img src="../media/lekiwi/motor_ids.webp?raw=true" alt="Motor ID's for mobile robot" title="Motor ID's for mobile robot" width="60%">
+
+### Assemble arms
+[Assemble arms instruction](https://github.com/huggingface/lerobot/blob/main/examples/10_use_so100.md#d-assemble-the-arms)
+
+## Mobile base (LeKiwi)
+[Assemble LeKiwi](https://github.com/SIGRobotics-UIUC/LeKiwi)
+
+### Update config
+Both config files on the LeKiwi LeRobot and on the laptop should be the same. First we should find the Ip address of the Raspberry Pi of the mobile manipulator. This is the same Ip address used in SSH. We also need the usb port of the control board of the leader arm on the laptop and the port of the control board on LeKiwi. We can find these ports with the following script.
+
+#### a. Run the script to find port
+
+<details>
+<summary><strong>Video finding port</strong></summary>
+  <video src="https://github.com/user-attachments/assets/4a21a14d-2046-4805-93c4-ee97a30ba33f"></video>
+  <video src="https://github.com/user-attachments/assets/1cc3aecf-c16d-4ff9-aec7-8c175afbbce2"></video>
+</details>
+
+To find the port for each bus servo adapter, run the utility script:
+```bash
+python lerobot/scripts/find_motors_bus_port.py
+```
+
+#### b. Example outputs
+
+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.
+```
+
+#### c. 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
+```
+
+#### d. Update config file
+
+IMPORTANTLY: Now that you have your ports of leader and follower arm and ip address of the mobile-so100, update the **ip** in Network configuration, **port** in leader_arms and **port** in lekiwi. In the [`LeKiwiRobotConfig`](../lerobot/common/robot_devices/robots/configs.py) file. Where you will find something like:
+```python
+@RobotConfig.register_subclass("lekiwi")
+@dataclass
+class LeKiwiRobotConfig(RobotConfig):
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
+
+    # Network Configuration
+    ip: str = "172.17.133.91"
+    port: int = 5555
+    video_port: int = 5556
+
+    cameras: dict[str, CameraConfig] = field(
+        default_factory=lambda: {
+            "mobile": OpenCVCameraConfig(camera_index="/dev/video0", fps=30, width=640, height=480),
+            "mobile2": OpenCVCameraConfig(camera_index="/dev/video2", fps=30, width=640, height=480),
+        }
+    )
+
+    calibration_dir: str = ".cache/calibration/lekiwi"
+
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0077581",
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/ttyACM0",
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                    "left_wheel": (7, "sts3215"),
+                    "back_wheel": (8, "sts3215"),
+                    "right_wheel": (9, "sts3215"),
+                },
+            ),
+        }
+    )
+
+    teleop_keys: dict[str, str] = field(
+        default_factory=lambda: {
+            # Movement
+            "forward": "w",
+            "backward": "s",
+            "left": "a",
+            "right": "d",
+            "rotate_left": "z",
+            "rotate_right": "x",
+            # Speed control
+            "speed_up": "r",
+            "speed_down": "f",
+            # quit teleop
+            "quit": "q",
+        }
+    )
+
+    mock: bool = False
+```
+
+## Wired version
+
+For the wired LeKiwi version your configured IP address should refer to your own laptop (127.0.0.1), because leader arm and LeKiwi are in this case connected to own laptop. Below and example configuration for this wired setup:
+```python
+@RobotConfig.register_subclass("lekiwi")
+@dataclass
+class LeKiwiRobotConfig(RobotConfig):
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
+
+    # Network Configuration
+    ip: str = "127.0.0.1"
+    port: int = 5555
+    video_port: int = 5556
+
+    cameras: dict[str, CameraConfig] = field(
+        default_factory=lambda: {
+            "front": OpenCVCameraConfig(
+                camera_index=0, fps=30, width=640, height=480, rotation=90
+            ),
+            "wrist": OpenCVCameraConfig(
+                camera_index=1, fps=30, width=640, height=480, rotation=180
+            ),
+        }
+    )
+
+    calibration_dir: str = ".cache/calibration/lekiwi"
+
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0077581",
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem58760431061",
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                    "left_wheel": (7, "sts3215"),
+                    "back_wheel": (8, "sts3215"),
+                    "right_wheel": (9, "sts3215"),
+                },
+            ),
+        }
+    )
+
+    teleop_keys: dict[str, str] = field(
+        default_factory=lambda: {
+            # Movement
+            "forward": "w",
+            "backward": "s",
+            "left": "a",
+            "right": "d",
+            "rotate_left": "z",
+            "rotate_right": "x",
+            # Speed control
+            "speed_up": "r",
+            "speed_down": "f",
+            # quit teleop
+            "quit": "q",
+        }
+    )
+
+    mock: bool = False
+```
+
+# E. Calibration
+Now we have to calibrate the leader arm and the follower arm. The wheel motors don't have to be calibrated.
+
+
+### Calibrate follower arm (on mobile base)
+> [!IMPORTANT]
+> Contrarily to step 6 of the [assembly video](https://youtu.be/FioA2oeFZ5I?t=724) 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/lekiwi/mobile_calib_zero.webp?raw=true" alt="SO-100 follower arm zero position" title="SO-100 follower arm zero position" style="width:100%;"> | <img src="../media/lekiwi/mobile_calib_rotated.webp?raw=true" alt="SO-100 follower arm rotated position" title="SO-100 follower arm rotated position" style="width:100%;"> | <img src="../media/lekiwi/mobile_calib_rest.webp?raw=true" alt="SO-100 follower arm rest position" title="SO-100 follower arm rest position" style="width:100%;"> |
+
+Make sure the arm is connected to the Raspberry Pi and run this script (on the Raspberry Pi) to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_follower"]'
+```
+
+### Wired version
+If you have the **wired** LeKiwi version please run all commands including this calibration command on your laptop.
+
+### Calibrate leader arm
+Then to calibrate the leader arm (which is attached to the laptop/pc). 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 (on your laptop/pc) to launch manual calibration:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.arms='["main_leader"]'
+```
+
+# F. Teleoperate
+
+> [!TIP]
+> If you're using a Mac, you might need to give Terminal permission to access your keyboard. Go to System Preferences > Security & Privacy > Input Monitoring and check the box for Terminal.
+
+To teleoperate SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=remote_robot
+```
+
+Then on your laptop, also run `conda activate lerobot` and this script:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=teleoperate \
+  --control.fps=30
+```
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`. For the `--control.type=remote_robot` you will also need to set `--control.viewer_ip` and `--control.viewer_port`
+
+You should see on your laptop something like this: ```[INFO] Connected to remote robot at tcp://172.17.133.91:5555 and video stream at tcp://172.17.133.91:5556.``` Now you can move the leader arm and use the keyboard (w,a,s,d) to drive forward, left, backwards, right. And use (z,x) to turn left or turn right. You can use (r,f) to increase and decrease the speed of the mobile robot. There are three speed modes, see the table below:
+| Speed Mode | Linear Speed (m/s) | Rotation Speed (deg/s) |
+| ---------- | ------------------ | ---------------------- |
+| Fast       | 0.4                | 90                     |
+| Medium     | 0.25               | 60                     |
+| Slow       | 0.1                | 30                     |
+
+
+| Key | Action         |
+| --- | -------------- |
+| W   | Move forward   |
+| A   | Move left      |
+| S   | Move backward  |
+| D   | Move right     |
+| Z   | Turn left      |
+| X   | Turn right     |
+| R   | Increase speed |
+| F   | Decrease speed |
+
+> [!TIP]
+>  If you use a different keyboard you can change the keys for each command in the [`LeKiwiRobotConfig`](../lerobot/common/robot_devices/robots/configs.py).
+
+### Wired version
+If you have the **wired** LeKiwi version please run all commands including both these teleoperation commands on your laptop.
+
+## Troubleshoot communication
+
+If you are having trouble connecting to the Mobile SO100, follow these steps to diagnose and resolve the issue.
+
+### 1. Verify IP Address Configuration
+Make sure that the correct ip for the Pi is set in the configuration file. To check the Raspberry Pi's IP address, run (on the Pi command line):
+```bash
+hostname -I
+```
+
+### 2. Check if Pi is reachable from laptop/pc
+Try pinging the Raspberry Pi from your laptop:
+```bach
+ping <your_pi_ip_address>
+```
+
+If the ping fails:
+- Ensure the Pi is powered on and connected to the same network.
+- Check if SSH is enabled on the Pi.
+
+### 3. Try SSH connection
+If you can't SSH into the Pi, it might not be properly connected. Use:
+```bash
+ssh <your_pi_user_name>@<your_pi_ip_address>
+```
+If you get a connection error:
+- Ensure SSH is enabled on the Pi by running:
+  ```bash
+  sudo raspi-config
+  ```
+  Then navigate to: **Interfacing Options -> SSH** and enable it.
+
+### 4. Same config file
+Make sure the configuration file on both your laptop/pc and the Raspberry Pi is the same.
+
+# G. Record a dataset
+Once you're familiar with teleoperation, you can record your first dataset with LeKiwi.
+
+To start the program on LeKiwi, SSH into your Raspberry Pi, and run `conda activate lerobot` and this script:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=remote_robot
+```
+
+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 $HF_USER
+```
+On your laptop then run this command to record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/lekiwi_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
+```
+
+Note: You can resume recording by adding `--control.resume=true`.
+
+### Wired version
+If you have the **wired** LeKiwi version please run all commands including both these record dataset commands on your laptop.
+
+# H. Visualize a dataset
+
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, 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}/lekiwi_test
+```
+
+If you didn't upload with `--control.push_to_hub=false`, you can also visualize it locally with (a window can be opened in the browser `http://127.0.0.1:9090` with the visualization tool):
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/lekiwi_test \
+  --local-files-only 1
+```
+
+# I. Replay an episode
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=lekiwi \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/lekiwi_test \
+  --control.episode=0
+```
+
+## J. 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=${HF_USER}/lekiwi_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_lekiwi_test \
+  --job_name=act_lekiwi_test \
+  --policy.device=cuda \
+  --wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/lekiwi_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
+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_lekiwi_test/checkpoints`.
+
+## K. 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 \
+  --robot.type=lekiwi \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Drive to the red block and pick it up" \
+  --control.repo_id=${HF_USER}/eval_act_lekiwi_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=10 \
+  --control.push_to_hub=true \
+  --control.policy.path=outputs/train/act_lekiwi_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_lekiwi_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_lekiwi_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${HF_USER}/eval_act_lekiwi_test`).

examples/11_use_moss.md

@@ -0,0 +1,337 @@
+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 advice 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 ffmpeg in your environment:
+When using `miniconda`, install `ffmpeg` in your environment:
+```bash
+conda install ffmpeg -c conda-forge
+```
+
+6. Install LeRobot with dependencies for the feetech motors:
+```bash
+cd ~/lerobot && pip install -e ".[feetech]"
+```
+
+## 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
+```
+
+#### Update config file
+
+IMPORTANTLY: Now that you have your ports, update the **port** default values of [`MossRobotConfig`](../lerobot/common/robot_devices/robots/configs.py). You will find something like:
+```python
+@RobotConfig.register_subclass("moss")
+@dataclass
+class MossRobotConfig(ManipulatorRobotConfig):
+    calibration_dir: str = ".cache/calibration/moss"
+    # `max_relative_target` limits the magnitude of the relative positional target vector for safety purposes.
+    # Set this to a positive scalar to have the same value for all motors, or a list that is the same length as
+    # the number of motors in your follower arms.
+    max_relative_target: int | None = None
+
+    leader_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem58760431091",  <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+
+    follower_arms: dict[str, MotorsBusConfig] = field(
+        default_factory=lambda: {
+            "main": FeetechMotorsBusConfig(
+                port="/dev/tty.usbmodem585A0076891",  <-- UPDATE HERE
+                motors={
+                    # name: (index, model)
+                    "shoulder_pan": [1, "sts3215"],
+                    "shoulder_lift": [2, "sts3215"],
+                    "elbow_flex": [3, "sts3215"],
+                    "wrist_flex": [4, "sts3215"],
+                    "wrist_roll": [5, "sts3215"],
+                    "gripper": [6, "sts3215"],
+                },
+            ),
+        }
+    )
+```
+
+**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 \
+  --robot.type=moss \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.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 \
+  --robot.type=moss \
+  --robot.cameras='{}' \
+  --control.type=calibrate \
+  --control.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 \
+  --robot.type=moss \
+  --robot.cameras='{}' \
+  --control.type=teleoperate
+```
+
+
+**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.
+
+> **NOTE:** To visualize the data, enable `--control.display_data=true`. This streams the data using `rerun`.
+
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --control.type=teleoperate
+```
+
+## 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 $HF_USER
+```
+
+Record 2 episodes and upload your dataset to the hub:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/moss_test \
+  --control.tags='["moss","tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=2 \
+  --control.push_to_hub=true
+```
+
+Note: You can resume recording by adding `--control.resume=true`.
+
+## Visualize a dataset
+
+If you uploaded your dataset to the hub with `--control.push_to_hub=true`, 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 `--control.push_to_hub=false`, you can also visualize it locally with:
+```bash
+python lerobot/scripts/visualize_dataset_html.py \
+  --repo-id ${HF_USER}/moss_test \
+  --local-files-only 1
+```
+
+## Replay an episode
+
+Now try to replay the first episode on your robot:
+```bash
+python lerobot/scripts/control_robot.py \
+  --robot.type=moss \
+  --control.type=replay \
+  --control.fps=30 \
+  --control.repo_id=${HF_USER}/moss_test \
+  --control.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=${HF_USER}/moss_test \
+  --policy.type=act \
+  --output_dir=outputs/train/act_moss_test \
+  --job_name=act_moss_test \
+  --policy.device=cuda \
+  --wandb.enable=true
+```
+
+Let's explain it:
+1. We provided the dataset as argument with `--dataset.repo_id=${HF_USER}/moss_test`.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
+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 \
+  --robot.type=moss \
+  --control.type=record \
+  --control.fps=30 \
+  --control.single_task="Grasp a lego block and put it in the bin." \
+  --control.repo_id=${HF_USER}/eval_act_moss_test \
+  --control.tags='["tutorial"]' \
+  --control.warmup_time_s=5 \
+  --control.episode_time_s=30 \
+  --control.reset_time_s=30 \
+  --control.num_episodes=10 \
+  --control.push_to_hub=true \
+  --control.policy.path=outputs/train/act_moss_test/checkpoints/last/pretrained_model
+```
+
+As you can see, it's almost the same command as previously used to record your training dataset. Two things changed:
+1. There is an additional `--control.policy.path` argument which indicates the path to your policy checkpoint with  (e.g. `outputs/train/eval_act_moss_test/checkpoints/last/pretrained_model`). You can also use the model repository if you uploaded a model checkpoint to the hub (e.g. `${HF_USER}/act_moss_test`).
+2. The name of dataset begins by `eval` to reflect that you are running inference (e.g. `${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/1_load_lerobot_dataset.py

@@ -92,11 +92,11 @@ print(dataset.hf_dataset)
 # 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
+# episodes, you can access the frame indices of any episode using dataset.meta.episodes. 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()
+from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
+to_idx = dataset.meta.episodes["dataset_to_index"][episode_index]
 
 # Then we grab all the image frames from the first camera:
 camera_key = dataset.meta.camera_keys[0]

examples/2_evaluate_pretrained_policy.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 """
-This script demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
+This scripts demonstrates how to evaluate a pretrained policy from the HuggingFace Hub or from your local
 training outputs directory. In the latter case, you might want to run examples/3_train_policy.py first.
 
 It requires the installation of the 'gym_pusht' simulation environment. Install it by running:
@@ -119,7 +119,7 @@ while not done:
     rewards.append(reward)
     frames.append(env.render())
 
-    # The rollout is considered done when the success state is reached (i.e. terminated is True),
+    # 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

examples/3_train_policy.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""This script demonstrates how to train Diffusion Policy on the PushT environment.
+"""This scripts demonstrates how to train Diffusion Policy on the PushT environment.
 
 Once you have trained a model with this script, you can try to evaluate it on
 examples/2_evaluate_pretrained_policy.py

examples/4_train_policy_with_script.md

@@ -1,5 +1,5 @@
 This tutorial will explain the training script, how to use it, and particularly how to configure everything needed for the training run.
-> **Note:** The following assumes you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
+> **Note:** The following assume you're running these commands on a machine equipped with a cuda GPU. If you don't have one (or if you're using a Mac), you can add `--policy.device=cpu` (`--policy.device=mps` respectively). However, be advised that the code executes much slower on cpu.
 
 
 ## The training script
@@ -23,7 +23,7 @@ def train(cfg: TrainPipelineConfig):
 
 You can inspect the `TrainPipelineConfig` defined in [`lerobot/configs/train.py`](../lerobot/configs/train.py) (which is heavily commented and meant to be a reference to understand any option)
 
-When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated to this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
+When running the script, inputs for the command line are parsed thanks to the `@parser.wrap()` decorator and an instance of this class is automatically generated. Under the hood, this is done with [Draccus](https://github.com/dlwh/draccus) which is a tool dedicated for this purpose. If you're familiar with Hydra, Draccus can similarly load configurations from config files (.json, .yaml) and also override their values through command line inputs. Unlike Hydra, these configurations are pre-defined in the code through dataclasses rather than being defined entirely in config files. This allows for more rigorous serialization/deserialization, typing, and to manipulate configuration as objects directly in the code and not as dictionaries or namespaces (which enables nice features in an IDE such as autocomplete, jump-to-def, etc.)
 
 Let's have a look at a simplified example. Amongst other attributes, the training config has the following attributes:
 ```python
@@ -43,7 +43,7 @@ class DatasetConfig:
 ```
 
 This creates a hierarchical relationship where, for example assuming we have a `cfg` instance of `TrainPipelineConfig`, we can access the `repo_id` value with `cfg.dataset.repo_id`.
-From the command line, we can specify this value by using a very similar syntax `--dataset.repo_id=repo/id`.
+From the command line, we can specify this value with using a very similar syntax `--dataset.repo_id=repo/id`.
 
 By default, every field takes its default value specified in the dataclass. If a field doesn't have a default value, it needs to be specified either from the command line or from a config file – which path is also given in the command line (more in this below). In the example above, the `dataset` field doesn't have a default value which means it must be specified.
 
@@ -135,7 +135,7 @@ will start a training run with the same configuration used for training [lerobot
 
 ## Resume training
 
-Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to do that here.
+Being able to resume a training run is important in case it crashed or aborted for any reason. We'll demonstrate how to that here.
 
 Let's reuse the command from the previous run and add a few more options:
 ```bash

examples/7_get_started_with_real_robot.md

@@ -83,7 +83,7 @@ python lerobot/scripts/configure_motor.py \
   --brand dynamixel \
   --model xl330-m288 \
   --baudrate 1000000 \
-  --id 1
+  --ID 1
 ```
 
 Then unplug your first motor and plug the second motor and set its ID to 2.
@@ -93,7 +93,7 @@ python lerobot/scripts/configure_motor.py \
   --brand dynamixel \
   --model xl330-m288 \
   --baudrate 1000000 \
-  --id 2
+  --ID 2
 ```
 
 Redo the process for all your motors until ID 6.
@@ -377,7 +377,7 @@ robot = ManipulatorRobot(robot_config)
 
 The `KochRobotConfig` is used to set the associated settings and calibration process. For instance, we activate the torque of the gripper of the leader Koch v1.1 arm and position it at a 40 degree angle to use it as a trigger.
 
-For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `AlohaRobotConfig` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected for Aloha.
+For the [Aloha bimanual robot](https://aloha-2.github.io), we would use `AlohaRobotConfig` to set different settings such as a secondary ID for shadow joints (shoulder, elbow). Specific to Aloha, LeRobot comes with default calibration files stored in in `.cache/calibration/aloha_default`. Assuming the motors have been properly assembled, no manual calibration step is expected for Aloha.
 
 **Calibrate and Connect the ManipulatorRobot**
 
@@ -399,7 +399,7 @@ And here are the corresponding positions for the leader arm:
 
 You can watch a [video tutorial of the calibration procedure](https://youtu.be/8drnU9uRY24) for more details.
 
-During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask you to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to measure if the values changed negatively or positively.
+During calibration, we count the number of full 360-degree rotations your motors have made since they were first used. That's why we ask yo to move to this arbitrary "zero" position. We don't actually "set" the zero position, so you don't need to be accurate. After calculating these "offsets" to shift the motor values around 0, we need to assess the rotation direction of each motor, which might differ. That's why we ask you to rotate all motors to roughly 90 degrees, to measure if the values changed negatively or positively.
 
 Finally, the rest position ensures that the follower and leader arms are roughly aligned after calibration, preventing sudden movements that could damage the motors when starting teleoperation.
 
@@ -622,7 +622,7 @@ camera_01_frame_000047.png
 
 Note: Some cameras may take a few seconds to warm up, and the first frame might be black or green.
 
-Finally, run this code to instantiate and connect your camera:
+Finally, run this code to instantiate and connectyour camera:
 ```python
 from lerobot.common.robot_devices.cameras.configs import OpenCVCameraConfig
 from lerobot.common.robot_devices.cameras.opencv import OpenCVCamera

examples/8_use_stretch.md

@@ -99,7 +99,7 @@ This is equivalent to running `stretch_robot_home.py`
 > **Note:** If you run any of the LeRobot scripts below and Stretch is not properly homed, it will automatically home/calibrate first.
 
 **Teleoperate**
-Before trying teleoperation, you need to 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).
+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):
 

examples/9_use_aloha.md

@@ -142,7 +142,7 @@ python lerobot/scripts/train.py \
 
 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.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor states, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
+2. We provided the policy with `policy.type=act`. This loads configurations from [`configuration_act.py`](../lerobot/common/policies/act/configuration_act.py). Importantly, this policy will automatically adapt to the number of motor sates, motor actions and cameras of your robot (e.g. `laptop` and `phone`) which have been saved in your dataset.
 4. We provided `policy.device=cuda` since we are training on a Nvidia GPU, but you could use `policy.device=mps` to train on Apple silicon.
 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`.
 

examples/advanced/1_add_image_transforms.py

@@ -31,7 +31,7 @@ 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()
+first_idx = dataset.meta.episodes["dataset_from_index"][0]
 
 # Get the frame corresponding to the first camera
 frame = dataset[first_idx][dataset.meta.camera_keys[0]]

examples/advanced/2_calculate_validation_loss.py

@@ -66,7 +66,7 @@ def main():
     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 and val datasets
+    # - Load train an val datasets
     train_dataset = LeRobotDataset(
         "lerobot/pusht", episodes=train_episodes, delta_timestamps=delta_timestamps
     )

examples/port_datasets/droid_rlds/README.md

@@ -0,0 +1,144 @@
+# Port DROID 1.0.1 dataset to LeRobotDataset
+
+## Download
+
+TODO
+
+It will take 2 TB in your local disk.
+
+## Port on a single computer
+
+First, install tensorflow dataset utilities to read from raw files:
+```bash
+pip install tensorflow
+pip install tensorflow_datasets
+```
+
+Then run this script to start porting the dataset:
+```bash
+python examples/port_datasets/droid_rlds/port_droid.py \
+    --raw-dir /your/data/droid/1.0.1 \
+    --repo-id your_id/droid_1.0.1 \
+    --push-to-hub
+```
+
+It will take 400GB in your local disk.
+
+As usual, your LeRobotDataset will be stored in your huggingface/lerobot cache folder.
+
+WARNING: it will take 7 days for porting the dataset locally and 3 days to upload, so we will need to parallelize over multiple nodes on a slurm cluster.
+
+NOTE: For development, run this script to start porting a shard:
+```bash
+python examples/port_datasets/droid_rlds/port.py \
+    --raw-dir /your/data/droid/1.0.1 \
+    --repo-id your_id/droid_1.0.1 \
+    --num-shards 2048 \
+    --shard-index 0
+```
+
+## Port over SLURM
+
+Install slurm utilities from Hugging Face:
+```bash
+pip install datatrove
+```
+
+
+### 1. Port one shard per job
+
+Run this script to start porting shards of the dataset:
+```bash
+python examples/port_datasets/droid_rlds/slurm_port_shards.py \
+    --raw-dir /your/data/droid/1.0.1 \
+    --repo-id your_id/droid_1.0.1 \
+    --logs-dir /your/logs \
+    --job-name port_droid \
+    --partition your_partition \
+    --workers 2048 \
+    --cpus-per-task 8 \
+    --mem-per-cpu 1950M
+```
+
+**Note on how to set your command line arguments**
+
+Regarding `--partition`, find yours by running:
+```bash
+info --format="%R"`
+```
+and select the CPU partition if you have one. No GPU needed.
+
+Regarding `--workers`, it is the number of slurm jobs you will launch in parallel. 2048 is the maximum number, since there is 2048 shards in Droid. This big number will certainly max-out your cluster.
+
+Regarding `--cpus-per-task` and `--mem-per-cpu`, by default it will use ~16GB of RAM (8*1950M) which is recommended to load the raw frames and 8 CPUs which can be useful to parallelize the encoding of the frames.
+
+Find the number of CPUs and Memory of the nodes of your partition by running:
+```bash
+sinfo -N -p your_partition -h -o "%N cpus=%c mem=%m"
+```
+
+**Useful commands to check progress and debug**
+
+Check if your jobs are running:
+```bash
+squeue -u $USER`
+```
+
+You should see a list with job indices like `15125385_155` where `15125385` is the index of the run and `155` is the worker index. The output/print of this worker is written in real time in `/your/logs/job_name/slurm_jobs/15125385_155.out`. For instance, you can inspect the content of this file by running `less /your/logs/job_name/slurm_jobs/15125385_155.out`.
+
+Check the progression of your jobs by running:
+```bash
+jobs_status /your/logs
+```
+
+If it's not 100% and no more slurm job is running, it means that some of them failed. Inspect the logs by running:
+```bash
+failed_logs /your/logs/job_name
+```
+
+If there is an issue in the code, you can fix it in debug mode with `--slurm 0` which allows to set breakpoint:
+```bash
+python examples/port_datasets/droid_rlds/slurm_port_shards.py --slurm 0 ...
+```
+
+And you can relaunch the same command, which will skip the completed jobs:
+```bash
+python examples/port_datasets/droid_rlds/slurm_port_shards.py --slurm 1 ...
+```
+
+Once all jobs are completed, you will have one dataset per shard (e.g. `droid_1.0.1_world_2048_rank_1594`) saved on disk in your `/lerobot/home/dir/your_id` directory. You can find your `/lerobot/home/dir` by running:
+```bash
+python -c "from lerobot.common.constants import HF_LEROBOT_HOME;print(HF_LEROBOT_HOME)"
+```
+
+
+### 2. Aggregate all shards
+
+Run this script to start aggregation:
+```bash
+python examples/port_datasets/droid_rlds/slurm_aggregate_shards.py \
+    --repo-id your_id/droid_1.0.1 \
+    --logs-dir /your/logs \
+    --job-name aggr_droid \
+    --partition your_partition \
+    --workers 2048 \
+    --cpus-per-task 8 \
+    --mem-per-cpu 1950M
+```
+
+Once all jobs are completed, you will have one dataset your `/lerobot/home/dir/your_id/droid_1.0.1` directory.
+
+
+### 3. Upload dataset
+
+Run this script to start uploading:
+```bash
+python examples/port_datasets/droid_rlds/slurm_upload.py \
+    --repo-id your_id/droid_1.0.1 \
+    --logs-dir /your/logs \
+    --job-name upload_droid \
+    --partition your_partition \
+    --workers 50 \
+    --cpus-per-task 4 \
+    --mem-per-cpu 1950M
+```

examples/port_datasets/droid_rlds/port_droid.py

@@ -0,0 +1,430 @@
+#!/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 argparse
+import logging
+import time
+from pathlib import Path
+
+import numpy as np
+import tensorflow_datasets as tfds
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.utils.utils import get_elapsed_time_in_days_hours_minutes_seconds
+
+DROID_SHARDS = 2048
+DROID_FPS = 15
+DROID_ROBOT_TYPE = "Franka"
+
+# Dataset schema slightly adapted from: https://droid-dataset.github.io/droid/the-droid-dataset.html#-dataset-schema
+DROID_FEATURES = {
+    # true on first step of the episode
+    "is_first": {
+        "dtype": "bool",
+        "shape": (1,),
+        "names": None,
+    },
+    # true on last step of the episode
+    "is_last": {
+        "dtype": "bool",
+        "shape": (1,),
+        "names": None,
+    },
+    # true on last step of the episode if it is a terminal step, True for demos
+    "is_terminal": {
+        "dtype": "bool",
+        "shape": (1,),
+        "names": None,
+    },
+    # language_instruction is also stored as "task" to follow LeRobot standard
+    "language_instruction": {
+        "dtype": "string",
+        "shape": (1,),
+        "names": None,
+    },
+    "language_instruction_2": {
+        "dtype": "string",
+        "shape": (1,),
+        "names": None,
+    },
+    "language_instruction_3": {
+        "dtype": "string",
+        "shape": (1,),
+        "names": None,
+    },
+    "observation.state.gripper_position": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": {
+            "axes": ["gripper"],
+        },
+    },
+    "observation.state.cartesian_position": {
+        "dtype": "float32",
+        "shape": (6,),
+        "names": {
+            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
+        },
+    },
+    "observation.state.joint_position": {
+        "dtype": "float32",
+        "shape": (7,),
+        "names": {
+            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
+        },
+    },
+    # Add this new feature to follow LeRobot standard of using joint position + gripper
+    "observation.state": {
+        "dtype": "float32",
+        "shape": (8,),
+        "names": {
+            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "gripper"],
+        },
+    },
+    # Initially called wrist_image_left
+    "observation.images.wrist_left": {
+        "dtype": "video",
+        "shape": (180, 320, 3),
+        "names": [
+            "height",
+            "width",
+            "channels",
+        ],
+    },
+    # Initially called exterior_image_1_left
+    "observation.images.exterior_1_left": {
+        "dtype": "video",
+        "shape": (180, 320, 3),
+        "names": [
+            "height",
+            "width",
+            "channels",
+        ],
+    },
+    # Initially called exterior_image_2_left
+    "observation.images.exterior_2_left": {
+        "dtype": "video",
+        "shape": (180, 320, 3),
+        "names": [
+            "height",
+            "width",
+            "channels",
+        ],
+    },
+    "action.gripper_position": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": {
+            "axes": ["gripper"],
+        },
+    },
+    "action.gripper_velocity": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": {
+            "axes": ["gripper"],
+        },
+    },
+    "action.cartesian_position": {
+        "dtype": "float32",
+        "shape": (6,),
+        "names": {
+            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
+        },
+    },
+    "action.cartesian_velocity": {
+        "dtype": "float32",
+        "shape": (6,),
+        "names": {
+            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
+        },
+    },
+    "action.joint_position": {
+        "dtype": "float32",
+        "shape": (7,),
+        "names": {
+            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
+        },
+    },
+    "action.joint_velocity": {
+        "dtype": "float32",
+        "shape": (7,),
+        "names": {
+            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6"],
+        },
+    },
+    # This feature was called "action" in RLDS dataset and consists of [6x joint velocities, 1x gripper position]
+    "action.original": {
+        "dtype": "float32",
+        "shape": (7,),
+        "names": {
+            "axes": ["x", "y", "z", "roll", "pitch", "yaw", "gripper"],
+        },
+    },
+    # Add this new feature to follow LeRobot standard of using joint position + gripper
+    "action": {
+        "dtype": "float32",
+        "shape": (8,),
+        "names": {
+            "axes": ["joint_0", "joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "gripper"],
+        },
+    },
+    "discount": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": None,
+    },
+    "reward": {
+        "dtype": "float32",
+        "shape": (1,),
+        "names": None,
+    },
+    # Meta data that are the same for all frames in the episode
+    "task_category": {
+        "dtype": "string",
+        "shape": (1,),
+        "names": None,
+    },
+    "building": {
+        "dtype": "string",
+        "shape": (1,),
+        "names": None,
+    },
+    "collector_id": {
+        "dtype": "string",
+        "shape": (1,),
+        "names": None,
+    },
+    "date": {
+        "dtype": "string",
+        "shape": (1,),
+        "names": None,
+    },
+    "camera_extrinsics.wrist_left": {
+        "dtype": "float32",
+        "shape": (6,),
+        "names": {
+            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
+        },
+    },
+    "camera_extrinsics.exterior_1_left": {
+        "dtype": "float32",
+        "shape": (6,),
+        "names": {
+            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
+        },
+    },
+    "camera_extrinsics.exterior_2_left": {
+        "dtype": "float32",
+        "shape": (6,),
+        "names": {
+            "axes": ["x", "y", "z", "roll", "pitch", "yaw"],
+        },
+    },
+    "is_episode_successful": {
+        "dtype": "bool",
+        "shape": (1,),
+        "names": None,
+    },
+}
+
+
+def is_episode_successful(tf_episode_metadata):
+    # Adapted from: https://github.com/droid-dataset/droid_policy_learning/blob/dd1020eb20d981f90b5ff07dc80d80d5c0cb108b/robomimic/utils/rlds_utils.py#L8
+    return "/success/" in tf_episode_metadata["file_path"].numpy().decode()
+
+
+def generate_lerobot_frames(tf_episode):
+    m = tf_episode["episode_metadata"]
+    frame_meta = {
+        "task_category": m["building"].numpy().decode(),
+        "building": m["building"].numpy().decode(),
+        "collector_id": m["collector_id"].numpy().decode(),
+        "date": m["date"].numpy().decode(),
+        "camera_extrinsics.wrist_left": m["extrinsics_wrist_cam"].numpy(),
+        "camera_extrinsics.exterior_1_left": m["extrinsics_exterior_cam_1"].numpy(),
+        "camera_extrinsics.exterior_2_left": m["extrinsics_exterior_cam_2"].numpy(),
+        "is_episode_successful": np.array([is_episode_successful(m)]),
+    }
+    for f in tf_episode["steps"]:
+        # Dataset schema slightly adapted from: https://droid-dataset.github.io/droid/the-droid-dataset.html#-dataset-schema
+        frame = {
+            "is_first": np.array([f["is_first"].numpy()]),
+            "is_last": np.array([f["is_last"].numpy()]),
+            "is_terminal": np.array([f["is_terminal"].numpy()]),
+            "language_instruction": f["language_instruction"].numpy().decode(),
+            "language_instruction_2": f["language_instruction_2"].numpy().decode(),
+            "language_instruction_3": f["language_instruction_3"].numpy().decode(),
+            "observation.state.gripper_position": f["observation"]["gripper_position"].numpy(),
+            "observation.state.cartesian_position": f["observation"]["cartesian_position"].numpy(),
+            "observation.state.joint_position": f["observation"]["joint_position"].numpy(),
+            "observation.images.wrist_left": f["observation"]["wrist_image_left"].numpy(),
+            "observation.images.exterior_1_left": f["observation"]["exterior_image_1_left"].numpy(),
+            "observation.images.exterior_2_left": f["observation"]["exterior_image_2_left"].numpy(),
+            "action.gripper_position": f["action_dict"]["gripper_position"].numpy(),
+            "action.gripper_velocity": f["action_dict"]["gripper_velocity"].numpy(),
+            "action.cartesian_position": f["action_dict"]["cartesian_position"].numpy(),
+            "action.cartesian_velocity": f["action_dict"]["cartesian_velocity"].numpy(),
+            "action.joint_position": f["action_dict"]["joint_position"].numpy(),
+            "action.joint_velocity": f["action_dict"]["joint_velocity"].numpy(),
+            "discount": np.array([f["discount"].numpy()]),
+            "reward": np.array([f["reward"].numpy()]),
+            "action.original": f["action"].numpy(),
+        }
+
+        # language_instruction is also stored as "task" to follow LeRobot standard
+        frame["task"] = frame["language_instruction"]
+
+        # Add this new feature to follow LeRobot standard of using joint position + gripper
+        frame["observation.state"] = np.concatenate(
+            [frame["observation.state.joint_position"], frame["observation.state.gripper_position"]]
+        )
+        frame["action"] = np.concatenate([frame["action.joint_position"], frame["action.gripper_position"]])
+
+        # Meta data that are the same for all frames in the episode
+        frame.update(frame_meta)
+
+        # Cast fp64 to fp32
+        for key in frame:
+            if isinstance(frame[key], np.ndarray) and frame[key].dtype == np.float64:
+                frame[key] = frame[key].astype(np.float32)
+
+        yield frame
+
+
+def port_droid(
+    raw_dir: Path,
+    repo_id: str,
+    push_to_hub: bool = False,
+    num_shards: int | None = None,
+    shard_index: int | None = None,
+):
+    dataset_name = raw_dir.parent.name
+    version = raw_dir.name
+    data_dir = raw_dir.parent.parent
+
+    builder = tfds.builder(f"{dataset_name}/{version}", data_dir=data_dir, version="")
+
+    if num_shards is not None:
+        tfds_num_shards = builder.info.splits["train"].num_shards
+        if tfds_num_shards != DROID_SHARDS:
+            raise ValueError(
+                f"Number of shards of Droid dataset is expected to be {DROID_SHARDS} but is {tfds_num_shards}."
+            )
+        if num_shards != tfds_num_shards:
+            raise ValueError(
+                f"We only shard over the fixed number of shards provided by tensorflow dataset ({tfds_num_shards}), but {num_shards} shards provided instead."
+            )
+        if shard_index >= tfds_num_shards:
+            raise ValueError(
+                f"Shard index is greater than the num of shards ({shard_index} >= {num_shards})."
+            )
+
+        raw_dataset = builder.as_dataset(split=f"train[{shard_index}shard]")
+    else:
+        raw_dataset = builder.as_dataset(split="train")
+
+    lerobot_dataset = LeRobotDataset.create(
+        repo_id=repo_id,
+        robot_type=DROID_ROBOT_TYPE,
+        fps=DROID_FPS,
+        features=DROID_FEATURES,
+    )
+
+    start_time = time.time()
+    num_episodes = raw_dataset.cardinality().numpy().item()
+    logging.info(f"Number of episodes {num_episodes}")
+
+    for episode_index, episode in enumerate(raw_dataset):
+        elapsed_time = time.time() - start_time
+        d, h, m, s = get_elapsed_time_in_days_hours_minutes_seconds(elapsed_time)
+
+        logging.info(
+            f"{episode_index} / {num_episodes} episodes processed (after {d} days, {h} hours, {m} minutes, {s:.3f} seconds)"
+        )
+
+        for frame in generate_lerobot_frames(episode):
+            lerobot_dataset.add_frame(frame)
+
+        lerobot_dataset.save_episode()
+        logging.info("Save_episode")
+
+    if push_to_hub:
+        lerobot_dataset.push_to_hub(
+            # Add openx tag, since it belongs to the openx collection of datasets
+            tags=["openx"],
+            private=False,
+        )
+
+
+def validate_dataset(repo_id):
+    """Sanity check that ensure meta data can be loaded and all files are present."""
+    meta = LeRobotDatasetMetadata(repo_id)
+
+    if meta.total_episodes == 0:
+        raise ValueError("Number of episodes is 0.")
+
+    for ep_idx in range(meta.total_episodes):
+        data_path = meta.root / meta.get_data_file_path(ep_idx)
+
+        if not data_path.exists():
+            raise ValueError(f"Parquet file is missing in: {data_path}")
+
+        for vid_key in meta.video_keys:
+            vid_path = meta.root / meta.get_video_file_path(ep_idx, vid_key)
+            if not vid_path.exists():
+                raise ValueError(f"Video file is missing in: {vid_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        required=True,
+        help="Directory containing input raw datasets (e.g. `path/to/dataset` or `path/to/dataset/version).",
+    )
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True",
+    )
+    parser.add_argument(
+        "--push-to-hub",
+        action="store_true",
+        help="Upload to hub.",
+    )
+    parser.add_argument(
+        "--num-shards",
+        type=int,
+        default=None,
+        help="Number of shards. Can be either None to load the full dataset, or 2048 to load one of the 2048 tensorflow dataset files.",
+    )
+    parser.add_argument(
+        "--shard-index",
+        type=int,
+        default=None,
+        help="Index of the shard. Can be either None to load the full dataset, or in [0,2047] to load one of the 2048 tensorflow dataset files.",
+    )
+
+    args = parser.parse_args()
+
+    port_droid(**vars(args))
+
+
+if __name__ == "__main__":
+    main()

examples/port_datasets/droid_rlds/slurm_aggregate_shards.py

@@ -0,0 +1,293 @@
+#!/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 argparse
+import logging
+from pathlib import Path
+
+import tqdm
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+
+from examples.port_datasets.droid_rlds.port_droid import DROID_SHARDS
+from lerobot.common.datasets.aggregate import validate_all_metadata
+from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.common.datasets.utils import (
+    legacy_write_episode_stats,
+    legacy_write_task,
+    write_episode,
+    write_info,
+)
+from lerobot.common.utils.utils import init_logging
+
+
+class AggregateDatasets(PipelineStep):
+    def __init__(
+        self,
+        repo_ids: list[str],
+        aggregated_repo_id: str,
+    ):
+        super().__init__()
+        self.repo_ids = repo_ids
+        self.aggr_repo_id = aggregated_repo_id
+
+        self.create_aggr_dataset()
+
+    def create_aggr_dataset(self):
+        init_logging()
+
+        logging.info("Start aggregate_datasets")
+
+        all_metadata = [LeRobotDatasetMetadata(repo_id) for repo_id in self.repo_ids]
+
+        fps, robot_type, features = validate_all_metadata(all_metadata)
+
+        # Create resulting dataset folder
+        aggr_meta = LeRobotDatasetMetadata.create(
+            repo_id=self.aggr_repo_id,
+            fps=fps,
+            robot_type=robot_type,
+            features=features,
+        )
+
+        logging.info("Find all tasks")
+        # find all tasks, deduplicate them, create new task indices for each dataset
+        # indexed by dataset index
+        datasets_task_index_to_aggr_task_index = {}
+        aggr_task_index = 0
+        for dataset_index, meta in enumerate(tqdm.tqdm(all_metadata, desc="Find all tasks")):
+            task_index_to_aggr_task_index = {}
+
+            for task_index, task in meta.tasks.items():
+                if task not in aggr_meta.task_to_task_index:
+                    # add the task to aggr tasks mappings
+                    aggr_meta.tasks[aggr_task_index] = task
+                    aggr_meta.task_to_task_index[task] = aggr_task_index
+                    aggr_task_index += 1
+
+                # add task_index anyway
+                task_index_to_aggr_task_index[task_index] = aggr_meta.task_to_task_index[task]
+
+            datasets_task_index_to_aggr_task_index[dataset_index] = task_index_to_aggr_task_index
+
+        logging.info("Prepare copy data and videos")
+        datasets_ep_idx_to_aggr_ep_idx = {}
+        datasets_aggr_episode_index_shift = {}
+        aggr_episode_index_shift = 0
+        for dataset_index, meta in enumerate(tqdm.tqdm(all_metadata, desc="Prepare copy data and videos")):
+            ep_idx_to_aggr_ep_idx = {}
+
+            for episode_index in range(meta.total_episodes):
+                aggr_episode_index = episode_index + aggr_episode_index_shift
+                ep_idx_to_aggr_ep_idx[episode_index] = aggr_episode_index
+
+            datasets_ep_idx_to_aggr_ep_idx[dataset_index] = ep_idx_to_aggr_ep_idx
+            datasets_aggr_episode_index_shift[dataset_index] = aggr_episode_index_shift
+
+            # populate episodes
+            for episode_index, episode_dict in meta.episodes.items():
+                aggr_episode_index = episode_index + aggr_episode_index_shift
+                episode_dict["episode_index"] = aggr_episode_index
+                aggr_meta.episodes[aggr_episode_index] = episode_dict
+
+            # populate episodes_stats
+            for episode_index, episode_stats in meta.episodes_stats.items():
+                aggr_episode_index = episode_index + aggr_episode_index_shift
+                aggr_meta.episodes_stats[aggr_episode_index] = episode_stats
+
+            # populate info
+            aggr_meta.info["total_episodes"] += meta.total_episodes
+            aggr_meta.info["total_frames"] += meta.total_frames
+            aggr_meta.info["total_videos"] += len(aggr_meta.video_keys) * meta.total_episodes
+
+            aggr_episode_index_shift += meta.total_episodes
+
+        logging.info("Write meta data")
+        aggr_meta.info["total_tasks"] = len(aggr_meta.tasks)
+        aggr_meta.info["total_chunks"] = aggr_meta.get_episode_chunk(aggr_episode_index_shift - 1)
+        aggr_meta.info["splits"] = {"train": f"0:{aggr_meta.info['total_episodes']}"}
+
+        # create a new episodes jsonl with updated episode_index using write_episode
+        for episode_dict in tqdm.tqdm(aggr_meta.episodes.values(), desc="Write episodes"):
+            write_episode(episode_dict, aggr_meta.root)
+
+        # create a new episode_stats jsonl with updated episode_index using write_episode_stats
+        for episode_index, episode_stats in tqdm.tqdm(
+            aggr_meta.episodes_stats.items(), desc="Write episodes stats"
+        ):
+            legacy_write_episode_stats(episode_index, episode_stats, aggr_meta.root)
+
+        # create a new task jsonl with updated episode_index using write_task
+        for task_index, task in tqdm.tqdm(aggr_meta.tasks.items(), desc="Write tasks"):
+            legacy_write_task(task_index, task, aggr_meta.root)
+
+        write_info(aggr_meta.info, aggr_meta.root)
+
+        self.datasets_task_index_to_aggr_task_index = datasets_task_index_to_aggr_task_index
+        self.datasets_ep_idx_to_aggr_ep_idx = datasets_ep_idx_to_aggr_ep_idx
+        self.datasets_aggr_episode_index_shift = datasets_aggr_episode_index_shift
+
+        logging.info("Meta data done writing!")
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+        import shutil
+
+        import pandas as pd
+
+        from lerobot.common.datasets.aggregate import get_update_episode_and_task_func
+        from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
+        from lerobot.common.utils.utils import init_logging
+
+        init_logging()
+
+        aggr_meta = LeRobotDatasetMetadata(self.aggr_repo_id)
+        all_metadata = [LeRobotDatasetMetadata(repo_id) for repo_id in self.repo_ids]
+
+        if world_size != len(all_metadata):
+            raise ValueError()
+
+        dataset_index = rank
+        meta = all_metadata[dataset_index]
+        aggr_episode_index_shift = self.datasets_aggr_episode_index_shift[dataset_index]
+
+        logging.info("Copy data")
+        for episode_index in range(meta.total_episodes):
+            aggr_episode_index = self.datasets_ep_idx_to_aggr_ep_idx[dataset_index][episode_index]
+            data_path = meta.root / meta.get_data_file_path(episode_index)
+            aggr_data_path = aggr_meta.root / aggr_meta.get_data_file_path(aggr_episode_index)
+
+            # update episode_index and task_index
+            df = pd.read_parquet(data_path)
+            update_row_func = get_update_episode_and_task_func(
+                aggr_episode_index_shift, self.datasets_task_index_to_aggr_task_index[dataset_index]
+            )
+            df = df.apply(update_row_func, axis=1)
+
+            aggr_data_path.parent.mkdir(parents=True, exist_ok=True)
+            df.to_parquet(aggr_data_path)
+
+        logging.info("Copy videos")
+        for episode_index in range(meta.total_episodes):
+            aggr_episode_index = episode_index + aggr_episode_index_shift
+            for vid_key in meta.video_keys:
+                video_path = meta.root / meta.get_video_file_path(episode_index, vid_key)
+                aggr_video_path = aggr_meta.root / aggr_meta.get_video_file_path(aggr_episode_index, vid_key)
+                aggr_video_path.parent.mkdir(parents=True, exist_ok=True)
+                shutil.copy(video_path, aggr_video_path)
+
+                # copy_command = f"cp {video_path} {aggr_video_path} &"
+                # subprocess.Popen(copy_command, shell=True)
+
+        logging.info("Done!")
+
+
+def make_aggregate_executor(
+    repo_ids, repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
+):
+    kwargs = {
+        "pipeline": [
+            AggregateDatasets(repo_ids, repo_id),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": DROID_SHARDS,
+                "workers": workers,
+                "time": "08:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        executor = SlurmPipelineExecutor(**kwargs)
+    else:
+        kwargs.update(
+            {
+                "tasks": DROID_SHARDS,
+                "workers": 1,
+            }
+        )
+        executor = LocalPipelineExecutor(**kwargs)
+
+    return executor
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
+    )
+    parser.add_argument(
+        "--logs-dir",
+        type=Path,
+        help="Path to logs directory for `datatrove`.",
+    )
+    parser.add_argument(
+        "--job-name",
+        type=str,
+        default="aggr_droid",
+        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
+    )
+    parser.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
+    )
+    parser.add_argument(
+        "--workers",
+        type=int,
+        default=2048,
+        help="Number of slurm workers. It should be less than the maximum number of shards.",
+    )
+    parser.add_argument(
+        "--partition",
+        type=str,
+        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
+    )
+    parser.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=8,
+        help="Number of cpus that each slurm worker will use.",
+    )
+    parser.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="1950M",
+        help="Memory per cpu that each worker will use.",
+    )
+
+    args = parser.parse_args()
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+
+    repo_ids = [f"{args.repo_id}_world_{DROID_SHARDS}_rank_{rank}" for rank in range(DROID_SHARDS)]
+    aggregate_executor = make_aggregate_executor(repo_ids, **kwargs)
+    aggregate_executor.run()
+
+
+if __name__ == "__main__":
+    main()

examples/port_datasets/droid_rlds/slurm_port_shards.py

@@ -0,0 +1,147 @@
+import argparse
+from pathlib import Path
+
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+
+from examples.port_datasets.droid_rlds.port_droid import DROID_SHARDS
+
+
+class PortDroidShards(PipelineStep):
+    def __init__(
+        self,
+        raw_dir: Path | str,
+        repo_id: str = None,
+    ):
+        super().__init__()
+        self.raw_dir = Path(raw_dir)
+        self.repo_id = repo_id
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        from datasets.utils.tqdm import disable_progress_bars
+
+        from examples.port_datasets.droid_rlds.port_droid import port_droid, validate_dataset
+        from lerobot.common.utils.utils import init_logging
+
+        init_logging()
+        disable_progress_bars()
+
+        shard_repo_id = f"{self.repo_id}_world_{world_size}_rank_{rank}"
+
+        try:
+            validate_dataset(shard_repo_id)
+            return
+        except:
+            pass
+
+        port_droid(
+            self.raw_dir,
+            shard_repo_id,
+            push_to_hub=False,
+            num_shards=world_size,
+            shard_index=rank,
+        )
+
+        validate_dataset(shard_repo_id)
+
+
+def make_port_executor(
+    raw_dir, repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
+):
+    kwargs = {
+        "pipeline": [
+            PortDroidShards(raw_dir, repo_id),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": DROID_SHARDS,
+                "workers": workers,
+                "time": "08:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        executor = SlurmPipelineExecutor(**kwargs)
+    else:
+        kwargs.update(
+            {
+                "tasks": 1,
+                "workers": 1,
+            }
+        )
+        executor = LocalPipelineExecutor(**kwargs)
+
+    return executor
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--raw-dir",
+        type=Path,
+        required=True,
+        help="Directory containing input raw datasets (e.g. `path/to/dataset` or `path/to/dataset/version).",
+    )
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
+    )
+    parser.add_argument(
+        "--logs-dir",
+        type=Path,
+        help="Path to logs directory for `datatrove`.",
+    )
+    parser.add_argument(
+        "--job-name",
+        type=str,
+        default="port_droid",
+        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
+    )
+    parser.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
+    )
+    parser.add_argument(
+        "--workers",
+        type=int,
+        default=2048,
+        help="Number of slurm workers. It should be less than the maximum number of shards.",
+    )
+    parser.add_argument(
+        "--partition",
+        type=str,
+        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
+    )
+    parser.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=8,
+        help="Number of cpus that each slurm worker will use.",
+    )
+    parser.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="1950M",
+        help="Memory per cpu that each worker will use.",
+    )
+
+    args = parser.parse_args()
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+    port_executor = make_port_executor(**kwargs)
+    port_executor.run()
+
+
+if __name__ == "__main__":
+    main()

examples/port_datasets/droid_rlds/slurm_upload.py

@@ -0,0 +1,263 @@
+import argparse
+import logging
+import os
+from pathlib import Path
+
+from datatrove.executor import LocalPipelineExecutor
+from datatrove.executor.slurm import SlurmPipelineExecutor
+from datatrove.pipeline.base import PipelineStep
+from huggingface_hub import HfApi
+from huggingface_hub.constants import REPOCARD_NAME
+
+from examples.port_datasets.droid_rlds.port_droid import DROID_SHARDS
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDatasetMetadata
+from lerobot.common.datasets.utils import create_lerobot_dataset_card
+from lerobot.common.utils.utils import init_logging
+
+
+class UploadDataset(PipelineStep):
+    def __init__(
+        self,
+        repo_id: str,
+        branch: str | None = None,
+        revision: str | None = None,
+        tags: list | None = None,
+        license: str | None = "apache-2.0",
+        private: bool = False,
+        distant_repo_id: str | None = None,
+        **card_kwargs,
+    ):
+        super().__init__()
+        self.repo_id = repo_id
+        self.distant_repo_id = self.repo_id if distant_repo_id is None else distant_repo_id
+        self.branch = branch
+        self.tags = tags
+        self.license = license
+        self.private = private
+        self.card_kwargs = card_kwargs
+        self.revision = revision if revision else CODEBASE_VERSION
+
+        if os.environ.get("HF_HUB_ENABLE_HF_TRANSFER", "0") != "1":
+            logging.warning(
+                'HF_HUB_ENABLE_HF_TRANSFER is not set to "1". Install hf_transfer and set the env '
+                "variable for faster uploads:\npip install hf-transfer\nexport HF_HUB_ENABLE_HF_TRANSFER=1"
+            )
+
+        self.create_repo()
+
+    def create_repo(self):
+        logging.info(f"Loading meta data from {self.repo_id}...")
+        meta = LeRobotDatasetMetadata(self.repo_id)
+
+        logging.info(f"Creating repo {self.distant_repo_id}...")
+        hub_api = HfApi()
+        hub_api.create_repo(
+            repo_id=self.distant_repo_id,
+            private=self.private,
+            repo_type="dataset",
+            exist_ok=True,
+        )
+        if self.branch:
+            hub_api.create_branch(
+                repo_id=self.distant_repo_id,
+                branch=self.branch,
+                revision=self.revision,
+                repo_type="dataset",
+                exist_ok=True,
+            )
+
+        if not hub_api.file_exists(
+            self.distant_repo_id, REPOCARD_NAME, repo_type="dataset", revision=self.branch
+        ):
+            card = create_lerobot_dataset_card(
+                tags=self.tags, dataset_info=meta.info, license=self.license, **self.card_kwargs
+            )
+            card.push_to_hub(repo_id=self.distant_repo_id, repo_type="dataset", revision=self.branch)
+
+        def list_files_recursively(directory):
+            base_path = Path(directory)
+            return [str(file.relative_to(base_path)) for file in base_path.rglob("*") if file.is_file()]
+
+        logging.info(f"Listing all local files from {self.repo_id}...")
+        self.file_paths = list_files_recursively(meta.root)
+        self.file_paths = sorted(self.file_paths)
+
+    def create_chunks(self, lst, n):
+        from itertools import islice
+
+        it = iter(lst)
+        return [list(islice(it, size)) for size in [len(lst) // n + (i < len(lst) % n) for i in range(n)]]
+
+    def create_commits(self, additions):
+        import logging
+        import math
+        import random
+        import time
+
+        from huggingface_hub import create_commit
+        from huggingface_hub.utils import HfHubHTTPError
+
+        FILES_BETWEEN_COMMITS = 10  # noqa: N806
+        BASE_DELAY = 0.1  # noqa: N806
+        MAX_RETRIES = 12  # noqa: N806
+
+        # Split the files into smaller chunks for faster commit
+        # and avoiding "A commit has happened since" error
+        num_chunks = math.ceil(len(additions) / FILES_BETWEEN_COMMITS)
+        chunks = self.create_chunks(additions, num_chunks)
+
+        for chunk in chunks:
+            retries = 0
+            while True:
+                try:
+                    create_commit(
+                        self.distant_repo_id,
+                        repo_type="dataset",
+                        operations=chunk,
+                        commit_message=f"DataTrove upload ({len(chunk)} files)",
+                        revision=self.branch,
+                    )
+                    # TODO: every 100 chunks super_squach_commits()
+                    logging.info("create_commit completed!")
+                    break
+                except HfHubHTTPError as e:
+                    if "A commit has happened since" in e.server_message:
+                        if retries >= MAX_RETRIES:
+                            logging.error(f"Failed to create commit after {MAX_RETRIES=}. Giving up.")
+                            raise e
+                        logging.info("Commit creation race condition issue. Waiting...")
+                        time.sleep(BASE_DELAY * 2**retries + random.uniform(0, 2))
+                        retries += 1
+                    else:
+                        raise e
+
+    def run(self, data=None, rank: int = 0, world_size: int = 1):
+        import logging
+
+        from datasets.utils.tqdm import disable_progress_bars
+        from huggingface_hub import CommitOperationAdd, preupload_lfs_files
+
+        from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
+        from lerobot.common.utils.utils import init_logging
+
+        init_logging()
+        disable_progress_bars()
+
+        chunks = self.create_chunks(self.file_paths, world_size)
+        file_paths = chunks[rank]
+
+        if len(file_paths) == 0:
+            raise ValueError(file_paths)
+
+        logging.info("Pre-uploading LFS files...")
+        for i, path in enumerate(file_paths):
+            logging.info(f"{i}: {path}")
+
+        meta = LeRobotDatasetMetadata(self.repo_id)
+        additions = [
+            CommitOperationAdd(path_in_repo=path, path_or_fileobj=meta.root / path) for path in file_paths
+        ]
+        preupload_lfs_files(
+            repo_id=self.distant_repo_id, repo_type="dataset", additions=additions, revision=self.branch
+        )
+
+        logging.info("Creating commits...")
+        self.create_commits(additions)
+        logging.info("Done!")
+
+
+def make_upload_executor(
+    repo_id, job_name, logs_dir, workers, partition, cpus_per_task, mem_per_cpu, slurm=True
+):
+    kwargs = {
+        "pipeline": [
+            UploadDataset(repo_id),
+        ],
+        "logging_dir": str(logs_dir / job_name),
+    }
+
+    if slurm:
+        kwargs.update(
+            {
+                "job_name": job_name,
+                "tasks": DROID_SHARDS,
+                "workers": workers,
+                "time": "08:00:00",
+                "partition": partition,
+                "cpus_per_task": cpus_per_task,
+                "sbatch_args": {"mem-per-cpu": mem_per_cpu},
+            }
+        )
+        executor = SlurmPipelineExecutor(**kwargs)
+    else:
+        kwargs.update(
+            {
+                "tasks": DROID_SHARDS,
+                "workers": 1,
+            }
+        )
+        executor = LocalPipelineExecutor(**kwargs)
+
+    return executor
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        help="Repositery identifier on Hugging Face: a community or a user name `/` the name of the dataset, required when push-to-hub is True.",
+    )
+    parser.add_argument(
+        "--logs-dir",
+        type=Path,
+        help="Path to logs directory for `datatrove`.",
+    )
+    parser.add_argument(
+        "--job-name",
+        type=str,
+        default="upload_droid",
+        help="Job name used in slurm, and name of the directory created inside the provided logs directory.",
+    )
+    parser.add_argument(
+        "--slurm",
+        type=int,
+        default=1,
+        help="Launch over slurm. Use `--slurm 0` to launch sequentially (useful to debug).",
+    )
+    parser.add_argument(
+        "--workers",
+        type=int,
+        default=50,
+        help="Number of slurm workers. It should be less than the maximum number of shards.",
+    )
+    parser.add_argument(
+        "--partition",
+        type=str,
+        help="Slurm partition. Ideally a CPU partition. No need for GPU partition.",
+    )
+    parser.add_argument(
+        "--cpus-per-task",
+        type=int,
+        default=8,
+        help="Number of cpus that each slurm worker will use.",
+    )
+    parser.add_argument(
+        "--mem-per-cpu",
+        type=str,
+        default="1950M",
+        help="Memory per cpu that each worker will use.",
+    )
+
+    init_logging()
+
+    args = parser.parse_args()
+    kwargs = vars(args)
+    kwargs["slurm"] = kwargs.pop("slurm") == 1
+    upload_executor = make_upload_executor(**kwargs)
+    upload_executor.run()
+
+
+if __name__ == "__main__":
+    main()

examples/robots/lekiwi_client_app.py

@@ -1,94 +0,0 @@
-# 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 time
-
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.common.datasets.utils import hw_to_dataset_features
-from lerobot.common.robots.lekiwi.config_lekiwi import LeKiwiClientConfig
-from lerobot.common.robots.lekiwi.lekiwi_client import LeKiwiClient
-from lerobot.common.teleoperators.keyboard import KeyboardTeleop, KeyboardTeleopConfig
-from lerobot.common.teleoperators.so100_leader import SO100Leader, SO100LeaderConfig
-
-NB_CYCLES_CLIENT_CONNECTION = 250
-
-
-def main():
-    logging.info("Configuring Teleop Devices")
-    leader_arm_config = SO100LeaderConfig(port="/dev/tty.usbmodem58760433331")
-    leader_arm = SO100Leader(leader_arm_config)
-
-    keyboard_config = KeyboardTeleopConfig()
-    keyboard = KeyboardTeleop(keyboard_config)
-
-    logging.info("Configuring LeKiwi Client")
-    robot_config = LeKiwiClientConfig(remote_ip="172.18.134.136", id="lekiwi")
-    robot = LeKiwiClient(robot_config)
-
-    logging.info("Creating LeRobot Dataset")
-
-    action_features = hw_to_dataset_features(robot.action_features, "action")
-    obs_features = hw_to_dataset_features(robot.observation_features, "observation")
-    dataset_features = {**action_features, **obs_features}
-
-    dataset = LeRobotDataset.create(
-        repo_id="user/lekiwi" + str(int(time.time())),
-        fps=10,
-        features=dataset_features,
-        robot_type=robot.name,
-    )
-
-    logging.info("Connecting Teleop Devices")
-    leader_arm.connect()
-    keyboard.connect()
-
-    logging.info("Connecting remote LeKiwi")
-    robot.connect()
-
-    if not robot.is_connected or not leader_arm.is_connected or not keyboard.is_connected:
-        logging.error("Failed to connect to all devices")
-        return
-
-    logging.info("Starting LeKiwi teleoperation")
-    i = 0
-    while i < NB_CYCLES_CLIENT_CONNECTION:
-        arm_action = leader_arm.get_action()
-        base_action = keyboard.get_action()
-        action = {**arm_action, **base_action} if len(base_action) > 0 else arm_action
-
-        action_sent = robot.send_action(action)
-        observation = robot.get_observation()
-
-        frame = {**action_sent, **observation}
-        task = "Dummy Example Task Dataset"
-
-        logging.info("Saved a frame into the dataset")
-        dataset.add_frame(frame, task)
-        i += 1
-
-    logging.info("Disconnecting Teleop Devices and LeKiwi Client")
-    robot.disconnect()
-    leader_arm.disconnect()
-    keyboard.disconnect()
-
-    logging.info("Uploading dataset to the hub")
-    dataset.save_episode()
-    dataset.push_to_hub()
-
-    logging.info("Finished LeKiwi cleanly")
-
-
-if __name__ == "__main__":
-    main()

lerobot/__init__.py

@@ -181,7 +181,7 @@ available_robots = [
     "koch_bimanual",
     "aloha",
     "so100",
-    "so101",
+    "moss",
 ]
 
 # lists all available cameras from `lerobot/common/robot_devices/cameras`

lerobot/calibrate.py

@@ -1,84 +0,0 @@
-# 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 to recalibrate your device (robot or teleoperator).
-
-Example:
-
-```shell
-python -m lerobot.calibrate \
-    --teleop.type=so100_leader \
-    --teleop.port=/dev/tty.usbmodem58760431551 \
-    --teleop.id=blue
-```
-"""
-
-import logging
-from dataclasses import asdict, dataclass
-from pprint import pformat
-
-import draccus
-
-from lerobot.common.cameras.opencv.configuration_opencv import OpenCVCameraConfig  # noqa: F401
-from lerobot.common.cameras.realsense.configuration_realsense import RealSenseCameraConfig  # noqa: F401
-from lerobot.common.robots import (  # noqa: F401
-    Robot,
-    RobotConfig,
-    koch_follower,
-    lekiwi,
-    make_robot_from_config,
-    so100_follower,
-    so100_follower_end_effector,
-)
-from lerobot.common.teleoperators import (  # noqa: F401
-    Teleoperator,
-    TeleoperatorConfig,
-    koch_leader,
-    make_teleoperator_from_config,
-    so100_leader,
-    so101_leader,
-)
-from lerobot.common.utils.utils import init_logging
-
-
-@dataclass
-class CalibrateConfig:
-    teleop: TeleoperatorConfig | None = None
-    robot: RobotConfig | None = None
-
-    def __post_init__(self):
-        if bool(self.teleop) == bool(self.robot):
-            raise ValueError("Choose either a teleop or a robot.")
-
-        self.device = self.robot if self.robot else self.teleop
-
-
-@draccus.wrap()
-def calibrate(cfg: CalibrateConfig):
-    init_logging()
-    logging.info(pformat(asdict(cfg)))
-
-    if isinstance(cfg.device, RobotConfig):
-        device = make_robot_from_config(cfg.device)
-    elif isinstance(cfg.device, TeleoperatorConfig):
-        device = make_teleoperator_from_config(cfg.device)
-
-    device.connect(calibrate=False)
-    device.calibrate()
-    device.disconnect()
-
-
-if __name__ == "__main__":
-    calibrate()

lerobot/common/cameras/__init__.py

@@ -1,17 +0,0 @@
-# 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 .camera import Camera
-from .configs import CameraConfig, ColorMode, Cv2Rotation
-from .utils import make_cameras_from_configs

lerobot/common/cameras/camera.py

@@ -1,120 +0,0 @@
-#!/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 abc
-from typing import Any, Dict, List
-
-import numpy as np
-
-from .configs import CameraConfig, ColorMode
-
-
-class Camera(abc.ABC):
-    """Base class for camera implementations.
-
-    Defines a standard interface for camera operations across different backends.
-    Subclasses must implement all abstract methods.
-
-    Manages basic camera properties (FPS, resolution) and core operations:
-    - Connection/disconnection
-    - Frame capture (sync/async)
-
-    Attributes:
-        fps (int | None): Configured frames per second
-        width (int | None): Frame width in pixels
-        height (int | None): Frame height in pixels
-
-    Example:
-        class MyCamera(Camera):
-            def __init__(self, config): ...
-            @property
-            def is_connected(self) -> bool: ...
-            def connect(self, warmup=True): ...
-            # Plus other required methods
-    """
-
-    def __init__(self, config: CameraConfig):
-        """Initialize the camera with the given configuration.
-
-        Args:
-            config: Camera configuration containing FPS and resolution.
-        """
-        self.fps: int | None = config.fps
-        self.width: int | None = config.width
-        self.height: int | None = config.height
-
-    @property
-    @abc.abstractmethod
-    def is_connected(self) -> bool:
-        """Check if the camera is currently connected.
-
-        Returns:
-            bool: True if the camera is connected and ready to capture frames,
-                  False otherwise.
-        """
-        pass
-
-    @staticmethod
-    @abc.abstractmethod
-    def find_cameras() -> List[Dict[str, Any]]:
-        """Detects available cameras connected to the system.
-        Returns:
-            List[Dict[str, Any]]: A list of dictionaries,
-            where each dictionary contains information about a detected camera.
-        """
-        pass
-
-    @abc.abstractmethod
-    def connect(self, warmup: bool = True) -> None:
-        """Establish connection to the camera.
-
-        Args:
-            warmup: If True (default), captures a warmup frame before returning. Useful
-                   for cameras that require time to adjust capture settings.
-                   If False, skips the warmup frame.
-        """
-        pass
-
-    @abc.abstractmethod
-    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
-        """Capture and return a single frame from the camera.
-
-        Args:
-            color_mode: Desired color mode for the output frame. If None,
-                        uses the camera's default color mode.
-
-        Returns:
-            np.ndarray: Captured frame as a numpy array.
-        """
-        pass
-
-    @abc.abstractmethod
-    def async_read(self, timeout_ms: float = ...) -> np.ndarray:
-        """Asynchronously capture and return a single frame from the camera.
-
-        Args:
-            timeout_ms: Maximum time to wait for a frame in milliseconds.
-                        Defaults to implementation-specific timeout.
-
-        Returns:
-            np.ndarray: Captured frame as a numpy array.
-        """
-        pass
-
-    @abc.abstractmethod
-    def disconnect(self) -> None:
-        """Disconnect from the camera and release resources."""
-        pass

lerobot/common/cameras/configs.py

@@ -1,44 +0,0 @@
-#!/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 abc
-from dataclasses import dataclass
-from enum import Enum
-
-import draccus
-
-
-class ColorMode(str, Enum):
-    RGB = "rgb"
-    BGR = "bgr"
-
-
-class Cv2Rotation(int, Enum):
-    NO_ROTATION = 0
-    ROTATE_90 = 90
-    ROTATE_180 = 180
-    ROTATE_270 = -90
-
-
-@dataclass(kw_only=True)
-class CameraConfig(draccus.ChoiceRegistry, abc.ABC):
-    fps: int | None = None
-    width: int | None = None
-    height: int | None = None
-
-    @property
-    def type(self) -> str:
-        return self.get_choice_name(self.__class__)

lerobot/common/cameras/opencv/__init__.py

@@ -1,16 +0,0 @@
-# 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 .camera_opencv import OpenCVCamera
-from .configuration_opencv import OpenCVCameraConfig

lerobot/common/cameras/opencv/camera_opencv.py

@@ -1,481 +0,0 @@
-# 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.
-
-"""
-Provides the OpenCVCamera class for capturing frames from cameras using OpenCV.
-"""
-
-import logging
-import math
-import platform
-import time
-from pathlib import Path
-from threading import Event, Lock, Thread
-from typing import Any, Dict, List
-
-import cv2
-import numpy as np
-
-from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
-
-from ..camera import Camera
-from ..utils import get_cv2_backend, get_cv2_rotation
-from .configuration_opencv import ColorMode, OpenCVCameraConfig
-
-# NOTE(Steven): The maximum opencv device index depends on your operating system. For instance,
-# if you have 3 cameras, they should be associated to index 0, 1, and 2. This is the case
-# on MacOS. However, on Ubuntu, the indices are different like 6, 16, 23.
-# When you change the USB port or reboot the computer, the operating system might
-# treat the same cameras as new devices. Thus we select a higher bound to search indices.
-MAX_OPENCV_INDEX = 60
-
-logger = logging.getLogger(__name__)
-
-
-class OpenCVCamera(Camera):
-    """
-    Manages camera interactions using OpenCV for efficient frame recording.
-
-    This class provides a high-level interface to connect to, configure, and read
-    frames from cameras compatible with OpenCV's VideoCapture. It supports both
-    synchronous and asynchronous frame reading.
-
-    An OpenCVCamera instance requires a camera index (e.g., 0) or a device path
-    (e.g., '/dev/video0' on Linux). Camera indices can be unstable across reboots
-    or port changes, especially on Linux. Use the provided utility script to find
-    available camera indices or paths:
-    ```bash
-    python -m lerobot.find_cameras opencv
-    ```
-
-    The camera's default settings (FPS, resolution, color mode) are used unless
-    overridden in the configuration.
-
-    Example:
-        ```python
-        from lerobot.common.cameras.opencv import OpenCVCamera
-        from lerobot.common.cameras.configuration_opencv import OpenCVCameraConfig, ColorMode, Cv2Rotation
-
-        # Basic usage with camera index 0
-        config = OpenCVCameraConfig(index_or_path=0)
-        camera = OpenCVCamera(config)
-        camera.connect()
-
-        # Read 1 frame synchronously
-        color_image = camera.read()
-        print(color_image.shape)
-
-        # Read 1 frame asynchronously
-        async_image = camera.async_read()
-
-        # When done, properly disconnect the camera using
-        camera.disconnect()
-
-        # Example with custom settings
-        custom_config = OpenCVCameraConfig(
-            index_or_path='/dev/video0', # Or use an index
-            fps=30,
-            width=1280,
-            height=720,
-            color_mode=ColorMode.RGB,
-            rotation=Cv2Rotation.ROTATE_90
-        )
-        custom_camera = OpenCVCamera(custom_config)
-        # ... connect, read, disconnect ...
-        ```
-    """
-
-    def __init__(self, config: OpenCVCameraConfig):
-        """
-        Initializes the OpenCVCamera instance.
-
-        Args:
-            config: The configuration settings for the camera.
-        """
-        super().__init__(config)
-
-        self.config = config
-        self.index_or_path = config.index_or_path
-
-        self.fps = config.fps
-        self.color_mode = config.color_mode
-        self.warmup_s = config.warmup_s
-
-        self.videocapture: cv2.VideoCapture | None = None
-
-        self.thread: Thread | None = None
-        self.stop_event: Event | None = None
-        self.frame_lock: Lock = Lock()
-        self.latest_frame: np.ndarray | None = None
-        self.new_frame_event: Event = Event()
-
-        self.rotation: int | None = get_cv2_rotation(config.rotation)
-        self.backend: int = get_cv2_backend()
-
-        if self.height and self.width:
-            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-                self.capture_width, self.capture_height = self.height, self.width
-            else:
-                self.capture_width, self.capture_height = self.width, self.height
-
-    def __str__(self) -> str:
-        return f"{self.__class__.__name__}({self.index_or_path})"
-
-    @property
-    def is_connected(self) -> bool:
-        """Checks if the camera is currently connected and opened."""
-        return isinstance(self.videocapture, cv2.VideoCapture) and self.videocapture.isOpened()
-
-    def connect(self, warmup: bool = True):
-        """
-        Connects to the OpenCV camera specified in the configuration.
-
-        Initializes the OpenCV VideoCapture object, sets desired camera properties
-        (FPS, width, height), and performs initial checks.
-
-        Raises:
-            DeviceAlreadyConnectedError: If the camera is already connected.
-            ConnectionError: If the specified camera index/path is not found or the camera is found but fails to open.
-            RuntimeError: If the camera opens but fails to apply requested FPS/resolution settings.
-        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
-
-        # Use 1 thread for OpenCV operations to avoid potential conflicts or
-        # blocking in multi-threaded applications, especially during data collection.
-        cv2.setNumThreads(1)
-
-        self.videocapture = cv2.VideoCapture(self.index_or_path, self.backend)
-
-        if not self.videocapture.isOpened():
-            self.videocapture.release()
-            self.videocapture = None
-            raise ConnectionError(
-                f"Failed to open {self}."
-                f"Run `python -m lerobot.find_cameras opencv` to find available cameras."
-            )
-
-        self._configure_capture_settings()
-
-        if warmup:
-            start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
-                self.read()
-                time.sleep(0.1)
-
-        logger.info(f"{self} connected.")
-
-    def _configure_capture_settings(self) -> None:
-        """
-        Applies the specified FPS, width, and height settings to the connected camera.
-
-        This method attempts to set the camera properties via OpenCV. It checks if
-        the camera successfully applied the settings and raises an error if not.
-
-        Args:
-            fps: The desired frames per second. If None, the setting is skipped.
-            width: The desired capture width. If None, the setting is skipped.
-            height: The desired capture height. If None, the setting is skipped.
-
-        Raises:
-            RuntimeError: If the camera fails to set any of the specified properties
-                          to the requested value.
-            DeviceNotConnectedError: If the camera is not connected when attempting
-                                     to configure settings.
-        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"Cannot configure settings for {self} as it is not connected.")
-
-        if self.fps is None:
-            self.fps = self.videocapture.get(cv2.CAP_PROP_FPS)
-        else:
-            self._validate_fps()
-
-        default_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
-        default_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
-
-        if self.width is None or self.height is None:
-            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-                self.width, self.height = default_height, default_width
-                self.capture_width, self.capture_height = default_width, default_height
-            else:
-                self.width, self.height = default_width, default_height
-                self.capture_width, self.capture_height = default_width, default_height
-        else:
-            self._validate_width_and_height()
-
-    def _validate_fps(self) -> None:
-        """Validates and sets the camera's frames per second (FPS)."""
-
-        success = self.videocapture.set(cv2.CAP_PROP_FPS, float(self.fps))
-        actual_fps = self.videocapture.get(cv2.CAP_PROP_FPS)
-        # Use math.isclose for robust float comparison
-        if not success or not math.isclose(self.fps, actual_fps, rel_tol=1e-3):
-            raise RuntimeError(f"{self} failed to set fps={self.fps} ({actual_fps=}).")
-
-    def _validate_width_and_height(self) -> None:
-        """Validates and sets the camera's frame capture width and height."""
-
-        success = self.videocapture.set(cv2.CAP_PROP_FRAME_WIDTH, float(self.capture_width))
-        actual_width = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_WIDTH)))
-        if not success or self.capture_width != actual_width:
-            raise RuntimeError(f"{self} failed to set capture_width={self.capture_width} ({actual_width=}).")
-
-        success = self.videocapture.set(cv2.CAP_PROP_FRAME_HEIGHT, float(self.capture_height))
-        actual_height = int(round(self.videocapture.get(cv2.CAP_PROP_FRAME_HEIGHT)))
-        if not success or self.capture_height != actual_height:
-            raise RuntimeError(
-                f"{self} failed to set capture_height={self.capture_height} ({actual_height})."
-            )
-
-    @staticmethod
-    def find_cameras() -> List[Dict[str, Any]]:
-        """
-        Detects available OpenCV cameras connected to the system.
-
-        On Linux, it scans '/dev/video*' paths. On other systems (like macOS, Windows),
-        it checks indices from 0 up to `MAX_OPENCV_INDEX`.
-
-        Returns:
-            List[Dict[str, Any]]: A list of dictionaries,
-            where each dictionary contains 'type', 'id' (port index or path),
-            and the default profile properties (width, height, fps, format).
-        """
-        found_cameras_info = []
-
-        if platform.system() == "Linux":
-            possible_paths = sorted(Path("/dev").glob("video*"), key=lambda p: p.name)
-            targets_to_scan = [str(p) for p in possible_paths]
-        else:
-            targets_to_scan = list(range(MAX_OPENCV_INDEX))
-
-        for target in targets_to_scan:
-            camera = cv2.VideoCapture(target)
-            if camera.isOpened():
-                default_width = int(camera.get(cv2.CAP_PROP_FRAME_WIDTH))
-                default_height = int(camera.get(cv2.CAP_PROP_FRAME_HEIGHT))
-                default_fps = camera.get(cv2.CAP_PROP_FPS)
-                default_format = camera.get(cv2.CAP_PROP_FORMAT)
-                camera_info = {
-                    "name": f"OpenCV Camera @ {target}",
-                    "type": "OpenCV",
-                    "id": target,
-                    "backend_api": camera.getBackendName(),
-                    "default_stream_profile": {
-                        "format": default_format,
-                        "width": default_width,
-                        "height": default_height,
-                        "fps": default_fps,
-                    },
-                }
-
-                found_cameras_info.append(camera_info)
-                camera.release()
-
-        return found_cameras_info
-
-    def read(self, color_mode: ColorMode | None = None) -> np.ndarray:
-        """
-        Reads a single frame synchronously from the camera.
-
-        This is a blocking call. It waits for the next available frame from the
-        camera hardware via OpenCV.
-
-        Args:
-            color_mode (Optional[ColorMode]): If specified, overrides the default
-                color mode (`self.color_mode`) for this read operation (e.g.,
-                request RGB even if default is BGR).
-
-        Returns:
-            np.ndarray: The captured frame as a NumPy array in the format
-                       (height, width, channels), using the specified or default
-                       color mode and applying any configured rotation.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If reading the frame from the camera fails or if the
-                          received frame dimensions don't match expectations before rotation.
-            ValueError: If an invalid `color_mode` is requested.
-        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
-        start_time = time.perf_counter()
-
-        ret, frame = self.videocapture.read()
-
-        if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
-
-        processed_frame = self._postprocess_image(frame, color_mode)
-
-        read_duration_ms = (time.perf_counter() - start_time) * 1e3
-        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
-
-        return processed_frame
-
-    def _postprocess_image(self, image: np.ndarray, color_mode: ColorMode | None = None) -> np.ndarray:
-        """
-        Applies color conversion, dimension validation, and rotation to a raw frame.
-
-        Args:
-            image (np.ndarray): The raw image frame (expected BGR format from OpenCV).
-            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
-                                             uses the instance's default `self.color_mode`.
-
-        Returns:
-            np.ndarray: The processed image frame.
-
-        Raises:
-            ValueError: If the requested `color_mode` is invalid.
-            RuntimeError: If the raw frame dimensions do not match the configured
-                          `width` and `height`.
-        """
-        requested_color_mode = self.color_mode if color_mode is None else color_mode
-
-        if requested_color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"Invalid color mode '{requested_color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
-            )
-
-        h, w, c = image.shape
-
-        if h != self.capture_height or w != self.capture_width:
-            raise RuntimeError(
-                f"{self} frame width={w} or height={h} do not match configured width={self.capture_width} or height={self.capture_height}."
-            )
-
-        if c != 3:
-            raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
-
-        processed_image = image
-        if requested_color_mode == ColorMode.RGB:
-            processed_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
-
-        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-            processed_image = cv2.rotate(processed_image, self.rotation)
-
-        return processed_image
-
-    def _read_loop(self):
-        """
-        Internal loop run by the background thread for asynchronous reading.
-
-        On each iteration:
-        1. Reads a color frame
-        2. Stores result in latest_frame (thread-safe)
-        3. Sets new_frame_event to notify listeners
-
-        Stops on DeviceNotConnectedError, logs other errors and continues.
-        """
-        while not self.stop_event.is_set():
-            try:
-                color_image = self.read()
-
-                with self.frame_lock:
-                    self.latest_frame = color_image
-                self.new_frame_event.set()
-
-            except DeviceNotConnectedError:
-                break
-            except Exception as e:
-                logger.warning(f"Error reading frame in background thread for {self}: {e}")
-
-    def _start_read_thread(self) -> None:
-        """Starts or restarts the background read thread if it's not running."""
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=0.1)
-        if self.stop_event is not None:
-            self.stop_event.set()
-
-        self.stop_event = Event()
-        self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
-        self.thread.daemon = True
-        self.thread.start()
-
-    def _stop_read_thread(self) -> None:
-        """Signals the background read thread to stop and waits for it to join."""
-        if self.stop_event is not None:
-            self.stop_event.set()
-
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=2.0)
-
-        self.thread = None
-        self.stop_event = None
-
-    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
-        """
-        Reads the latest available frame asynchronously.
-
-        This method retrieves the most recent frame captured by the background
-        read thread. It does not block waiting for the camera hardware directly,
-        but may wait up to timeout_ms for the background thread to provide a frame.
-
-        Args:
-            timeout_ms (float): Maximum time in milliseconds to wait for a frame
-                to become available. Defaults to 200ms (0.2 seconds).
-
-        Returns:
-            np.ndarray: The latest captured frame as a NumPy array in the format
-                       (height, width, channels), processed according to configuration.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is not connected.
-            TimeoutError: If no frame becomes available within the specified timeout.
-            RuntimeError: If an unexpected error occurs.
-        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
-        if self.thread is None or not self.thread.is_alive():
-            self._start_read_thread()
-
-        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
-            thread_alive = self.thread is not None and self.thread.is_alive()
-            raise TimeoutError(
-                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {thread_alive}."
-            )
-
-        with self.frame_lock:
-            frame = self.latest_frame
-            self.new_frame_event.clear()
-
-        if frame is None:
-            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
-
-        return frame
-
-    def disconnect(self):
-        """
-        Disconnects from the camera and cleans up resources.
-
-        Stops the background read thread (if running) and releases the OpenCV
-        VideoCapture object.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is already disconnected.
-        """
-        if not self.is_connected and self.thread is None:
-            raise DeviceNotConnectedError(f"{self} not connected.")
-
-        if self.thread is not None:
-            self._stop_read_thread()
-
-        if self.videocapture is not None:
-            self.videocapture.release()
-            self.videocapture = None
-
-        logger.info(f"{self} disconnected.")

lerobot/common/cameras/opencv/configuration_opencv.py

@@ -1,73 +0,0 @@
-# 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 dataclasses import dataclass
-from pathlib import Path
-
-from ..configs import CameraConfig, ColorMode, Cv2Rotation
-
-
-@CameraConfig.register_subclass("opencv")
-@dataclass
-class OpenCVCameraConfig(CameraConfig):
-    """Configuration class for OpenCV-based camera devices or video files.
-
-    This class provides configuration options for cameras accessed through OpenCV,
-    supporting both physical camera devices and video files. It includes settings
-    for resolution, frame rate, color mode, and image rotation.
-
-    Example configurations:
-    ```python
-    # Basic configurations
-    OpenCVCameraConfig(0, 30, 1280, 720)   # 1280x720 @ 30FPS
-    OpenCVCameraConfig(/dev/video4, 60, 640, 480)   # 640x480 @ 60FPS
-
-    # Advanced configurations
-    OpenCVCameraConfig(128422271347, 30, 640, 480, rotation=Cv2Rotation.ROTATE_90)     # With 90° rotation
-    ```
-
-    Attributes:
-        index_or_path: Either an integer representing the camera device index,
-                      or a Path object pointing to a video file.
-        fps: Requested frames per second for the color stream.
-        width: Requested frame width in pixels for the color stream.
-        height: Requested frame height in pixels for the color stream.
-        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
-        rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
-        warmup_s: Time reading frames before returning from connect (in seconds)
-
-    Note:
-        - Only 3-channel color output (RGB/BGR) is currently supported.
-    """
-
-    index_or_path: int | Path
-    color_mode: ColorMode = ColorMode.RGB
-    rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
-    warmup_s: int = 1
-
-    def __post_init__(self):
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
-            )
-
-        if self.rotation not in (
-            Cv2Rotation.NO_ROTATION,
-            Cv2Rotation.ROTATE_90,
-            Cv2Rotation.ROTATE_180,
-            Cv2Rotation.ROTATE_270,
-        ):
-            raise ValueError(
-                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
-            )

lerobot/common/cameras/realsense/__init__.py

@@ -1,16 +0,0 @@
-# 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 .camera_realsense import RealSenseCamera
-from .configuration_realsense import RealSenseCameraConfig

lerobot/common/cameras/realsense/camera_realsense.py

@@ -1,557 +0,0 @@
-# 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.
-
-"""
-Provides the RealSenseCamera class for capturing frames from Intel RealSense cameras.
-"""
-
-import logging
-import time
-from threading import Event, Lock, Thread
-from typing import Any, Dict, List
-
-import cv2
-import numpy as np
-
-try:
-    import pyrealsense2 as rs
-except Exception as e:
-    logging.info(f"Could not import realsense: {e}")
-
-from lerobot.common.errors import DeviceAlreadyConnectedError, DeviceNotConnectedError
-
-from ..camera import Camera
-from ..configs import ColorMode
-from ..utils import get_cv2_rotation
-from .configuration_realsense import RealSenseCameraConfig
-
-logger = logging.getLogger(__name__)
-
-
-class RealSenseCamera(Camera):
-    """
-    Manages interactions with Intel RealSense cameras for frame and depth recording.
-
-    This class provides an interface similar to `OpenCVCamera` but tailored for
-    RealSense devices, leveraging the `pyrealsense2` library. It uses the camera's
-    unique serial number for identification, offering more stability than device
-    indices, especially on Linux. It also supports capturing depth maps alongside
-    color frames.
-
-    Use the provided utility script to find available camera indices and default profiles:
-    ```bash
-    python -m lerobot.find_cameras realsense
-    ```
-
-    A `RealSenseCamera` instance requires a configuration object specifying the
-    camera's serial number or a unique device name. If using the name, ensure only
-    one camera with that name is connected.
-
-    The camera's default settings (FPS, resolution, color mode) from the stream
-    profile are used unless overridden in the configuration.
-
-    Example:
-        ```python
-        from lerobot.common.cameras.realsense import RealSenseCamera, RealSenseCameraConfig
-        from lerobot.common.cameras import ColorMode, Cv2Rotation
-
-        # Basic usage with serial number
-        config = RealSenseCameraConfig(serial_number_or_name="0123456789") # Replace with actual SN
-        camera = RealSenseCamera(config)
-        camera.connect()
-
-        # Read 1 frame synchronously
-        color_image = camera.read()
-        print(color_image.shape)
-
-        # Read 1 frame asynchronously
-        async_image = camera.async_read()
-
-        # When done, properly disconnect the camera using
-        camera.disconnect()
-
-        # Example with depth capture and custom settings
-        custom_config = RealSenseCameraConfig(
-            serial_number_or_name="0123456789", # Replace with actual SN
-            fps=30,
-            width=1280,
-            height=720,
-            color_mode=ColorMode.BGR, # Request BGR output
-            rotation=Cv2Rotation.NO_ROTATION,
-            use_depth=True
-        )
-        depth_camera = RealSenseCamera(custom_config)
-        depth_camera.connect()
-
-        # Read 1 depth frame
-        depth_map = depth_camera.read_depth()
-
-        # Example using a unique camera name
-        name_config = RealSenseCameraConfig(serial_number_or_name="Intel RealSense D435") # If unique
-        name_camera = RealSenseCamera(name_config)
-        # ... connect, read, disconnect ...
-        ```
-    """
-
-    def __init__(self, config: RealSenseCameraConfig):
-        """
-        Initializes the RealSenseCamera instance.
-
-        Args:
-            config: The configuration settings for the camera.
-        """
-
-        super().__init__(config)
-
-        self.config = config
-
-        if config.serial_number_or_name.isdigit():
-            self.serial_number = config.serial_number_or_name
-        else:
-            self.serial_number = self._find_serial_number_from_name(config.serial_number_or_name)
-
-        self.fps = config.fps
-        self.color_mode = config.color_mode
-        self.use_depth = config.use_depth
-        self.warmup_s = config.warmup_s
-
-        self.rs_pipeline: rs.pipeline | None = None
-        self.rs_profile: rs.pipeline_profile | None = None
-
-        self.thread: Thread | None = None
-        self.stop_event: Event | None = None
-        self.frame_lock: Lock = Lock()
-        self.latest_frame: np.ndarray | None = None
-        self.new_frame_event: Event = Event()
-
-        self.rotation: int | None = get_cv2_rotation(config.rotation)
-
-        if self.height and self.width:
-            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-                self.capture_width, self.capture_height = self.height, self.width
-            else:
-                self.capture_width, self.capture_height = self.width, self.height
-
-    def __str__(self) -> str:
-        return f"{self.__class__.__name__}({self.serial_number})"
-
-    @property
-    def is_connected(self) -> bool:
-        """Checks if the camera pipeline is started and streams are active."""
-        return self.rs_pipeline is not None and self.rs_profile is not None
-
-    def connect(self, warmup: bool = True):
-        """
-        Connects to the RealSense camera specified in the configuration.
-
-        Initializes the RealSense pipeline, configures the required streams (color
-        and optionally depth), starts the pipeline, and validates the actual stream settings.
-
-        Raises:
-            DeviceAlreadyConnectedError: If the camera is already connected.
-            ValueError: If the configuration is invalid (e.g., missing serial/name, name not unique).
-            ConnectionError: If the camera is found but fails to start the pipeline or no RealSense devices are detected at all.
-            RuntimeError: If the pipeline starts but fails to apply requested settings.
-        """
-        if self.is_connected:
-            raise DeviceAlreadyConnectedError(f"{self} is already connected.")
-
-        self.rs_pipeline = rs.pipeline()
-        rs_config = rs.config()
-        self._configure_rs_pipeline_config(rs_config)
-
-        try:
-            self.rs_profile = self.rs_pipeline.start(rs_config)
-        except RuntimeError as e:
-            self.rs_profile = None
-            self.rs_pipeline = None
-            raise ConnectionError(
-                f"Failed to open {self}."
-                "Run `python -m lerobot.find_cameras realsense` to find available cameras."
-            ) from e
-
-        self._configure_capture_settings()
-
-        if warmup:
-            time.sleep(
-                1
-            )  # NOTE(Steven): RS cameras need a bit of time to warm up before the first read. If we don't wait, the first read from the warmup will raise.
-            start_time = time.time()
-            while time.time() - start_time < self.warmup_s:
-                self.read()
-                time.sleep(0.1)
-
-        logger.info(f"{self} connected.")
-
-    @staticmethod
-    def find_cameras() -> List[Dict[str, Any]]:
-        """
-        Detects available Intel RealSense cameras connected to the system.
-
-        Returns:
-            List[Dict[str, Any]]: A list of dictionaries,
-            where each dictionary contains 'type', 'id' (serial number), 'name',
-            firmware version, USB type, and other available specs, and the default profile properties (width, height, fps, format).
-
-        Raises:
-            OSError: If pyrealsense2 is not installed.
-            ImportError: If pyrealsense2 is not installed.
-        """
-        found_cameras_info = []
-        context = rs.context()
-        devices = context.query_devices()
-
-        for device in devices:
-            camera_info = {
-                "name": device.get_info(rs.camera_info.name),
-                "type": "RealSense",
-                "id": device.get_info(rs.camera_info.serial_number),
-                "firmware_version": device.get_info(rs.camera_info.firmware_version),
-                "usb_type_descriptor": device.get_info(rs.camera_info.usb_type_descriptor),
-                "physical_port": device.get_info(rs.camera_info.physical_port),
-                "product_id": device.get_info(rs.camera_info.product_id),
-                "product_line": device.get_info(rs.camera_info.product_line),
-            }
-
-            # Get stream profiles for each sensor
-            sensors = device.query_sensors()
-            for sensor in sensors:
-                profiles = sensor.get_stream_profiles()
-
-                for profile in profiles:
-                    if profile.is_video_stream_profile() and profile.is_default():
-                        vprofile = profile.as_video_stream_profile()
-                        stream_info = {
-                            "stream_type": vprofile.stream_name(),
-                            "format": vprofile.format().name,
-                            "width": vprofile.width(),
-                            "height": vprofile.height(),
-                            "fps": vprofile.fps(),
-                        }
-                        camera_info["default_stream_profile"] = stream_info
-
-            found_cameras_info.append(camera_info)
-
-        return found_cameras_info
-
-    def _find_serial_number_from_name(self, name: str) -> str:
-        """Finds the serial number for a given unique camera name."""
-        camera_infos = self.find_cameras()
-        found_devices = [cam for cam in camera_infos if str(cam["name"]) == name]
-
-        if not found_devices:
-            available_names = [cam["name"] for cam in camera_infos]
-            raise ValueError(
-                f"No RealSense camera found with name '{name}'. Available camera names: {available_names}"
-            )
-
-        if len(found_devices) > 1:
-            serial_numbers = [dev["serial_number"] for dev in found_devices]
-            raise ValueError(
-                f"Multiple RealSense cameras found with name '{name}'. "
-                f"Please use a unique serial number instead. Found SNs: {serial_numbers}"
-            )
-
-        serial_number = str(found_devices[0]["serial_number"])
-        return serial_number
-
-    def _configure_rs_pipeline_config(self, rs_config):
-        """Creates and configures the RealSense pipeline configuration object."""
-        rs.config.enable_device(rs_config, self.serial_number)
-
-        if self.width and self.height and self.fps:
-            rs_config.enable_stream(
-                rs.stream.color, self.capture_width, self.capture_height, rs.format.rgb8, self.fps
-            )
-            if self.use_depth:
-                rs_config.enable_stream(
-                    rs.stream.depth, self.capture_width, self.capture_height, rs.format.z16, self.fps
-                )
-        else:
-            rs_config.enable_stream(rs.stream.color)
-            if self.use_depth:
-                rs_config.enable_stream(rs.stream.depth)
-
-    def _configure_capture_settings(self) -> None:
-        """Sets fps, width, and height from device stream if not already configured.
-
-        Uses the color stream profile to update unset attributes. Handles rotation by
-        swapping width/height when needed. Original capture dimensions are always stored.
-
-        Raises:
-            DeviceNotConnectedError: If device is not connected.
-        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"Cannot validate settings for {self} as it is not connected.")
-
-        stream = self.rs_profile.get_stream(rs.stream.color).as_video_stream_profile()
-
-        if self.fps is None:
-            self.fps = stream.fps()
-
-        if self.width is None or self.height is None:
-            actual_width = int(round(stream.width()))
-            actual_height = int(round(stream.height()))
-            if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-                self.width, self.height = actual_height, actual_width
-                self.capture_width, self.capture_height = actual_width, actual_height
-            else:
-                self.width, self.height = actual_width, actual_height
-                self.capture_width, self.capture_height = actual_width, actual_height
-
-    def read_depth(self, timeout_ms: int = 200) -> np.ndarray:
-        """
-        Reads a single frame (depth) synchronously from the camera.
-
-        This is a blocking call. It waits for a coherent set of frames (depth)
-        from the camera hardware via the RealSense pipeline.
-
-        Args:
-            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
-
-        Returns:
-            np.ndarray: The depth map as a NumPy array (height, width)
-                  of type `np.uint16` (raw depth values in millimeters) and rotation.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If reading frames from the pipeline fails or frames are invalid.
-        """
-
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-        if not self.use_depth:
-            raise RuntimeError(
-                f"Failed to capture depth frame '.read_depth()'. Depth stream is not enabled for {self}."
-            )
-
-        start_time = time.perf_counter()
-
-        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
-
-        if not ret or frame is None:
-            raise RuntimeError(f"{self} read_depth failed (status={ret}).")
-
-        depth_frame = frame.get_depth_frame()
-        depth_map = np.asanyarray(depth_frame.get_data())
-
-        depth_map_processed = self._postprocess_image(depth_map, depth_frame=True)
-
-        read_duration_ms = (time.perf_counter() - start_time) * 1e3
-        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
-
-        return depth_map_processed
-
-    def read(self, color_mode: ColorMode | None = None, timeout_ms: int = 200) -> np.ndarray:
-        """
-        Reads a single frame (color) synchronously from the camera.
-
-        This is a blocking call. It waits for a coherent set of frames (color)
-        from the camera hardware via the RealSense pipeline.
-
-        Args:
-            timeout_ms (int): Maximum time in milliseconds to wait for a frame. Defaults to 200ms.
-
-        Returns:
-            np.ndarray: The captured color frame as a NumPy array
-              (height, width, channels), processed according to `color_mode` and rotation.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is not connected.
-            RuntimeError: If reading frames from the pipeline fails or frames are invalid.
-            ValueError: If an invalid `color_mode` is requested.
-        """
-
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
-        start_time = time.perf_counter()
-
-        ret, frame = self.rs_pipeline.try_wait_for_frames(timeout_ms=timeout_ms)
-
-        if not ret or frame is None:
-            raise RuntimeError(f"{self} read failed (status={ret}).")
-
-        color_frame = frame.get_color_frame()
-        color_image_raw = np.asanyarray(color_frame.get_data())
-
-        color_image_processed = self._postprocess_image(color_image_raw, color_mode)
-
-        read_duration_ms = (time.perf_counter() - start_time) * 1e3
-        logger.debug(f"{self} read took: {read_duration_ms:.1f}ms")
-
-        return color_image_processed
-
-    def _postprocess_image(
-        self, image: np.ndarray, color_mode: ColorMode | None = None, depth_frame: bool = False
-    ) -> np.ndarray:
-        """
-        Applies color conversion, dimension validation, and rotation to a raw color frame.
-
-        Args:
-            image (np.ndarray): The raw image frame (expected RGB format from RealSense).
-            color_mode (Optional[ColorMode]): The target color mode (RGB or BGR). If None,
-                                             uses the instance's default `self.color_mode`.
-
-        Returns:
-            np.ndarray: The processed image frame according to `self.color_mode` and `self.rotation`.
-
-        Raises:
-            ValueError: If the requested `color_mode` is invalid.
-            RuntimeError: If the raw frame dimensions do not match the configured
-                          `width` and `height`.
-        """
-
-        if color_mode and color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"Invalid requested color mode '{color_mode}'. Expected {ColorMode.RGB} or {ColorMode.BGR}."
-            )
-
-        if depth_frame:
-            h, w = image.shape
-        else:
-            h, w, c = image.shape
-
-            if c != 3:
-                raise RuntimeError(f"{self} frame channels={c} do not match expected 3 channels (RGB/BGR).")
-
-        if h != self.capture_height or w != self.capture_width:
-            raise RuntimeError(
-                f"{self} frame width={w} or height={h} do not match configured width={self.capture_width} or height={self.capture_height}."
-            )
-
-        processed_image = image
-        if self.color_mode == ColorMode.BGR:
-            processed_image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
-
-        if self.rotation in [cv2.ROTATE_90_CLOCKWISE, cv2.ROTATE_90_COUNTERCLOCKWISE]:
-            processed_image = cv2.rotate(processed_image, self.rotation)
-
-        return processed_image
-
-    def _read_loop(self):
-        """
-        Internal loop run by the background thread for asynchronous reading.
-
-        On each iteration:
-        1. Reads a color frame with 500ms timeout
-        2. Stores result in latest_frame (thread-safe)
-        3. Sets new_frame_event to notify listeners
-
-        Stops on DeviceNotConnectedError, logs other errors and continues.
-        """
-        while not self.stop_event.is_set():
-            try:
-                color_image = self.read(timeout_ms=500)
-
-                with self.frame_lock:
-                    self.latest_frame = color_image
-                self.new_frame_event.set()
-
-            except DeviceNotConnectedError:
-                break
-            except Exception as e:
-                logger.warning(f"Error reading frame in background thread for {self}: {e}")
-
-    def _start_read_thread(self) -> None:
-        """Starts or restarts the background read thread if it's not running."""
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=0.1)
-        if self.stop_event is not None:
-            self.stop_event.set()
-
-        self.stop_event = Event()
-        self.thread = Thread(target=self._read_loop, args=(), name=f"{self}_read_loop")
-        self.thread.daemon = True
-        self.thread.start()
-
-    def _stop_read_thread(self):
-        """Signals the background read thread to stop and waits for it to join."""
-        if self.stop_event is not None:
-            self.stop_event.set()
-
-        if self.thread is not None and self.thread.is_alive():
-            self.thread.join(timeout=2.0)
-
-        self.thread = None
-        self.stop_event = None
-
-    # NOTE(Steven): Missing implementation for depth for now
-    def async_read(self, timeout_ms: float = 200) -> np.ndarray:
-        """
-        Reads the latest available frame data (color) asynchronously.
-
-        This method retrieves the most recent color frame captured by the background
-        read thread. It does not block waiting for the camera hardware directly,
-        but may wait up to timeout_ms for the background thread to provide a frame.
-
-        Args:
-            timeout_ms (float): Maximum time in milliseconds to wait for a frame
-                to become available. Defaults to 200ms (0.2 seconds).
-
-        Returns:
-            np.ndarray:
-            The latest captured frame data (color image), processed according to configuration.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is not connected.
-            TimeoutError: If no frame data becomes available within the specified timeout.
-            RuntimeError: If the background thread died unexpectedly or another error occurs.
-        """
-        if not self.is_connected:
-            raise DeviceNotConnectedError(f"{self} is not connected.")
-
-        if self.thread is None or not self.thread.is_alive():
-            self._start_read_thread()
-
-        if not self.new_frame_event.wait(timeout=timeout_ms / 1000.0):
-            thread_alive = self.thread is not None and self.thread.is_alive()
-            raise TimeoutError(
-                f"Timed out waiting for frame from camera {self} after {timeout_ms} ms. "
-                f"Read thread alive: {thread_alive}."
-            )
-
-        with self.frame_lock:
-            frame = self.latest_frame
-            self.new_frame_event.clear()
-
-        if frame is None:
-            raise RuntimeError(f"Internal error: Event set but no frame available for {self}.")
-
-        return frame
-
-    def disconnect(self):
-        """
-        Disconnects from the camera, stops the pipeline, and cleans up resources.
-
-        Stops the background read thread (if running) and stops the RealSense pipeline.
-
-        Raises:
-            DeviceNotConnectedError: If the camera is already disconnected (pipeline not running).
-        """
-
-        if not self.is_connected and self.thread is None:
-            raise DeviceNotConnectedError(
-                f"Attempted to disconnect {self}, but it appears already disconnected."
-            )
-
-        if self.thread is not None:
-            self._stop_read_thread()
-
-        if self.rs_pipeline is not None:
-            self.rs_pipeline.stop()
-            self.rs_pipeline = None
-            self.rs_profile = None
-
-        logger.info(f"{self} disconnected.")

lerobot/common/cameras/realsense/configuration_realsense.py

@@ -1,82 +0,0 @@
-# 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 dataclasses import dataclass
-
-from ..configs import CameraConfig, ColorMode, Cv2Rotation
-
-
-@CameraConfig.register_subclass("intelrealsense")
-@dataclass
-class RealSenseCameraConfig(CameraConfig):
-    """Configuration class for Intel RealSense cameras.
-
-    This class provides specialized configuration options for Intel RealSense cameras,
-    including support for depth sensing and device identification via serial number or name.
-
-    Example configurations for Intel RealSense D405:
-    ```python
-    # Basic configurations
-    RealSenseCameraConfig("0123456789", 30, 1280, 720)   # 1280x720 @ 30FPS
-    RealSenseCameraConfig("0123456789", 60, 640, 480)   # 640x480 @ 60FPS
-
-    # Advanced configurations
-    RealSenseCameraConfig("0123456789", 30, 640, 480, use_depth=True)  # With depth sensing
-    RealSenseCameraConfig("0123456789", 30, 640, 480, rotation=Cv2Rotation.ROTATE_90)     # With 90° rotation
-    ```
-
-    Attributes:
-        fps: Requested frames per second for the color stream.
-        width: Requested frame width in pixels for the color stream.
-        height: Requested frame height in pixels for the color stream.
-        serial_number_or_name: Unique serial number or human-readable name to identify the camera.
-        color_mode: Color mode for image output (RGB or BGR). Defaults to RGB.
-        use_depth: Whether to enable depth stream. Defaults to False.
-        rotation: Image rotation setting (0°, 90°, 180°, or 270°). Defaults to no rotation.
-        warmup_s: Time reading frames before returning from connect (in seconds)
-
-    Note:
-        - Either name or serial_number must be specified.
-        - Depth stream configuration (if enabled) will use the same FPS as the color stream.
-        - The actual resolution and FPS may be adjusted by the camera to the nearest supported mode.
-        - For `fps`, `width` and `height`, either all of them need to be set, or none of them.
-    """
-
-    serial_number_or_name: str
-    color_mode: ColorMode = ColorMode.RGB
-    use_depth: bool = False
-    rotation: Cv2Rotation = Cv2Rotation.NO_ROTATION
-    warmup_s: int = 1
-
-    def __post_init__(self):
-        if self.color_mode not in (ColorMode.RGB, ColorMode.BGR):
-            raise ValueError(
-                f"`color_mode` is expected to be {ColorMode.RGB.value} or {ColorMode.BGR.value}, but {self.color_mode} is provided."
-            )
-
-        if self.rotation not in (
-            Cv2Rotation.NO_ROTATION,
-            Cv2Rotation.ROTATE_90,
-            Cv2Rotation.ROTATE_180,
-            Cv2Rotation.ROTATE_270,
-        ):
-            raise ValueError(
-                f"`rotation` is expected to be in {(Cv2Rotation.NO_ROTATION, Cv2Rotation.ROTATE_90, Cv2Rotation.ROTATE_180, Cv2Rotation.ROTATE_270)}, but {self.rotation} is provided."
-            )
-
-        values = (self.fps, self.width, self.height)
-        if any(v is not None for v in values) and any(v is None for v in values):
-            raise ValueError(
-                "For `fps`, `width` and `height`, either all of them need to be set, or none of them."
-            )

lerobot/common/cameras/utils.py

@@ -1,65 +0,0 @@
-#!/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 platform
-from pathlib import Path
-from typing import TypeAlias
-
-from .camera import Camera
-from .configs import CameraConfig, Cv2Rotation
-
-IndexOrPath: TypeAlias = int | Path
-
-
-def make_cameras_from_configs(camera_configs: dict[str, CameraConfig]) -> dict[str, Camera]:
-    cameras = {}
-
-    for key, cfg in camera_configs.items():
-        if cfg.type == "opencv":
-            from .opencv import OpenCVCamera
-
-            cameras[key] = OpenCVCamera(cfg)
-
-        elif cfg.type == "intelrealsense":
-            from .realsense.camera_realsense import RealSenseCamera
-
-            cameras[key] = RealSenseCamera(cfg)
-        else:
-            raise ValueError(f"The motor type '{cfg.type}' is not valid.")
-
-    return cameras
-
-
-def get_cv2_rotation(rotation: Cv2Rotation) -> int | None:
-    import cv2
-
-    if rotation == Cv2Rotation.ROTATE_90:
-        return cv2.ROTATE_90_CLOCKWISE
-    elif rotation == Cv2Rotation.ROTATE_180:
-        return cv2.ROTATE_180
-    elif rotation == Cv2Rotation.ROTATE_270:
-        return cv2.ROTATE_90_COUNTERCLOCKWISE
-    else:
-        return None
-
-
-def get_cv2_backend() -> int:
-    import cv2
-
-    if platform.system() == "Windows":
-        return cv2.CAP_AVFOUNDATION
-    else:
-        return cv2.CAP_ANY

lerobot/common/constants.py

@@ -17,15 +17,12 @@ from pathlib import Path
 
 from huggingface_hub.constants import HF_HOME
 
-OBS_ENV_STATE = "observation.environment_state"
-OBS_STATE = "observation.state"
+OBS_ENV = "observation.environment_state"
+OBS_ROBOT = "observation.state"
 OBS_IMAGE = "observation.image"
 OBS_IMAGES = "observation.images"
 ACTION = "action"
 
-ROBOTS = "robots"
-TELEOPERATORS = "teleoperators"
-
 # files & directories
 CHECKPOINTS_DIR = "checkpoints"
 LAST_CHECKPOINT_LINK = "last"
@@ -37,16 +34,12 @@ OPTIMIZER_STATE = "optimizer_state.safetensors"
 OPTIMIZER_PARAM_GROUPS = "optimizer_param_groups.json"
 SCHEDULER_STATE = "scheduler_state.json"
 
+# cache dir
+default_cache_path = Path(HF_HOME) / "lerobot"
+HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
+
 if "LEROBOT_HOME" in os.environ:
     raise ValueError(
         f"You have a 'LEROBOT_HOME' environment variable set to '{os.getenv('LEROBOT_HOME')}'.\n"
         "'LEROBOT_HOME' is deprecated, please use 'HF_LEROBOT_HOME' instead."
     )
-
-# cache dir
-default_cache_path = Path(HF_HOME) / "lerobot"
-HF_LEROBOT_HOME = Path(os.getenv("HF_LEROBOT_HOME", default_cache_path)).expanduser()
-
-# calibration dir
-default_calibration_path = HF_LEROBOT_HOME / "calibration"
-HF_LEROBOT_CALIBRATION = Path(os.getenv("HF_LEROBOT_CALIBRATION", default_calibration_path)).expanduser()

lerobot/common/datasets/aggregate.py

@@ -0,0 +1,416 @@
+import logging
+import shutil
+from pathlib import Path
+
+import pandas as pd
+import tqdm
+
+from lerobot.common.constants import HF_LEROBOT_HOME
+from lerobot.common.datasets.compute_stats import aggregate_stats
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, LeRobotDatasetMetadata
+from lerobot.common.datasets.utils import (
+    DEFAULT_CHUNK_SIZE,
+    DEFAULT_DATA_FILE_SIZE_IN_MB,
+    DEFAULT_DATA_PATH,
+    DEFAULT_EPISODES_PATH,
+    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
+    DEFAULT_VIDEO_PATH,
+    concat_video_files,
+    get_parquet_file_size_in_mb,
+    get_video_size_in_mb,
+    to_parquet_with_hf_images,
+    update_chunk_file_indices,
+    write_info,
+    write_stats,
+    write_tasks,
+)
+from lerobot.common.utils.utils import init_logging
+
+
+def validate_all_metadata(all_metadata: list[LeRobotDatasetMetadata]):
+    # validate same fps, robot_type, features
+
+    fps = all_metadata[0].fps
+    robot_type = all_metadata[0].robot_type
+    features = all_metadata[0].features
+
+    for meta in tqdm.tqdm(all_metadata, desc="Validate all meta data"):
+        if fps != meta.fps:
+            raise ValueError(f"Same fps is expected, but got fps={meta.fps} instead of {fps}.")
+        if robot_type != meta.robot_type:
+            raise ValueError(
+                f"Same robot_type is expected, but got robot_type={meta.robot_type} instead of {robot_type}."
+            )
+        if features != meta.features:
+            raise ValueError(
+                f"Same features is expected, but got features={meta.features} instead of {features}."
+            )
+
+    return fps, robot_type, features
+
+
+def update_data_df(df, src_meta, dst_meta):
+    def _update(row):
+        row["episode_index"] = row["episode_index"] + dst_meta["total_episodes"]
+        row["index"] = row["index"] + dst_meta["total_frames"]
+        task = src_meta.tasks.iloc[row["task_index"]].name
+        row["task_index"] = dst_meta.tasks.loc[task].task_index.item()
+        return row
+
+    return df.apply(_update, axis=1)
+
+
+def update_meta_data(
+    df,
+    dst_meta,
+    meta_idx,
+    data_idx,
+    videos_idx,
+):
+    def _update(row):
+        row["meta/episodes/chunk_index"] = row["meta/episodes/chunk_index"] + meta_idx["chunk_index"]
+        row["meta/episodes/file_index"] = row["meta/episodes/file_index"] + meta_idx["file_index"]
+        row["data/chunk_index"] = row["data/chunk_index"] + data_idx["chunk_index"]
+        row["data/file_index"] = row["data/file_index"] + data_idx["file_index"]
+        for key, video_idx in videos_idx.items():
+            row[f"videos/{key}/chunk_index"] = row[f"videos/{key}/chunk_index"] + video_idx["chunk_index"]
+            row[f"videos/{key}/file_index"] = row[f"videos/{key}/file_index"] + video_idx["file_index"]
+            row[f"videos/{key}/from_timestamp"] = (
+                row[f"videos/{key}/from_timestamp"] + video_idx["latest_duration"]
+            )
+            row[f"videos/{key}/to_timestamp"] = (
+                row[f"videos/{key}/to_timestamp"] + video_idx["latest_duration"]
+            )
+        row["dataset_from_index"] = row["dataset_from_index"] + dst_meta.info["total_frames"]
+        row["dataset_to_index"] = row["dataset_to_index"] + dst_meta.info["total_frames"]
+        row["episode_index"] = row["episode_index"] + dst_meta.info["total_episodes"]
+        return row
+
+    return df.apply(_update, axis=1)
+
+
+def aggregate_datasets(repo_ids: list[str], aggr_repo_id: str, roots: list[Path] = None, aggr_root=None):
+    logging.info("Start aggregate_datasets")
+
+    # Load metadata
+    all_metadata = (
+        [LeRobotDatasetMetadata(repo_id) for repo_id in repo_ids]
+        if roots is None
+        else [
+            LeRobotDatasetMetadata(repo_id, root=root) for repo_id, root in zip(repo_ids, roots, strict=False)
+        ]
+    )
+    fps, robot_type, features = validate_all_metadata(all_metadata)
+    video_keys = [key for key in features if features[key]["dtype"] == "video"]
+    image_keys = [key for key in features if features[key]["dtype"] == "image"]
+
+    # Initialize output dataset metadata
+    dst_meta = LeRobotDatasetMetadata.create(
+        repo_id=aggr_repo_id,
+        fps=fps,
+        robot_type=robot_type,
+        features=features,
+        root=aggr_root,
+    )
+
+    # Aggregate task info
+    logging.info("Find all tasks")
+    unique_tasks = pd.concat([m.tasks for m in all_metadata]).index.unique()
+    dst_meta.tasks = pd.DataFrame({"task_index": range(len(unique_tasks))}, index=unique_tasks)
+
+    # Track counters and indices
+    meta_idx = {"chunk": 0, "file": 0}
+    data_idx = {"chunk": 0, "file": 0}
+    videos_idx = {
+        key: {"chunk": 0, "file": 0, "latest_duration": 0, "episode_duration": 0} for key in video_keys
+    }
+
+    dst_meta.episodes = {}
+
+    # Process each dataset
+    for src_meta in tqdm.tqdm(all_metadata, desc="Copy data and videos"):
+        videos_idx = aggregate_videos(src_meta, dst_meta, videos_idx)
+        data_idx = aggregate_data(src_meta, dst_meta, data_idx)
+    
+        meta_idx = aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx, video_keys, image_keys)
+
+        dst_meta.info["total_episodes"] += src_meta.total_episodes
+        dst_meta.info["total_frames"] += src_meta.total_frames
+
+    finalize_aggregation(dst_meta, all_metadata)
+    logging.info("Aggregation complete.")
+
+
+# -------------------------------
+# Helper Functions
+# -------------------------------
+
+
+def aggregate_videos(src_meta, dst_meta, videos_idx):
+    """
+    Aggregates video chunks from a dataset into the aggregated dataset folder.
+    """
+    for key, video_idx in videos_idx.items():
+        # Get unique (chunk, file) combinations
+        unique_chunk_file_pairs = {
+            (chunk, file)
+            for chunk, file in zip(
+                src_meta.episodes[f"videos/{key}/chunk_index"],
+                src_meta.episodes[f"videos/{key}/file_index"],
+                strict=False,
+            )
+        }
+
+        # Current target chunk/file index
+        chunk_idx = video_idx["chunk_idx"]
+        file_idx = video_idx["file_idx"]
+
+        for src_chunk_idx, src_file_idx in unique_chunk_file_pairs:
+            src_path = src_meta.root / DEFAULT_VIDEO_PATH.format(
+                video_key=key,
+                chunk_index=src_chunk_idx,
+                file_index=src_file_idx,
+            )
+
+            dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
+                video_key=key,
+                chunk_index=chunk_idx,
+                file_index=file_idx,
+            )
+
+            if not dst_path.exists():
+                # First write to this destination file
+                dst_path.parent.mkdir(parents=True, exist_ok=True)
+                shutil.copy(str(src_path), str(dst_path))
+                continue
+
+            # Check file sizes before appending
+            src_size = get_video_size_in_mb(src_path)
+            dst_size = get_video_size_in_mb(dst_path)
+
+            if dst_size + src_size >= DEFAULT_VIDEO_FILE_SIZE_IN_MB:
+                # Rotate to a new chunk/file
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
+                dst_path = dst_meta.root / DEFAULT_VIDEO_PATH.format(
+                    video_key=key,
+                    chunk_index=chunk_idx,
+                    file_index=file_idx,
+                )
+                dst_path.parent.mkdir(parents=True, exist_ok=True)
+                shutil.copy(str(src_path), str(dst_path))
+            else:
+                # Append to existing video file
+                concat_video_files(
+                    [dst_path, src_path],
+                    dst_meta.root,
+                    key,
+                    chunk_idx,
+                    file_idx,
+                )
+
+                if src_size + dst_size >= DEFAULT_DATA_FILE_SIZE_IN_MB:
+                    # Size limit is reached, prepare new parquet file
+                    aggr_data_chunk_idx, aggr_data_file_idx = update_chunk_file_indices(
+                        aggr_data_chunk_idx, aggr_data_file_idx, DEFAULT_CHUNK_SIZE
+                    )
+                    aggr_path = aggr_root / DEFAULT_DATA_PATH.format(
+                        chunk_index=aggr_data_chunk_idx, file_index=aggr_data_file_idx
+                    )
+                    aggr_path.parent.mkdir(parents=True, exist_ok=True)
+                    df.to_parquet(aggr_path)
+                else:
+                    # Update the existing parquet file with new rows
+                    aggr_df = pd.read_parquet(aggr_path)
+                    df = pd.concat([aggr_df, df], ignore_index=True)
+                    to_parquet_with_hf_images(df, aggr_path, dst_meta.image_keys)
+
+        return videos_idx
+
+
+def aggregate_data(src_meta, dst_meta, data_idx):
+    unique_chunk_file_ids = {
+        (c, f)
+        for c, f in zip(
+            src_meta.episodes["data/chunk_index"], src_meta.episodes["data/file_index"], strict=False
+        )
+    }
+    for src_chunk_idx, src_file_idx in unique_chunk_file_ids:
+        src_path = src_meta.root / DEFAULT_DATA_PATH.format(
+            chunk_index=src_chunk_idx, file_index=src_file_idx
+        )
+        df = pd.read_parquet(src_path)
+        df = update_data_df(df, src_meta, dst_meta)
+
+        data_idx = append_or_create_parquet_file(
+            df,
+            src_path,
+            data_idx,
+            DEFAULT_DATA_FILE_SIZE_IN_MB,
+            DEFAULT_CHUNK_SIZE,
+            DEFAULT_DATA_PATH,
+            contains_images=len(dst_meta.image_keys) > 0
+        )
+
+    return data_idx
+
+
+def aggregate_metadata(src_meta, dst_meta, meta_idx, data_idx, videos_idx):
+    chunk_file_ids = {
+        (c, f)
+        for c, f in zip(
+            src_meta.episodes["meta/episodes/chunk_index"],
+            src_meta.episodes["meta/episodes/file_index"],
+            strict=False,
+        )
+    }
+
+    for chunk_idx, file_idx in chunk_file_ids:
+        src_path = src_meta.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+        df = pd.read_parquet(src_path)
+        df = update_meta_data(
+            df,
+            dst_meta,
+            meta_idx,
+            data_idx,
+            videos_idx,
+        )
+
+        # for k in video_keys:
+        #     video_idx[k]["latest_duration"] += video_idx[k]["episode_duration"]
+
+        append_or_create_parquet_file(
+            df,
+            src_path,
+            meta_idx,
+            DEFAULT_DATA_FILE_SIZE_IN_MB,
+            DEFAULT_CHUNK_SIZE,
+            DEFAULT_EPISODES_PATH,
+        )
+
+    return meta_idx
+
+
+def append_or_create_parquet_file(
+    df: pd.DataFrame,
+    src_path: Path,
+    idx: dict[str, int],
+    max_mb: float,
+    chunk_size: int,
+    default_path: str,
+    contains_images: bool = False,
+):
+    """
+    Safely appends or creates a Parquet file at dst_path based on size constraints.
+
+    Parameters:
+        df (pd.DataFrame): Data to write.
+        src_path (Path): Path to source file (used to get size).
+        idx (dict): Dictionary containing 'chunk' and 'file' indices.
+        max_mb (float): Maximum allowed file size in MB.
+        chunk_size (int): Maximum number of files per chunk.
+        default_path (str): Format string for generating a new file path.
+
+    Returns:
+        dict: Updated index dictionary.
+    """
+    # Initial destination path
+    dst_path = aggr_root / DEFAULT_DATA_PATH.format(
+        chunk_index=idx["chunk"], file_index=idx["file"]
+    )
+
+    # If destination file doesn't exist, just write the new one
+    if not dst_path.exists():
+        dst_path.parent.mkdir(parents=True, exist_ok=True)
+        df.to_parquet(dst_path)
+        return idx
+
+    # Otherwise, check if we exceed the size limit
+    src_size = get_parquet_file_size_in_mb(src_path)
+    dst_size = get_parquet_file_size_in_mb(dst_path)
+
+    if dst_size + src_size >= max_mb:
+        # File is too large, move to a new one
+        idx["chunk"], idx["file"] = update_chunk_file_indices(idx["chunk"], idx["file"], chunk_size)
+        new_path = dst_path.parent / default_path.format(chunk_index=idx["chunk"], file_index=idx["file"])
+        new_path.parent.mkdir(parents=True, exist_ok=True)
+        final_df = df
+    else:
+        # Append to existing file
+        existing_df = pd.read_parquet(dst_path)
+        final_df = pd.concat([existing_df, df], ignore_index=True)
+
+    if contains_images:
+        to_parquet_with_hf_images(final_df, new_path)
+    else:
+        final_df.to_parquet(new_path)
+
+    return idx
+
+
+def finalize_aggregation(aggr_meta, all_metadata):
+    logging.info("write tasks")
+    write_tasks(aggr_meta.tasks, aggr_meta.root)
+
+    logging.info("write info")
+    aggr_meta.info.update(
+        {
+            "total_tasks": len(aggr_meta.tasks),
+            "total_episodes": sum(m.total_episodes for m in all_metadata),
+            "total_frames": sum(m.total_frames for m in all_metadata),
+            "splits": {"train": f"0:{sum(m.total_episodes for m in all_metadata)}"},
+        }
+    )
+    write_info(aggr_meta.info, aggr_meta.root)
+
+    logging.info("write stats")
+    aggr_meta.stats = aggregate_stats([m.stats for m in all_metadata])
+    write_stats(aggr_meta.stats, aggr_meta.root)
+
+
+if __name__ == "__main__":
+    init_logging()
+
+    num_shards = 2048
+    repo_id = "cadene/droid_1.0.1_v30"
+    aggr_repo_id = f"{repo_id}_compact_6"
+    tags = ["openx"]
+
+    # num_shards = 210
+    # repo_id = "cadene/agibot_alpha_v30"
+    # aggr_repo_id = f"{repo_id}"
+    # tags = None
+
+    # aggr_root = Path(f"/tmp/{aggr_repo_id}")
+    aggr_root = HF_LEROBOT_HOME / aggr_repo_id
+    if aggr_root.exists():
+        shutil.rmtree(aggr_root)
+
+    repo_ids = []
+    roots = []
+    for rank in range(num_shards):
+        shard_repo_id = f"{repo_id}_world_{num_shards}_rank_{rank}"
+        shard_root = HF_LEROBOT_HOME / shard_repo_id
+        try:
+            meta = LeRobotDatasetMetadata(shard_repo_id, root=shard_root)
+            if len(meta.video_keys) == 0:
+                continue
+            repo_ids.append(shard_repo_id)
+            roots.append(shard_root)
+        except:
+            pass
+
+        if rank == 1:
+            break
+
+    aggregate_datasets(
+        repo_ids,
+        aggr_repo_id,
+        roots=roots,
+        aggr_root=aggr_root,
+    )
+
+    aggr_dataset = LeRobotDataset(repo_id=aggr_repo_id, root=aggr_root)
+    # for i in tqdm.tqdm(range(len(aggr_dataset))):
+    #     aggr_dataset[i]
+    #     pass
+    aggr_dataset.push_to_hub(tags=tags, upload_large_folder=True)

lerobot/common/datasets/backward_compatibility.py

@@ -47,6 +47,18 @@ If you encounter a problem, contact LeRobot maintainers on [Discord](https://dis
 or open an [issue on GitHub](https://github.com/huggingface/lerobot/issues/new/choose).
 """
 
+V30_MESSAGE = """
+The dataset you requested ({repo_id}) is in {version} format.
+While current version of LeRobot is backward-compatible with it, the version of your dataset still uses global
+stats instead of per-episode stats. Update your dataset stats to the new format using this command:
+```
+python lerobot/common/datasets/v30/convert_dataset_v21_to_v30.py --repo-id={repo_id}
+```
+
+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).
+"""
+
 FUTURE_MESSAGE = """
 The dataset you requested ({repo_id}) is only available in {version} format.
 As we cannot ensure forward compatibility with it, please update your current version of lerobot.
@@ -58,7 +70,14 @@ class CompatibilityError(Exception): ...
 
 class BackwardCompatibilityError(CompatibilityError):
     def __init__(self, repo_id: str, version: packaging.version.Version):
-        message = V2_MESSAGE.format(repo_id=repo_id, version=version)
+        if version.major == 3:
+            message = V30_MESSAGE.format(repo_id=repo_id, version=version)
+        elif version.major == 2:
+            message = V2_MESSAGE.format(repo_id=repo_id, version=version)
+        else:
+            raise NotImplementedError(
+                "Contact the maintainer on [Discord](https://discord.com/invite/s3KuuzsPFb)."
+            )
         super().__init__(message)
 
 

lerobot/common/datasets/factory.py

@@ -49,7 +49,7 @@ def resolve_delta_timestamps(
                 "observation.state": [-0.04, -0.02, 0]
                 "observation.action": [-0.02, 0, 0.02]
             }
-            returns `None` if the resulting dict is empty.
+            returns `None` if the the resulting dict is empty.
     """
     delta_timestamps = {}
     for key in ds_meta.features:

lerobot/common/datasets/image_writer.py

@@ -106,7 +106,7 @@ def worker_process(queue: queue.Queue, num_threads: int):
 class AsyncImageWriter:
     """
     This class abstract away the initialisation of processes or/and threads to
-    save images on disk asynchronously, which is critical to control a robot and record data
+    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`.

lerobot/common/datasets/lerobot_dataset.py

@@ -16,16 +16,18 @@
 import contextlib
 import logging
 import shutil
+import tempfile
 from pathlib import Path
 from typing import Callable
 
 import datasets
 import numpy as np
 import packaging.version
+import pandas as pd
 import PIL.Image
 import torch
 import torch.utils
-from datasets import concatenate_datasets, load_dataset
+from datasets import Dataset
 from huggingface_hub import HfApi, snapshot_download
 from huggingface_hub.constants import REPOCARD_NAME
 from huggingface_hub.errors import RevisionNotFoundError
@@ -34,35 +36,41 @@ from lerobot.common.constants import HF_LEROBOT_HOME
 from lerobot.common.datasets.compute_stats import aggregate_stats, compute_episode_stats
 from lerobot.common.datasets.image_writer import AsyncImageWriter, write_image
 from lerobot.common.datasets.utils import (
+    DEFAULT_EPISODES_PATH,
     DEFAULT_FEATURES,
     DEFAULT_IMAGE_PATH,
     INFO_PATH,
-    TASKS_PATH,
-    append_jsonlines,
-    backward_compatible_episodes_stats,
     check_delta_timestamps,
-    check_timestamps_sync,
     check_version_compatibility,
+    concat_video_files,
     create_empty_dataset_info,
     create_lerobot_dataset_card,
     embed_images,
+    flatten_dict,
     get_delta_indices,
-    get_episode_data_index,
+    get_features_from_robot,
+    get_hf_dataset_size_in_mb,
     get_hf_features_from_features,
+    get_parquet_file_size_in_mb,
+    get_parquet_num_frames,
     get_safe_version,
+    get_video_duration_in_s,
+    get_video_size_in_mb,
     hf_transform_to_torch,
     is_valid_version,
     load_episodes,
-    load_episodes_stats,
     load_info,
+    load_nested_dataset,
     load_stats,
     load_tasks,
+    to_parquet_with_hf_images,
+    update_chunk_file_indices,
     validate_episode_buffer,
     validate_frame,
-    write_episode,
-    write_episode_stats,
     write_info,
     write_json,
+    write_stats,
+    write_tasks,
 )
 from lerobot.common.datasets.video_utils import (
     VideoFrame,
@@ -71,8 +79,9 @@ from lerobot.common.datasets.video_utils import (
     get_safe_default_codec,
     get_video_info,
 )
+from lerobot.common.robot_devices.robots.utils import Robot
 
-CODEBASE_VERSION = "v2.1"
+CODEBASE_VERSION = "v3.0"
 
 
 class LeRobotDatasetMetadata:
@@ -96,20 +105,18 @@ class LeRobotDatasetMetadata:
                 self.revision = get_safe_version(self.repo_id, self.revision)
 
             (self.root / "meta").mkdir(exist_ok=True, parents=True)
+            # TODO(rcadene): instead of downloading all episodes metadata files,
+            # download only the ones associated to the requested episodes. This would
+            # require adding `episodes: list[int]` as argument.
             self.pull_from_repo(allow_patterns="meta/")
             self.load_metadata()
 
     def load_metadata(self):
         self.info = load_info(self.root)
         check_version_compatibility(self.repo_id, self._version, CODEBASE_VERSION)
-        self.tasks, self.task_to_task_index = load_tasks(self.root)
+        self.tasks = load_tasks(self.root)
         self.episodes = load_episodes(self.root)
-        if self._version < packaging.version.parse("v2.1"):
-            self.stats = load_stats(self.root)
-            self.episodes_stats = backward_compatible_episodes_stats(self.stats, self.episodes)
-        else:
-            self.episodes_stats = load_episodes_stats(self.root)
-            self.stats = aggregate_stats(list(self.episodes_stats.values()))
+        self.stats = load_stats(self.root)
 
     def pull_from_repo(
         self,
@@ -131,18 +138,19 @@ class LeRobotDatasetMetadata:
         return packaging.version.parse(self.info["codebase_version"])
 
     def get_data_file_path(self, ep_index: int) -> Path:
-        ep_chunk = self.get_episode_chunk(ep_index)
-        fpath = self.data_path.format(episode_chunk=ep_chunk, episode_index=ep_index)
+        ep = self.episodes[ep_index]
+        chunk_idx = ep["data/chunk_index"]
+        file_idx = ep["data/file_index"]
+        fpath = self.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
         return Path(fpath)
 
     def get_video_file_path(self, ep_index: int, vid_key: str) -> Path:
-        ep_chunk = self.get_episode_chunk(ep_index)
-        fpath = self.video_path.format(episode_chunk=ep_chunk, video_key=vid_key, episode_index=ep_index)
+        ep = self.episodes[ep_index]
+        chunk_idx = ep[f"videos/{vid_key}/chunk_index"]
+        file_idx = ep[f"videos/{vid_key}/file_index"]
+        fpath = self.video_path.format(video_key=vid_key, chunk_index=chunk_idx, file_index=file_idx)
         return Path(fpath)
 
-    def get_episode_chunk(self, ep_index: int) -> int:
-        return ep_index // self.chunks_size
-
     @property
     def data_path(self) -> str:
         """Formattable string for the parquet files."""
@@ -209,39 +217,108 @@ class LeRobotDatasetMetadata:
         return self.info["total_tasks"]
 
     @property
-    def total_chunks(self) -> int:
-        """Total number of chunks (groups of episodes)."""
-        return self.info["total_chunks"]
+    def chunks_size(self) -> int:
+        """Max number of files per chunk."""
+        return self.info["chunks_size"]
 
     @property
-    def chunks_size(self) -> int:
-        """Max number of episodes per chunk."""
-        return self.info["chunks_size"]
+    def data_files_size_in_mb(self) -> int:
+        """Max size of data file in mega bytes."""
+        return self.info["data_files_size_in_mb"]
+
+    @property
+    def video_files_size_in_mb(self) -> int:
+        """Max size of video file in mega bytes."""
+        return self.info["video_files_size_in_mb"]
 
     def get_task_index(self, task: str) -> int | None:
         """
         Given a task in natural language, returns its task_index if the task already exists in the dataset,
         otherwise return None.
         """
-        return self.task_to_task_index.get(task, None)
+        if task in self.tasks.index:
+            return int(self.tasks.loc[task].task_index)
+        else:
+            return None
 
-    def add_task(self, task: str):
+    def save_episode_tasks(self, tasks: list[str]):
+        if len(set(tasks)) != len(tasks):
+            raise ValueError(f"Tasks are not unique: {tasks}")
+
+        if self.tasks is None:
+            new_tasks = tasks
+            task_indices = range(len(tasks))
+            self.tasks = pd.DataFrame({"task_index": task_indices}, index=tasks)
+        else:
+            new_tasks = [task for task in tasks if task not in self.tasks.index]
+            new_task_indices = range(len(self.tasks), len(self.tasks) + len(new_tasks))
+            for task_idx, task in zip(new_task_indices, new_tasks, strict=False):
+                self.tasks.loc[task] = task_idx
+
+        if len(new_tasks) > 0:
+            # Update on disk
+            write_tasks(self.tasks, self.root)
+
+    def _save_episode_metadata(self, episode_dict: dict) -> None:
+        """Save episode metadata to a parquet file and update the Hugging Face dataset of episodes metadata.
+
+        This function processes episodes metadata from a dictionary, converts it into a Hugging Face dataset,
+        and saves it as a parquet file. It handles both the creation of new parquet files and the
+        updating of existing ones based on size constraints. After saving the metadata, it reloads
+        the Hugging Face dataset to ensure it is up-to-date.
+
+        Notes: We both need to update parquet files and HF dataset:
+        - `pandas` loads parquet file in RAM
+        - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
+          or loads directly from pyarrow cache.
         """
-        Given a task in natural language, add it to the dictionary of tasks.
-        """
-        if task in self.task_to_task_index:
-            raise ValueError(f"The task '{task}' already exists and can't be added twice.")
+        # Convert buffer into HF Dataset
+        episode_dict = {key: [value] for key, value in episode_dict.items()}
+        ep_dataset = Dataset.from_dict(episode_dict)
+        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
+        df = pd.DataFrame(ep_dataset)
+        num_frames = episode_dict["length"][0]
 
-        task_index = self.info["total_tasks"]
-        self.task_to_task_index[task] = task_index
-        self.tasks[task_index] = task
-        self.info["total_tasks"] += 1
+        if self.episodes is None:
+            # Initialize indices and frame count for a new dataset made of the first episode data
+            chunk_idx, file_idx = 0, 0
+            df["meta/episodes/chunk_index"] = [chunk_idx]
+            df["meta/episodes/file_index"] = [file_idx]
+            df["dataset_from_index"] = [0]
+            df["dataset_to_index"] = [num_frames]
+        else:
+            # Retrieve information from the latest parquet file
+            latest_ep = self.episodes[-1]
+            chunk_idx = latest_ep["meta/episodes/chunk_index"]
+            file_idx = latest_ep["meta/episodes/file_index"]
 
-        task_dict = {
-            "task_index": task_index,
-            "task": task,
-        }
-        append_jsonlines(task_dict, self.root / TASKS_PATH)
+            latest_path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
+
+            if latest_size_in_mb + ep_size_in_mb >= self.data_files_size_in_mb:
+                # Size limit is reached, prepare new parquet file
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.chunks_size)
+
+            # Update the existing pandas dataframe with new row
+            df["meta/episodes/chunk_index"] = [chunk_idx]
+            df["meta/episodes/file_index"] = [file_idx]
+            df["dataset_from_index"] = [latest_ep["dataset_to_index"]]
+            df["dataset_to_index"] = [latest_ep["dataset_to_index"] + num_frames]
+
+            if latest_size_in_mb + ep_size_in_mb < self.data_files_size_in_mb:
+                # Size limit wasnt reached, concatenate latest dataframe with new one
+                latest_df = pd.read_parquet(latest_path)
+                df = pd.concat([latest_df, df], ignore_index=True)
+
+        # Write the resulting dataframe from RAM to disk
+        path = self.root / DEFAULT_EPISODES_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+        path.parent.mkdir(parents=True, exist_ok=True)
+        df.to_parquet(path, index=False)
+
+        # Update the Hugging Face dataset by reloading it.
+        # This process should be fast because only the latest Parquet file has been modified.
+        # Therefore, only this file needs to be converted to PyArrow; the rest is loaded from the PyArrow memory-mapped cache.
+        self.episodes = load_episodes(self.root)
 
     def save_episode(
         self,
@@ -249,32 +326,28 @@ class LeRobotDatasetMetadata:
         episode_length: int,
         episode_tasks: list[str],
         episode_stats: dict[str, dict],
+        episode_metadata: dict,
     ) -> None:
-        self.info["total_episodes"] += 1
-        self.info["total_frames"] += episode_length
-
-        chunk = self.get_episode_chunk(episode_index)
-        if chunk >= self.total_chunks:
-            self.info["total_chunks"] += 1
-
-        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
-        self.info["total_videos"] += len(self.video_keys)
-        if len(self.video_keys) > 0:
-            self.update_video_info()
-
-        write_info(self.info, self.root)
-
         episode_dict = {
             "episode_index": episode_index,
             "tasks": episode_tasks,
             "length": episode_length,
         }
-        self.episodes[episode_index] = episode_dict
-        write_episode(episode_dict, self.root)
+        episode_dict.update(episode_metadata)
+        episode_dict.update(flatten_dict({"stats": episode_stats}))
+        self._save_episode_metadata(episode_dict)
 
-        self.episodes_stats[episode_index] = episode_stats
-        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats else episode_stats
-        write_episode_stats(episode_index, episode_stats, self.root)
+        # Update info
+        self.info["total_episodes"] += 1
+        self.info["total_frames"] += episode_length
+        self.info["total_tasks"] = len(self.tasks)
+        self.info["splits"] = {"train": f"0:{self.info['total_episodes']}"}
+        if len(self.video_keys) > 0:
+            self.update_video_info()
+        write_info(self.info, self.root)
+
+        self.stats = aggregate_stats([self.stats, episode_stats]) if self.stats is not None else episode_stats
+        write_stats(self.stats, self.root)
 
     def update_video_info(self) -> None:
         """
@@ -302,9 +375,10 @@ class LeRobotDatasetMetadata:
         cls,
         repo_id: str,
         fps: int,
-        features: dict,
-        robot_type: str | None = None,
         root: str | Path | None = None,
+        robot: Robot | None = None,
+        robot_type: str | None = None,
+        features: dict | None = None,
         use_videos: bool = True,
     ) -> "LeRobotDatasetMetadata":
         """Creates metadata for a LeRobotDataset."""
@@ -314,27 +388,34 @@ class LeRobotDatasetMetadata:
 
         obj.root.mkdir(parents=True, exist_ok=False)
 
-        # if robot is not None:
-        #     features = get_features_from_robot(robot, use_videos)
-        #     robot_type = robot.robot_type
-        #     if not all(cam.fps == fps for cam in robot.cameras.values()):
-        #         logging.warning(
-        #             f"Some cameras in your {robot.robot_type} robot don't have an fps matching the fps of your dataset."
-        #             "In this case, frames from lower fps cameras will be repeated to fill in the blanks."
-        #         )
+        if robot is not None:
+            features = get_features_from_robot(robot, use_videos)
+            robot_type = robot.robot_type
+            if not all(cam.fps == fps for cam in robot.cameras.values()):
+                logging.warning(
+                    f"Some cameras in your {robot.robot_type} robot don't have an fps matching the fps of your dataset."
+                    "In this case, frames from lower fps cameras will be repeated to fill in the blanks."
+                )
+        elif features is None:
+            raise ValueError(
+                "Dataset features must either come from a Robot or explicitly passed upon creation."
+            )
+        else:
+            # TODO(aliberts, rcadene): implement sanity check for features
+            features = {**features, **DEFAULT_FEATURES}
 
-        # TODO(aliberts, rcadene): implement sanity check for features
-        features = {**features, **DEFAULT_FEATURES}
+            # check if none of the features contains a "/" in their names,
+            # as this would break the dict flattening in the stats computation, which uses '/' as separator
+            for key in features:
+                if "/" in key:
+                    raise ValueError(f"Feature names should not contain '/'. Found '/' in feature '{key}'.")
 
-        # check if none of the features contains a "/" in their names,
-        # as this would break the dict flattening in the stats computation, which uses '/' as separator
-        for key in features:
-            if "/" in key:
-                raise ValueError(f"Feature names should not contain '/'. Found '/' in feature '{key}'.")
+            features = {**features, **DEFAULT_FEATURES}
 
-        obj.tasks, obj.task_to_task_index = {}, {}
-        obj.episodes_stats, obj.stats, obj.episodes = {}, {}, {}
-        obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, features, use_videos, robot_type)
+        obj.tasks = None
+        obj.episodes = None
+        obj.stats = None
+        obj.info = create_empty_dataset_info(CODEBASE_VERSION, fps, robot_type, features, use_videos)
         if len(obj.video_keys) > 0 and not use_videos:
             raise ValueError()
         write_json(obj.info, obj.root / INFO_PATH)
@@ -478,29 +559,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.meta = LeRobotDatasetMetadata(
             self.repo_id, self.root, self.revision, force_cache_sync=force_cache_sync
         )
-        if self.episodes is not None and self.meta._version >= packaging.version.parse("v2.1"):
-            episodes_stats = [self.meta.episodes_stats[ep_idx] for ep_idx in self.episodes]
-            self.stats = aggregate_stats(episodes_stats)
 
         # Load actual data
         try:
             if force_cache_sync:
                 raise FileNotFoundError
-            assert all((self.root / fpath).is_file() for fpath in self.get_episodes_file_paths())
             self.hf_dataset = self.load_hf_dataset()
         except (AssertionError, FileNotFoundError, NotADirectoryError):
             self.revision = get_safe_version(self.repo_id, self.revision)
-            self.download_episodes(download_videos)
+            self.download(download_videos)
             self.hf_dataset = self.load_hf_dataset()
 
-        self.episode_data_index = get_episode_data_index(self.meta.episodes, self.episodes)
-
-        # Check timestamps
-        timestamps = torch.stack(self.hf_dataset["timestamp"]).numpy()
-        episode_indices = torch.stack(self.hf_dataset["episode_index"]).numpy()
-        ep_data_index_np = {k: t.numpy() for k, t in self.episode_data_index.items()}
-        check_timestamps_sync(timestamps, episode_indices, ep_data_index_np, self.fps, self.tolerance_s)
-
         # Setup delta_indices
         if self.delta_timestamps is not None:
             check_delta_timestamps(self.delta_timestamps, self.fps, self.tolerance_s)
@@ -576,7 +645,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
             ignore_patterns=ignore_patterns,
         )
 
-    def download_episodes(self, download_videos: bool = True) -> None:
+    def download(self, download_videos: bool = True) -> None:
         """Downloads the dataset from the given 'repo_id' at the provided version. If 'episodes' is given, this
         will only download those episodes (selected by their episode_index). If 'episodes' is None, the whole
         dataset will be downloaded. Thanks to the behavior of snapshot_download, if the files are already present
@@ -584,11 +653,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """
         # TODO(rcadene, aliberts): implement faster transfer
         # https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads
-        files = None
         ignore_patterns = None if download_videos else "videos/"
+        files = None
         if self.episodes is not None:
             files = self.get_episodes_file_paths()
-
         self.pull_from_repo(allow_patterns=files, ignore_patterns=ignore_patterns)
 
     def get_episodes_file_paths(self) -> list[Path]:
@@ -601,19 +669,13 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 for ep_idx in episodes
             ]
             fpaths += video_files
-
+        # episodes are stored in the same files, so we return unique paths only
+        fpaths = list(set(fpaths))
         return fpaths
 
     def load_hf_dataset(self) -> datasets.Dataset:
         """hf_dataset contains all the observations, states, actions, rewards, etc."""
-        if self.episodes is None:
-            path = str(self.root / "data")
-            hf_dataset = load_dataset("parquet", data_dir=path, split="train")
-        else:
-            files = [str(self.root / self.meta.get_data_file_path(ep_idx)) for ep_idx in self.episodes]
-            hf_dataset = load_dataset("parquet", data_files=files, split="train")
-
-        # TODO(aliberts): hf_dataset.set_format("torch")
+        hf_dataset = load_nested_dataset(self.root / "data")
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
@@ -621,8 +683,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         features = get_hf_features_from_features(self.features)
         ft_dict = {col: [] for col in features}
         hf_dataset = datasets.Dataset.from_dict(ft_dict, features=features, split="train")
-
-        # TODO(aliberts): hf_dataset.set_format("torch")
         hf_dataset.set_transform(hf_transform_to_torch)
         return hf_dataset
 
@@ -654,15 +714,16 @@ class LeRobotDataset(torch.utils.data.Dataset):
             return get_hf_features_from_features(self.features)
 
     def _get_query_indices(self, idx: int, ep_idx: int) -> tuple[dict[str, list[int | bool]]]:
-        ep_start = self.episode_data_index["from"][ep_idx]
-        ep_end = self.episode_data_index["to"][ep_idx]
+        ep = self.meta.episodes[ep_idx]
+        ep_start = ep["dataset_from_index"]
+        ep_end = ep["dataset_to_index"]
         query_indices = {
-            key: [max(ep_start.item(), min(ep_end.item() - 1, idx + delta)) for delta in delta_idx]
+            key: [max(ep_start, min(ep_end - 1, idx + delta)) for delta in delta_idx]
             for key, delta_idx in self.delta_indices.items()
         }
         padding = {  # Pad values outside of current episode range
             f"{key}_is_pad": torch.BoolTensor(
-                [(idx + delta < ep_start.item()) | (idx + delta >= ep_end.item()) for delta in delta_idx]
+                [(idx + delta < ep_start) | (idx + delta >= ep_end) for delta in delta_idx]
             )
             for key, delta_idx in self.delta_indices.items()
         }
@@ -676,7 +737,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         query_timestamps = {}
         for key in self.meta.video_keys:
             if query_indices is not None and key in query_indices:
-                timestamps = self.hf_dataset.select(query_indices[key])["timestamp"]
+                timestamps = self.hf_dataset[query_indices[key]]["timestamp"]
                 query_timestamps[key] = torch.stack(timestamps).tolist()
             else:
                 query_timestamps[key] = [current_ts]
@@ -685,7 +746,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     def _query_hf_dataset(self, query_indices: dict[str, list[int]]) -> dict:
         return {
-            key: torch.stack(self.hf_dataset.select(q_idx)[key])
+            key: torch.stack(self.hf_dataset[q_idx][key])
             for key, q_idx in query_indices.items()
             if key not in self.meta.video_keys
         }
@@ -696,10 +757,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
         Segmentation Fault. This probably happens because a memory reference to the video loader is created in
         the main process and a subprocess fails to access it.
         """
+        ep = self.meta.episodes[ep_idx]
         item = {}
         for vid_key, query_ts in query_timestamps.items():
+            # Episodes are stored sequentially on a single mp4 to reduce the number of files.
+            # Thus we load the start timestamp of the episode on this mp4 and
+            # shift the query timestamp accordingly.
+            from_timestamp = ep[f"videos/{vid_key}/from_timestamp"]
+            shifted_query_ts = [from_timestamp + ts for ts in query_ts]
+
             video_path = self.root / self.meta.get_video_file_path(ep_idx, vid_key)
-            frames = decode_video_frames(video_path, query_ts, self.tolerance_s, self.video_backend)
+            frames = decode_video_frames(video_path, shifted_query_ts, self.tolerance_s, self.video_backend)
             item[vid_key] = frames.squeeze(0)
 
         return item
@@ -737,8 +805,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Add task as a string
         task_idx = item["task_index"].item()
-        item["task"] = self.meta.tasks[task_idx]
-
+        item["task"] = self.meta.tasks.iloc[task_idx].name
         return item
 
     def __repr__(self):
@@ -768,6 +835,9 @@ class LeRobotDataset(torch.utils.data.Dataset):
         )
         return self.root / fpath
 
+    def _get_image_file_dir(self, episode_index: int, image_key: str) -> Path:
+        return self._get_image_file_path(episode_index, image_key, frame_index=0).parent
+
     def _save_image(self, image: torch.Tensor | np.ndarray | PIL.Image.Image, fpath: Path) -> None:
         if self.image_writer is None:
             if isinstance(image, torch.Tensor):
@@ -776,7 +846,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         else:
             self.image_writer.save_image(image=image, fpath=fpath)
 
-    def add_frame(self, frame: dict, task: str, timestamp: float | None = None) -> None:
+    def add_frame(self, frame: dict) -> None:
         """
         This function only adds the frame to the episode_buffer. Apart from images — which are written in a
         temporary directory — nothing is written to disk. To save those frames, the 'save_episode()' method
@@ -794,14 +864,17 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
         # Automatically add frame_index and timestamp to episode buffer
         frame_index = self.episode_buffer["size"]
-        if timestamp is None:
-            timestamp = frame_index / self.fps
+        timestamp = frame.pop("timestamp") if "timestamp" in frame else frame_index / self.fps
         self.episode_buffer["frame_index"].append(frame_index)
         self.episode_buffer["timestamp"].append(timestamp)
-        self.episode_buffer["task"].append(task)
 
         # Add frame features to episode_buffer
         for key in frame:
+            if key == "task":
+                # Note: we associate the task in natural language to its task index during `save_episode`
+                self.episode_buffer["task"].append(frame["task"])
+                continue
+
             if key not in self.features:
                 raise ValueError(
                     f"An element of the frame is not in the features. '{key}' not in '{self.features.keys()}'."
@@ -843,11 +916,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
         episode_buffer["index"] = np.arange(self.meta.total_frames, self.meta.total_frames + episode_length)
         episode_buffer["episode_index"] = np.full((episode_length,), episode_index)
 
-        # Add new tasks to the tasks dictionary
-        for task in episode_tasks:
-            task_index = self.meta.get_task_index(task)
-            if task_index is None:
-                self.meta.add_task(task)
+        # Update tasks and task indices with new tasks if any
+        self.meta.save_episode_tasks(episode_tasks)
 
         # Given tasks in natural language, find their corresponding task indices
         episode_buffer["task_index"] = np.array([self.meta.get_task_index(task) for task in tasks])
@@ -859,51 +929,154 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 continue
             episode_buffer[key] = np.stack(episode_buffer[key])
 
+        # Wait for image writer to end, so that episode stats over images can be computed
         self._wait_image_writer()
-        self._save_episode_table(episode_buffer, episode_index)
         ep_stats = compute_episode_stats(episode_buffer, self.features)
 
-        if len(self.meta.video_keys) > 0:
-            video_paths = self.encode_episode_videos(episode_index)
-            for key in self.meta.video_keys:
-                episode_buffer[key] = video_paths[key]
+        ep_metadata = self._save_episode_data(episode_buffer)
+        for video_key in self.meta.video_keys:
+            ep_metadata.update(self._save_episode_video(video_key, episode_index))
 
-        # `meta.save_episode` be executed after encoding the videos
-        self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats)
+        # `meta.save_episode` need to be executed after encoding the videos
+        self.meta.save_episode(episode_index, episode_length, episode_tasks, ep_stats, ep_metadata)
 
-        ep_data_index = get_episode_data_index(self.meta.episodes, [episode_index])
-        ep_data_index_np = {k: t.numpy() for k, t in ep_data_index.items()}
-        check_timestamps_sync(
-            episode_buffer["timestamp"],
-            episode_buffer["episode_index"],
-            ep_data_index_np,
-            self.fps,
-            self.tolerance_s,
-        )
-
-        video_files = list(self.root.rglob("*.mp4"))
-        assert len(video_files) == self.num_episodes * len(self.meta.video_keys)
-
-        parquet_files = list(self.root.rglob("*.parquet"))
-        assert len(parquet_files) == self.num_episodes
+        # TODO(rcadene): remove? there is only one episode in the episode buffer, no need for ep_data_index
+        # ep_data_index = get_episode_data_index(self.meta.episodes, [episode_index])
+        # ep_data_index_np = {k: t.numpy() for k, t in ep_data_index.items()}
+        # check_timestamps_sync(
+        #     episode_buffer["timestamp"],
+        #     episode_buffer["episode_index"],
+        #     ep_data_index_np,
+        #     self.fps,
+        #     self.tolerance_s,
+        # )
 
+        # TODO(rcadene): images are also deleted in clear_episode_buffer
         # delete images
         img_dir = self.root / "images"
         if img_dir.is_dir():
             shutil.rmtree(self.root / "images")
 
-        if not episode_data:  # Reset the buffer
+        if not episode_data:
+            # Reset episode buffer
             self.episode_buffer = self.create_episode_buffer()
 
-    def _save_episode_table(self, episode_buffer: dict, episode_index: int) -> None:
-        episode_dict = {key: episode_buffer[key] for key in self.hf_features}
-        ep_dataset = datasets.Dataset.from_dict(episode_dict, features=self.hf_features, split="train")
+    def _save_episode_data(self, episode_buffer: dict) -> dict:
+        """Save episode data to a parquet file and update the Hugging Face dataset of frames data.
+
+        This function processes episodes data from a buffer, converts it into a Hugging Face dataset,
+        and saves it as a parquet file. It handles both the creation of new parquet files and the
+        updating of existing ones based on size constraints. After saving the data, it reloads
+        the Hugging Face dataset to ensure it is up-to-date.
+
+        Notes: We both need to update parquet files and HF dataset:
+        - `pandas` loads parquet file in RAM
+        - `datasets` relies on a memory mapping from pyarrow (no RAM). It either converts parquet files to a pyarrow cache on disk,
+          or loads directly from pyarrow cache.
+        """
+        # Convert buffer into HF Dataset
+        ep_dict = {key: episode_buffer[key] for key in self.hf_features}
+        ep_dataset = datasets.Dataset.from_dict(ep_dict, features=self.hf_features, split="train")
         ep_dataset = embed_images(ep_dataset)
-        self.hf_dataset = concatenate_datasets([self.hf_dataset, ep_dataset])
-        self.hf_dataset.set_transform(hf_transform_to_torch)
-        ep_data_path = self.root / self.meta.get_data_file_path(ep_index=episode_index)
-        ep_data_path.parent.mkdir(parents=True, exist_ok=True)
-        ep_dataset.to_parquet(ep_data_path)
+        ep_size_in_mb = get_hf_dataset_size_in_mb(ep_dataset)
+        ep_num_frames = len(ep_dataset)
+        df = pd.DataFrame(ep_dataset)
+
+        if self.meta.episodes is None:
+            # Initialize indices and frame count for a new dataset made of the first episode data
+            chunk_idx, file_idx = 0, 0
+            latest_num_frames = 0
+        else:
+            # Retrieve information from the latest parquet file
+            latest_ep = self.meta.episodes[-1]
+            chunk_idx = latest_ep["data/chunk_index"]
+            file_idx = latest_ep["data/file_index"]
+
+            latest_path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
+            latest_size_in_mb = get_parquet_file_size_in_mb(latest_path)
+            latest_num_frames = get_parquet_num_frames(latest_path)
+
+            # Determine if a new parquet file is needed
+            if latest_size_in_mb + ep_size_in_mb >= self.meta.data_files_size_in_mb:
+                # Size limit is reached, prepare new parquet file
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
+                latest_num_frames = 0
+            else:
+                # Update the existing parquet file with new rows
+                latest_df = pd.read_parquet(latest_path)
+                df = pd.concat([latest_df, df], ignore_index=True)
+
+        # Write the resulting dataframe from RAM to disk
+        path = self.root / self.meta.data_path.format(chunk_index=chunk_idx, file_index=file_idx)
+        path.parent.mkdir(parents=True, exist_ok=True)
+        if len(self.meta.image_keys) > 0:
+            to_parquet_with_hf_images(df, path)
+        else:
+            df.to_parquet(path)
+
+        # Update the Hugging Face dataset by reloading it.
+        # This process should be fast because only the latest Parquet file has been modified.
+        # Therefore, only this file needs to be converted to PyArrow; the rest is loaded from the PyArrow memory-mapped cache.
+        self.hf_dataset = self.load_hf_dataset()
+
+        metadata = {
+            "data/chunk_index": chunk_idx,
+            "data/file_index": file_idx,
+            "dataset_from_index": latest_num_frames,
+            "dataset_to_index": latest_num_frames + ep_num_frames,
+        }
+        return metadata
+
+    def _save_episode_video(self, video_key: str, episode_index: int):
+        # Encode episode frames into a temporary video
+        ep_path = self._encode_temporary_episode_video(video_key, episode_index)
+        ep_size_in_mb = get_video_size_in_mb(ep_path)
+        ep_duration_in_s = get_video_duration_in_s(ep_path)
+
+        if self.meta.episodes is None:
+            # Initialize indices for a new dataset made of the first episode data
+            chunk_idx, file_idx = 0, 0
+            latest_duration_in_s = 0
+            new_path = self.root / self.meta.video_path.format(
+                video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
+            )
+            new_path.parent.mkdir(parents=True, exist_ok=True)
+            shutil.move(str(ep_path), str(new_path))
+        else:
+            # Retrieve information from the latest video file
+            latest_ep = self.meta.episodes[-1]
+            chunk_idx = latest_ep[f"videos/{video_key}/chunk_index"]
+            file_idx = latest_ep[f"videos/{video_key}/file_index"]
+
+            latest_path = self.root / self.meta.video_path.format(
+                video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
+            )
+            latest_size_in_mb = get_video_size_in_mb(latest_path)
+            latest_duration_in_s = get_video_duration_in_s(latest_path)
+
+            if latest_size_in_mb + ep_size_in_mb >= self.meta.video_files_size_in_mb:
+                # Move temporary episode video to a new video file in the dataset
+                chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, self.meta.chunks_size)
+                new_path = self.root / self.meta.video_path.format(
+                    video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
+                )
+                new_path.parent.mkdir(parents=True, exist_ok=True)
+                shutil.move(str(ep_path), str(new_path))
+            else:
+                # Update latest video file
+                concat_video_files([latest_path, ep_path], self.root, video_key, chunk_idx, file_idx)
+
+        # Remove temporary directory
+        shutil.rmtree(str(ep_path.parent))
+
+        metadata = {
+            "episode_index": episode_index,
+            f"videos/{video_key}/chunk_index": chunk_idx,
+            f"videos/{video_key}/file_index": file_idx,
+            f"videos/{video_key}/from_timestamp": latest_duration_in_s,
+            f"videos/{video_key}/to_timestamp": latest_duration_in_s + ep_duration_in_s,
+        }
+        return metadata
 
     def clear_episode_buffer(self) -> None:
         episode_index = self.episode_buffer["episode_index"]
@@ -932,7 +1105,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
     def stop_image_writer(self) -> None:
         """
         Whenever wrapping this dataset inside a parallelized DataLoader, this needs to be called first to
-        remove the image_writer in order for the LeRobotDataset object to be picklable and parallelized.
+        remove the image_writer in order for the LeRobotDataset object to be pickleable and parallelized.
         """
         if self.image_writer is not None:
             self.image_writer.stop()
@@ -943,43 +1116,26 @@ class LeRobotDataset(torch.utils.data.Dataset):
         if self.image_writer is not None:
             self.image_writer.wait_until_done()
 
-    def encode_videos(self) -> None:
+    def _encode_temporary_episode_video(self, video_key: str, episode_index: int) -> dict:
         """
         Use ffmpeg to convert frames stored as png into mp4 videos.
         Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
         since video encoding with ffmpeg is already using multithreading.
         """
-        for ep_idx in range(self.meta.total_episodes):
-            self.encode_episode_videos(ep_idx)
-
-    def encode_episode_videos(self, episode_index: int) -> dict:
-        """
-        Use ffmpeg to convert frames stored as png into mp4 videos.
-        Note: `encode_video_frames` is a blocking call. Making it asynchronous shouldn't speedup encoding,
-        since video encoding with ffmpeg is already using multithreading.
-        """
-        video_paths = {}
-        for key in self.meta.video_keys:
-            video_path = self.root / self.meta.get_video_file_path(episode_index, key)
-            video_paths[key] = str(video_path)
-            if video_path.is_file():
-                # Skip if video is already encoded. Could be the case when resuming data recording.
-                continue
-            img_dir = self._get_image_file_path(
-                episode_index=episode_index, image_key=key, frame_index=0
-            ).parent
-            encode_video_frames(img_dir, video_path, self.fps, overwrite=True)
-
-        return video_paths
+        temp_path = Path(tempfile.mkdtemp(dir=self.root)) / f"{video_key}_{episode_index:03d}.mp4"
+        img_dir = self._get_image_file_dir(episode_index, video_key)
+        encode_video_frames(img_dir, temp_path, self.fps, overwrite=True)
+        return temp_path
 
     @classmethod
     def create(
         cls,
         repo_id: str,
         fps: int,
-        features: dict,
         root: str | Path | None = None,
+        robot: Robot | None = None,
         robot_type: str | None = None,
+        features: dict | None = None,
         use_videos: bool = True,
         tolerance_s: float = 1e-4,
         image_writer_processes: int = 0,
@@ -991,9 +1147,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.meta = LeRobotDatasetMetadata.create(
             repo_id=repo_id,
             fps=fps,
+            root=root,
+            robot=robot,
             robot_type=robot_type,
             features=features,
-            root=root,
             use_videos=use_videos,
         )
         obj.repo_id = obj.meta.repo_id
@@ -1013,7 +1170,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.image_transforms = None
         obj.delta_timestamps = None
         obj.delta_indices = None
-        obj.episode_data_index = None
         obj.video_backend = video_backend if video_backend is not None else get_safe_default_codec()
         return obj
 

lerobot/common/datasets/online_buffer.py

@@ -337,13 +337,11 @@ def compute_sampler_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"],
+            offline_dataset.meta.episodes["dataset_from_index"],
+            offline_dataset.meta.episodes["dataset_to_index"],
             strict=True,
         ):
-            offline_data_mask_indices.extend(
-                range(start_index.item(), end_index.item() - offline_drop_n_last_frames)
-            )
+            offline_data_mask_indices.extend(range(start_index, end_index - 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(

lerobot/common/datasets/sampler.py

@@ -21,7 +21,8 @@ import torch
 class EpisodeAwareSampler:
     def __init__(
         self,
-        episode_data_index: dict,
+        dataset_from_indices: list[int],
+        dataset_to_indices: list[int],
         episode_indices_to_use: Union[list, None] = None,
         drop_n_first_frames: int = 0,
         drop_n_last_frames: int = 0,
@@ -30,7 +31,8 @@ class EpisodeAwareSampler:
         """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.
+            dataset_from_indices: List of indices containing the start of each episode in the dataset.
+            dataset_to_indices: List of indices containing the end of each episode in the dataset.
             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.
@@ -39,12 +41,10 @@ class EpisodeAwareSampler:
         """
         indices = []
         for episode_idx, (start_index, end_index) in enumerate(
-            zip(episode_data_index["from"], episode_data_index["to"], strict=True)
+            zip(dataset_from_indices, dataset_to_indices, 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)
-                )
+                indices.extend(range(start_index + drop_n_first_frames, end_index - drop_n_last_frames))
 
         self.indices = indices
         self.shuffle = shuffle

lerobot/common/datasets/transforms.py

@@ -128,7 +128,7 @@ class SharpnessJitter(Transform):
             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"sharpness values should be between (0., inf), but got {sharpness}.")
+            raise ValueError(f"sharpnesss values should be between (0., inf), but got {sharpness}.")
 
         return float(sharpness[0]), float(sharpness[1])
 

lerobot/common/datasets/utils.py

@@ -17,18 +17,23 @@ import contextlib
 import importlib.resources
 import json
 import logging
+import shutil
+import subprocess
+import tempfile
 from collections.abc import Iterator
-from itertools import accumulate
 from pathlib import Path
 from pprint import pformat
 from types import SimpleNamespace
 from typing import Any
 
 import datasets
-import jsonlines
 import numpy as np
 import packaging.version
+import pandas
+import pandas as pd
+import pyarrow.parquet as pq
 import torch
+from datasets import Dataset, concatenate_datasets
 from datasets.table import embed_table_storage
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi
 from huggingface_hub.errors import RevisionNotFoundError
@@ -40,21 +45,27 @@ from lerobot.common.datasets.backward_compatibility import (
     BackwardCompatibilityError,
     ForwardCompatibilityError,
 )
-from lerobot.common.robots import Robot
+from lerobot.common.robot_devices.robots.utils import Robot
 from lerobot.common.utils.utils import is_valid_numpy_dtype_string
-from lerobot.configs.types import DictLike, FeatureType, PolicyFeature
+from lerobot.configs.types import FeatureType, PolicyFeature
 
-DEFAULT_CHUNK_SIZE = 1000  # Max number of episodes per chunk
+DEFAULT_CHUNK_SIZE = 1000  # Max number of files per chunk
+DEFAULT_DATA_FILE_SIZE_IN_MB = 100  # Max size per file
+DEFAULT_VIDEO_FILE_SIZE_IN_MB = 500  # Max size per file
 
 INFO_PATH = "meta/info.json"
-EPISODES_PATH = "meta/episodes.jsonl"
 STATS_PATH = "meta/stats.json"
-EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
-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"
+EPISODES_DIR = "meta/episodes"
+DATA_DIR = "data"
+VIDEO_DIR = "videos"
+
+CHUNK_FILE_PATTERN = "chunk-{chunk_index:03d}/file-{file_index:03d}"
+DEFAULT_TASKS_PATH = "meta/tasks.parquet"
+DEFAULT_EPISODES_PATH = EPISODES_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
+DEFAULT_DATA_PATH = DATA_DIR + "/" + CHUNK_FILE_PATTERN + ".parquet"
+DEFAULT_VIDEO_PATH = VIDEO_DIR + "/{video_key}/" + CHUNK_FILE_PATTERN + ".mp4"
+DEFAULT_IMAGE_PATH = "images/{image_key}/episode-{episode_index:06d}/frame-{frame_index:06d}.png"
 
 DATASET_CARD_TEMPLATE = """
 ---
@@ -75,6 +86,115 @@ DEFAULT_FEATURES = {
 }
 
 
+def get_parquet_file_size_in_mb(parquet_path):
+    metadata = pq.read_metadata(parquet_path)
+    total_uncompressed_size = 0
+    for row_group in range(metadata.num_row_groups):
+        rg_metadata = metadata.row_group(row_group)
+        for column in range(rg_metadata.num_columns):
+            col_metadata = rg_metadata.column(column)
+            total_uncompressed_size += col_metadata.total_uncompressed_size
+    return total_uncompressed_size / (1024**2)
+
+
+def get_hf_dataset_size_in_mb(hf_ds: Dataset) -> int:
+    return hf_ds.data.nbytes / (1024**2)
+
+
+def get_pd_dataframe_size_in_mb(df: pandas.DataFrame) -> int:
+    # TODO(rcadene): unused?
+    memory_usage_bytes = df.memory_usage(deep=True).sum()
+    return memory_usage_bytes / (1024**2)
+
+
+def update_chunk_file_indices(chunk_idx: int, file_idx: int, chunks_size: int):
+    if file_idx == chunks_size - 1:
+        file_idx = 0
+        chunk_idx += 1
+    else:
+        file_idx += 1
+    return chunk_idx, file_idx
+
+
+def load_nested_dataset(pq_dir: Path) -> Dataset:
+    """Find parquet files in provided directory {pq_dir}/chunk-xxx/file-xxx.parquet
+    Convert parquet files to pyarrow memory mapped in a cache folder for efficient RAM usage
+    Concatenate all pyarrow references to return HF Dataset format
+    """
+    paths = sorted(pq_dir.glob("*/*.parquet"))
+    if len(paths) == 0:
+        raise FileNotFoundError(f"Provided directory does not contain any parquet file: {pq_dir}")
+
+    # TODO(rcadene): set num_proc to accelerate conversion to pyarrow
+    datasets = [Dataset.from_parquet(str(path)) for path in paths]
+    return concatenate_datasets(datasets)
+
+
+def get_parquet_num_frames(parquet_path):
+    metadata = pq.read_metadata(parquet_path)
+    return metadata.num_rows
+
+
+def get_video_size_in_mb(mp4_path: Path):
+    file_size_bytes = mp4_path.stat().st_size
+    file_size_mb = file_size_bytes / (1024**2)
+    return file_size_mb
+
+
+def concat_video_files(paths_to_cat: list[Path], root: Path, video_key: str, chunk_idx: int, file_idx: int):
+    # TODO(rcadene): move to video_utils.py
+    # TODO(rcadene): add docstring
+    tmp_dir = Path(tempfile.mkdtemp(dir=root))
+    # Create a text file with the list of files to concatenate
+    path_concat_video_files = tmp_dir / "concat_video_files.txt"
+    with open(path_concat_video_files, "w") as f:
+        for ep_path in paths_to_cat:
+            f.write(f"file '{str(ep_path)}'\n")
+
+    path_tmp_output = tmp_dir / "tmp_output.mp4"
+    command = [
+        "ffmpeg",
+        "-y",
+        "-f",
+        "concat",
+        "-safe",
+        "0",
+        "-i",
+        str(path_concat_video_files),
+        "-c",
+        "copy",
+        str(path_tmp_output),
+    ]
+    subprocess.run(command, check=True)
+
+    output_path = root / DEFAULT_VIDEO_PATH.format(
+        video_key=video_key, chunk_index=chunk_idx, file_index=file_idx
+    )
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    shutil.move(str(path_tmp_output), str(output_path))
+    shutil.rmtree(str(tmp_dir))
+
+
+def get_video_duration_in_s(mp4_file: Path):
+    # TODO(rcadene): move to video_utils.py
+    command = [
+        "ffprobe",
+        "-v",
+        "error",
+        "-show_entries",
+        "format=duration",
+        "-of",
+        "default=noprint_wrappers=1:nokey=1",
+        str(mp4_file),
+    ]
+    result = subprocess.run(
+        command,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.STDOUT,
+    )
+    return float(result.stdout)
+
+
 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.
 
@@ -107,23 +227,13 @@ def unflatten_dict(d: dict, sep: str = "/") -> dict:
     return outdict
 
 
-def get_nested_item(obj: DictLike, flattened_key: str, sep: str = "/") -> Any:
-    split_keys = flattened_key.split(sep)
-    getter = obj[split_keys[0]]
-    if len(split_keys) == 1:
-        return getter
-
-    for key in split_keys[1:]:
-        getter = getter[key]
-
-    return getter
-
-
 def serialize_dict(stats: dict[str, torch.Tensor | np.ndarray | dict]) -> dict:
     serialized_dict = {}
     for key, value in flatten_dict(stats).items():
         if isinstance(value, (torch.Tensor, np.ndarray)):
             serialized_dict[key] = value.tolist()
+        elif isinstance(value, list) and isinstance(value[0], (int, float, list)):
+            serialized_dict[key] = value
         elif isinstance(value, np.generic):
             serialized_dict[key] = value.item()
         elif isinstance(value, (int, float)):
@@ -153,23 +263,6 @@ def write_json(data: dict, fpath: Path) -> None:
         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 write_info(info: dict, local_dir: Path):
     write_json(info, local_dir / INFO_PATH)
 
@@ -198,43 +291,42 @@ def load_stats(local_dir: Path) -> dict[str, dict[str, np.ndarray]]:
     return cast_stats_to_numpy(stats)
 
 
-def write_task(task_index: int, task: dict, local_dir: Path):
-    task_dict = {
-        "task_index": task_index,
-        "task": task,
-    }
-    append_jsonlines(task_dict, local_dir / TASKS_PATH)
+def write_hf_dataset(hf_dataset: Dataset, local_dir: Path):
+    if get_hf_dataset_size_in_mb(hf_dataset) > DEFAULT_DATA_FILE_SIZE_IN_MB:
+        raise NotImplementedError("Contact a maintainer.")
+
+    path = local_dir / DEFAULT_DATA_PATH.format(chunk_index=0, file_index=0)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    hf_dataset.to_parquet(path)
 
 
-def load_tasks(local_dir: Path) -> tuple[dict, dict]:
-    tasks = load_jsonlines(local_dir / TASKS_PATH)
-    tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
-    task_to_task_index = {task: task_index for task_index, task in tasks.items()}
-    return tasks, task_to_task_index
+def write_tasks(tasks: pandas.DataFrame, local_dir: Path):
+    path = local_dir / DEFAULT_TASKS_PATH
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tasks.to_parquet(path)
 
 
-def write_episode(episode: dict, local_dir: Path):
-    append_jsonlines(episode, local_dir / EPISODES_PATH)
+def load_tasks(local_dir: Path):
+    tasks = pd.read_parquet(local_dir / DEFAULT_TASKS_PATH)
+    return tasks
 
 
-def load_episodes(local_dir: Path) -> dict:
-    episodes = load_jsonlines(local_dir / EPISODES_PATH)
-    return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
+def write_episodes(episodes: Dataset, local_dir: Path):
+    if get_hf_dataset_size_in_mb(episodes) > DEFAULT_DATA_FILE_SIZE_IN_MB:
+        raise NotImplementedError("Contact a maintainer.")
+
+    fpath = local_dir / DEFAULT_EPISODES_PATH.format(chunk_index=0, file_index=0)
+    fpath.parent.mkdir(parents=True, exist_ok=True)
+    episodes.to_parquet(fpath)
 
 
-def write_episode_stats(episode_index: int, episode_stats: dict, local_dir: Path):
-    # We wrap episode_stats in a dictionary since `episode_stats["episode_index"]`
-    # is a dictionary of stats and not an integer.
-    episode_stats = {"episode_index": episode_index, "stats": serialize_dict(episode_stats)}
-    append_jsonlines(episode_stats, local_dir / EPISODES_STATS_PATH)
-
-
-def load_episodes_stats(local_dir: Path) -> dict:
-    episodes_stats = load_jsonlines(local_dir / EPISODES_STATS_PATH)
-    return {
-        item["episode_index"]: cast_stats_to_numpy(item["stats"])
-        for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
-    }
+def load_episodes(local_dir: Path) -> datasets.Dataset:
+    episodes = load_nested_dataset(local_dir / EPISODES_DIR)
+    # Select episode features/columns containing references to episode data and videos
+    # (e.g. tasks, dataset_from_index, dataset_to_index, data/chunk_index, data/file_index, etc.)
+    # This is to speedup access to these data, instead of having to load episode stats.
+    episodes = episodes.select_columns([key for key in episodes.features if not key.startswith("stats/")])
+    return episodes
 
 
 def backward_compatible_episodes_stats(
@@ -387,60 +479,8 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
     return datasets.Features(hf_features)
 
 
-def _validate_feature_names(features: dict[str, dict]) -> None:
-    invalid_features = {name: ft for name, ft in features.items() if "/" in name}
-    if invalid_features:
-        raise ValueError(f"Feature names should not contain '/'. Found '/' in '{invalid_features}'.")
-
-
-def hw_to_dataset_features(
-    hw_features: dict[str, type | tuple], prefix: str, use_video: bool = True
-) -> dict[str, dict]:
-    features = {}
-    joint_fts = {key: ftype for key, ftype in hw_features.items() if ftype is float}
-    cam_fts = {key: shape for key, shape in hw_features.items() if isinstance(shape, tuple)}
-
-    if joint_fts and prefix == "action":
-        features[prefix] = {
-            "dtype": "float32",
-            "shape": (len(joint_fts),),
-            "names": list(joint_fts),
-        }
-
-    if joint_fts and prefix == "observation":
-        features[f"{prefix}.state"] = {
-            "dtype": "float32",
-            "shape": (len(joint_fts),),
-            "names": list(joint_fts),
-        }
-
-    for key, shape in cam_fts.items():
-        features[f"{prefix}.images.{key}"] = {
-            "dtype": "video" if use_video else "image",
-            "shape": shape,
-            "names": ["height", "width", "channels"],
-        }
-
-    _validate_feature_names(features)
-    return features
-
-
-def build_dataset_frame(
-    ds_features: dict[str, dict], values: dict[str, Any], prefix: str
-) -> dict[str, np.ndarray]:
-    frame = {}
-    for key, ft in ds_features.items():
-        if key in DEFAULT_FEATURES or not key.startswith(prefix):
-            continue
-        elif ft["dtype"] == "float32" and len(ft["shape"]) == 1:
-            frame[key] = np.array([values[name] for name in ft["names"]], dtype=np.float32)
-        elif ft["dtype"] in ["image", "video"]:
-            frame[key] = values[key.removeprefix(f"{prefix}.images.")]
-
-    return frame
-
-
 def get_features_from_robot(robot: Robot, use_videos: bool = True) -> dict:
+    # TODO(rcadene): add fps for each feature
     camera_ft = {}
     if robot.cameras:
         camera_ft = {
@@ -468,7 +508,7 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
             type = FeatureType.ENV
         elif key.startswith("observation"):
             type = FeatureType.STATE
-        elif key.startswith("action"):
+        elif key == "action":
             type = FeatureType.ACTION
         else:
             continue
@@ -484,9 +524,9 @@ def dataset_to_policy_features(features: dict[str, dict]) -> dict[str, PolicyFea
 def create_empty_dataset_info(
     codebase_version: str,
     fps: int,
+    robot_type: str,
     features: dict,
     use_videos: bool,
-    robot_type: str | None = None,
 ) -> dict:
     return {
         "codebase_version": codebase_version,
@@ -494,31 +534,17 @@ def create_empty_dataset_info(
         "total_episodes": 0,
         "total_frames": 0,
         "total_tasks": 0,
-        "total_videos": 0,
-        "total_chunks": 0,
         "chunks_size": DEFAULT_CHUNK_SIZE,
+        "data_files_size_in_mb": DEFAULT_DATA_FILE_SIZE_IN_MB,
+        "video_files_size_in_mb": DEFAULT_VIDEO_FILE_SIZE_IN_MB,
         "fps": fps,
         "splits": {},
-        "data_path": DEFAULT_PARQUET_PATH,
+        "data_path": DEFAULT_DATA_PATH,
         "video_path": DEFAULT_VIDEO_PATH if use_videos else None,
         "features": features,
     }
 
 
-def get_episode_data_index(
-    episode_dicts: dict[dict], episodes: list[int] | None = None
-) -> dict[str, torch.Tensor]:
-    episode_lengths = {ep_idx: ep_dict["length"] for ep_idx, ep_dict in episode_dicts.items()}
-    if episodes is not None:
-        episode_lengths = {ep_idx: episode_lengths[ep_idx] for ep_idx in episodes}
-
-    cumulative_lengths = list(accumulate(episode_lengths.values()))
-    return {
-        "from": torch.LongTensor([0] + cumulative_lengths[:-1]),
-        "to": torch.LongTensor(cumulative_lengths),
-    }
-
-
 def check_timestamps_sync(
     timestamps: np.ndarray,
     episode_indices: np.ndarray,
@@ -752,12 +778,16 @@ class IterableNamespace(SimpleNamespace):
 
 
 def validate_frame(frame: dict, features: dict):
-    expected_features = set(features) - set(DEFAULT_FEATURES)
-    actual_features = set(frame)
+    optional_features = {"timestamp"}
+    expected_features = (set(features) - set(DEFAULT_FEATURES.keys())) | {"task"}
+    actual_features = set(frame.keys())
 
-    error_message = validate_features_presence(actual_features, expected_features)
+    error_message = validate_features_presence(actual_features, expected_features, optional_features)
 
-    common_features = actual_features & expected_features
+    if "task" in frame:
+        error_message += validate_feature_string("task", frame["task"])
+
+    common_features = actual_features & (expected_features | optional_features)
     for name in common_features - {"task"}:
         error_message += validate_feature_dtype_and_shape(name, features[name], frame[name])
 
@@ -765,10 +795,12 @@ def validate_frame(frame: dict, features: dict):
         raise ValueError(error_message)
 
 
-def validate_features_presence(actual_features: set[str], expected_features: set[str]):
+def validate_features_presence(
+    actual_features: set[str], expected_features: set[str], optional_features: set[str]
+):
     error_message = ""
     missing_features = expected_features - actual_features
-    extra_features = actual_features - expected_features
+    extra_features = actual_features - (expected_features | optional_features)
 
     if missing_features or extra_features:
         error_message += "Feature mismatch in `frame` dictionary:\n"
@@ -858,3 +890,11 @@ def validate_episode_buffer(episode_buffer: dict, total_episodes: int, features:
             f"In episode_buffer not in features: {buffer_keys - set(features)}"
             f"In features not in episode_buffer: {set(features) - buffer_keys}"
         )
+
+
+def to_parquet_with_hf_images(df: pandas.DataFrame, path: Path):
+    """ This function correctly writes to parquet a panda DataFrame that contains images encoded by HF dataset.
+        This way, it can be loaded by HF dataset and correctly formated images are returned.
+    """
+    # TODO(qlhoest): replace this weird synthax by `df.to_parquet(path)` only
+    datasets.Dataset.from_dict(df.to_dict(orient="list")).to_parquet(path)

lerobot/common/datasets/v2/batch_convert_dataset_v1_to_v2.py

@@ -27,7 +27,7 @@ from textwrap import dedent
 
 from lerobot import available_datasets
 from lerobot.common.datasets.v2.convert_dataset_v1_to_v2 import convert_dataset
-from lerobot.common.robots.aloha.configuration_aloha import AlohaRobotConfig
+from lerobot.common.robot_devices.robots.configs import AlohaRobotConfig
 
 LOCAL_DIR = Path("data/")
 

lerobot/common/datasets/v2/convert_dataset_v1_to_v2.py

@@ -121,12 +121,12 @@ from safetensors.torch import load_file
 
 from lerobot.common.datasets.utils import (
     DEFAULT_CHUNK_SIZE,
-    DEFAULT_PARQUET_PATH,
+    DEFAULT_DATA_PATH,
     DEFAULT_VIDEO_PATH,
-    EPISODES_PATH,
     INFO_PATH,
+    LEGACY_EPISODES_PATH,
+    LEGACY_TASKS_PATH,
     STATS_PATH,
-    TASKS_PATH,
     create_branch,
     create_lerobot_dataset_card,
     flatten_dict,
@@ -141,7 +141,8 @@ from lerobot.common.datasets.video_utils import (
     get_image_pixel_channels,
     get_video_info,
 )
-from lerobot.common.robots import RobotConfig
+from lerobot.common.robot_devices.robots.configs import RobotConfig
+from lerobot.common.robot_devices.robots.utils import make_robot_config
 
 V16 = "v1.6"
 V20 = "v2.0"
@@ -290,14 +291,12 @@ def split_parquet_by_episodes(
     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)
+        chunk_dir = "/".join(DEFAULT_DATA_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
-            )
+            output_file = output_dir / DEFAULT_DATA_PATH.format(episode_chunk=ep_chunk, episode_index=ep_idx)
             pq.write_table(ep_table, output_file)
 
     return episode_lengths
@@ -495,7 +494,7 @@ def convert_dataset(
 
     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)
+    write_jsonlines(tasks, v20_dir / LEGACY_TASKS_PATH)
     features["task_index"] = {
         "dtype": "int64",
         "shape": (1,),
@@ -545,7 +544,7 @@ def convert_dataset(
         {"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)
+    write_jsonlines(episodes, v20_dir / LEGACY_EPISODES_PATH)
 
     # Assemble metadata v2.0
     metadata_v2_0 = {
@@ -559,7 +558,7 @@ def convert_dataset(
         "chunks_size": DEFAULT_CHUNK_SIZE,
         "fps": metadata_v1["fps"],
         "splits": {"train": f"0:{total_episodes}"},
-        "data_path": DEFAULT_PARQUET_PATH,
+        "data_path": DEFAULT_DATA_PATH,
         "video_path": DEFAULT_VIDEO_PATH if video_keys else None,
         "features": features,
     }
@@ -597,30 +596,6 @@ def convert_dataset(
         create_branch(repo_id=repo_id, branch=V20, repo_type="dataset")
 
 
-def make_robot_config(robot_type: str, **kwargs) -> RobotConfig:
-    if robot_type == "aloha":
-        raise NotImplementedError  # TODO
-
-    elif robot_type == "koch_follower":
-        from lerobot.common.robots.koch_follower import KochFollowerConfig
-
-        return KochFollowerConfig(**kwargs)
-    elif robot_type == "so100_follower":
-        from lerobot.common.robots.so100_follower import SO100FollowerConfig
-
-        return SO100FollowerConfig(**kwargs)
-    elif robot_type == "stretch":
-        from lerobot.common.robots.stretch3 import Stretch3RobotConfig
-
-        return Stretch3RobotConfig(**kwargs)
-    elif robot_type == "lekiwi":
-        from lerobot.common.robots.lekiwi import LeKiwiConfig
-
-        return LeKiwiConfig(**kwargs)
-    else:
-        raise ValueError(f"Robot type '{robot_type}' is not available.")
-
-
 def main():
     parser = argparse.ArgumentParser()
     task_args = parser.add_mutually_exclusive_group(required=True)

lerobot/common/datasets/v21/convert_dataset_v20_to_v21.py

@@ -37,7 +37,7 @@ import logging
 from huggingface_hub import HfApi
 
 from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
-from lerobot.common.datasets.utils import EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
+from lerobot.common.datasets.utils import LEGACY_EPISODES_STATS_PATH, STATS_PATH, load_stats, write_info
 from lerobot.common.datasets.v21.convert_stats import check_aggregate_stats, convert_stats
 
 V20 = "v2.0"
@@ -61,8 +61,8 @@ def convert_dataset(
     with SuppressWarnings():
         dataset = LeRobotDataset(repo_id, revision=V20, force_cache_sync=True)
 
-    if (dataset.root / EPISODES_STATS_PATH).is_file():
-        (dataset.root / EPISODES_STATS_PATH).unlink()
+    if (dataset.root / LEGACY_EPISODES_STATS_PATH).is_file():
+        (dataset.root / LEGACY_EPISODES_STATS_PATH).unlink()
 
     convert_stats(dataset, num_workers=num_workers)
     ref_stats = load_stats(dataset.root)

lerobot/common/datasets/v21/convert_stats.py

@@ -19,7 +19,7 @@ from tqdm import tqdm
 
 from lerobot.common.datasets.compute_stats import aggregate_stats, get_feature_stats, sample_indices
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
-from lerobot.common.datasets.utils import write_episode_stats
+from lerobot.common.datasets.utils import legacy_write_episode_stats
 
 
 def sample_episode_video_frames(dataset: LeRobotDataset, episode_index: int, ft_key: str) -> np.ndarray:
@@ -72,7 +72,7 @@ def convert_stats(dataset: LeRobotDataset, num_workers: int = 0):
             convert_episode_stats(dataset, ep_idx)
 
     for ep_idx in tqdm(range(total_episodes)):
-        write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
+        legacy_write_episode_stats(ep_idx, dataset.meta.episodes_stats[ep_idx], dataset.root)
 
 
 def check_aggregate_stats(

lerobot/common/datasets/v30/convert_dataset_v21_to_v30.py

@@ -0,0 +1,452 @@
+"""
+This script will help you convert any LeRobot dataset already pushed to the hub from codebase version 2.1 to
+3.0. It will:
+
+- Generate per-episodes stats and writes them in `episodes_stats.jsonl`
+- Check consistency between these new stats and the old ones.
+- Remove the deprecated `stats.json`.
+- Update codebase_version in `info.json`.
+- Push this new version to the hub on the 'main' branch and tags it with "v2.1".
+
+Usage:
+
+```bash
+python lerobot/common/datasets/v30/convert_dataset_v21_to_v30.py \
+    --repo-id=lerobot/pusht
+```
+
+"""
+
+import argparse
+import shutil
+from pathlib import Path
+from typing import Any
+
+import jsonlines
+import pandas as pd
+import pyarrow as pa
+import tqdm
+from datasets import Dataset, Features, Image
+from huggingface_hub import HfApi, snapshot_download
+from requests import HTTPError
+
+from lerobot.common.constants import HF_LEROBOT_HOME
+from lerobot.common.datasets.compute_stats import aggregate_stats
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
+from lerobot.common.datasets.utils import (
+    DEFAULT_CHUNK_SIZE,
+    DEFAULT_DATA_FILE_SIZE_IN_MB,
+    DEFAULT_DATA_PATH,
+    DEFAULT_VIDEO_FILE_SIZE_IN_MB,
+    DEFAULT_VIDEO_PATH,
+    cast_stats_to_numpy,
+    concat_video_files,
+    flatten_dict,
+    get_parquet_file_size_in_mb,
+    get_parquet_num_frames,
+    get_video_duration_in_s,
+    get_video_size_in_mb,
+    load_info,
+    update_chunk_file_indices,
+    write_episodes,
+    write_info,
+    write_stats,
+    write_tasks,
+)
+
+LEGACY_EPISODES_PATH = "meta/episodes.jsonl"
+LEGACY_EPISODES_STATS_PATH = "meta/episodes_stats.jsonl"
+LEGACY_TASKS_PATH = "meta/tasks.jsonl"
+LEGACY_DEFAULT_VIDEO_PATH = "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4"
+LEGACY_DEFAULT_PARQUET_PATH = "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet"
+
+V21 = "v2.1"
+
+
+"""
+-------------------------
+OLD
+data/chunk-000/episode_000000.parquet
+
+NEW
+data/chunk-000/file_000.parquet
+-------------------------
+OLD
+videos/chunk-000/CAMERA/episode_000000.mp4
+
+NEW
+videos/chunk-000/file_000.mp4
+-------------------------
+OLD
+episodes.jsonl
+{"episode_index": 1, "tasks": ["Put the blue block in the green bowl"], "length": 266}
+
+NEW
+meta/episodes/chunk-000/episodes_000.parquet
+episode_index | video_chunk_index | video_file_index | data_chunk_index | data_file_index | tasks | length
+-------------------------
+OLD
+tasks.jsonl
+{"task_index": 1, "task": "Put the blue block in the green bowl"}
+
+NEW
+meta/tasks/chunk-000/file_000.parquet
+task_index | task
+-------------------------
+OLD
+episodes_stats.jsonl
+
+NEW
+meta/episodes_stats/chunk-000/file_000.parquet
+episode_index | mean | std | min | max
+-------------------------
+UPDATE
+meta/info.json
+-------------------------
+"""
+
+
+def load_jsonlines(fpath: Path) -> list[Any]:
+    with jsonlines.open(fpath, "r") as reader:
+        return list(reader)
+
+
+def legacy_load_episodes(local_dir: Path) -> dict:
+    episodes = load_jsonlines(local_dir / LEGACY_EPISODES_PATH)
+    return {item["episode_index"]: item for item in sorted(episodes, key=lambda x: x["episode_index"])}
+
+
+def legacy_load_episodes_stats(local_dir: Path) -> dict:
+    episodes_stats = load_jsonlines(local_dir / LEGACY_EPISODES_STATS_PATH)
+    return {
+        item["episode_index"]: cast_stats_to_numpy(item["stats"])
+        for item in sorted(episodes_stats, key=lambda x: x["episode_index"])
+    }
+
+
+def legacy_load_tasks(local_dir: Path) -> tuple[dict, dict]:
+    tasks = load_jsonlines(local_dir / LEGACY_TASKS_PATH)
+    tasks = {item["task_index"]: item["task"] for item in sorted(tasks, key=lambda x: x["task_index"])}
+    task_to_task_index = {task: task_index for task_index, task in tasks.items()}
+    return tasks, task_to_task_index
+
+
+def convert_tasks(root, new_root):
+    tasks, _ = legacy_load_tasks(root)
+    task_indices = tasks.keys()
+    task_strings = tasks.values()
+    df_tasks = pd.DataFrame({"task_index": task_indices}, index=task_strings)
+    write_tasks(df_tasks, new_root)
+
+
+def concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys):
+    # TODO(rcadene): to save RAM use Dataset.from_parquet(file) and concatenate_datasets
+    dataframes = [pd.read_parquet(file) for file in paths_to_cat]
+    # Concatenate all DataFrames along rows
+    concatenated_df = pd.concat(dataframes, ignore_index=True)
+
+    path = new_root / DEFAULT_DATA_PATH.format(chunk_index=chunk_idx, file_index=file_idx)
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    if len(image_keys) > 0:
+        schema = pa.Schema.from_pandas(concatenated_df)
+        features = Features.from_arrow_schema(schema)
+        for key in image_keys:
+            features[key] = Image()
+        schema = features.arrow_schema
+    else:
+        schema = None
+
+    concatenated_df.to_parquet(path, index=False, schema=schema)
+
+
+def convert_data(root, new_root):
+    data_dir = root / "data"
+    ep_paths = sorted(data_dir.glob("*/*.parquet"))
+
+    image_keys = get_image_keys(root)
+
+    ep_idx = 0
+    chunk_idx = 0
+    file_idx = 0
+    size_in_mb = 0
+    num_frames = 0
+    paths_to_cat = []
+    episodes_metadata = []
+    for ep_path in ep_paths:
+        ep_size_in_mb = get_parquet_file_size_in_mb(ep_path)
+        ep_num_frames = get_parquet_num_frames(ep_path)
+        ep_metadata = {
+            "episode_index": ep_idx,
+            "data/chunk_index": chunk_idx,
+            "data/file_index": file_idx,
+            "dataset_from_index": num_frames,
+            "dataset_to_index": num_frames + ep_num_frames,
+        }
+        size_in_mb += ep_size_in_mb
+        num_frames += ep_num_frames
+        episodes_metadata.append(ep_metadata)
+        ep_idx += 1
+
+        if size_in_mb < DEFAULT_DATA_FILE_SIZE_IN_MB:
+            paths_to_cat.append(ep_path)
+            continue
+
+        concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
+
+        # Reset for the next file
+        size_in_mb = ep_size_in_mb
+        num_frames = ep_num_frames
+        paths_to_cat = [ep_path]
+
+        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
+
+    # Write remaining data if any
+    if paths_to_cat:
+        concat_data_files(paths_to_cat, new_root, chunk_idx, file_idx, image_keys)
+
+    return episodes_metadata
+
+
+def get_video_keys(root):
+    info = load_info(root)
+    features = info["features"]
+    video_keys = [key for key, ft in features.items() if ft["dtype"] == "video"]
+    return video_keys
+
+
+def get_image_keys(root):
+    info = load_info(root)
+    features = info["features"]
+    image_keys = [key for key, ft in features.items() if ft["dtype"] == "image"]
+    return image_keys
+
+
+def convert_videos(root: Path, new_root: Path):
+    video_keys = get_video_keys(root)
+    if len(video_keys) == 0:
+        return None
+
+    video_keys = sorted(video_keys)
+
+    eps_metadata_per_cam = []
+    for camera in video_keys:
+        eps_metadata = convert_videos_of_camera(root, new_root, camera)
+        eps_metadata_per_cam.append(eps_metadata)
+
+    num_eps_per_cam = [len(eps_cam_map) for eps_cam_map in eps_metadata_per_cam]
+    if len(set(num_eps_per_cam)) != 1:
+        raise ValueError(f"All cams dont have same number of episodes ({num_eps_per_cam}).")
+
+    episods_metadata = []
+    num_cameras = len(video_keys)
+    num_episodes = num_eps_per_cam[0]
+    for ep_idx in range(num_episodes):
+        # Sanity check
+        ep_ids = [eps_metadata_per_cam[cam_idx][ep_idx]["episode_index"] for cam_idx in range(num_cameras)]
+        ep_ids += [ep_idx]
+        if len(set(ep_ids)) != 1:
+            raise ValueError(f"All episode indices need to match ({ep_ids}).")
+
+        ep_dict = {}
+        for cam_idx in range(num_cameras):
+            ep_dict.update(eps_metadata_per_cam[cam_idx][ep_idx])
+        episods_metadata.append(ep_dict)
+
+    return episods_metadata
+
+
+def convert_videos_of_camera(root: Path, new_root: Path, video_key):
+    # Access old paths to mp4
+    videos_dir = root / "videos"
+    ep_paths = sorted(videos_dir.glob(f"*/{video_key}/*.mp4"))
+
+    ep_idx = 0
+    chunk_idx = 0
+    file_idx = 0
+    size_in_mb = 0
+    duration_in_s = 0.0
+    paths_to_cat = []
+    episodes_metadata = []
+    for ep_path in tqdm.tqdm(ep_paths, desc=f"convert videos of {video_key}"):
+        ep_size_in_mb = get_video_size_in_mb(ep_path)
+        ep_duration_in_s = get_video_duration_in_s(ep_path)
+        ep_metadata = {
+            "episode_index": ep_idx,
+            f"videos/{video_key}/chunk_index": chunk_idx,
+            f"videos/{video_key}/file_index": file_idx,
+            f"videos/{video_key}/from_timestamp": duration_in_s,
+            f"videos/{video_key}/to_timestamp": duration_in_s + ep_duration_in_s,
+        }
+        size_in_mb += ep_size_in_mb
+        duration_in_s += ep_duration_in_s
+        episodes_metadata.append(ep_metadata)
+        ep_idx += 1
+
+        if size_in_mb < DEFAULT_VIDEO_FILE_SIZE_IN_MB:
+            paths_to_cat.append(ep_path)
+            continue
+
+        concat_video_files(paths_to_cat, new_root, video_key, chunk_idx, file_idx)
+
+        # Reset for the next file
+        size_in_mb = ep_size_in_mb
+        duration_in_s = ep_duration_in_s
+        paths_to_cat = [ep_path]
+
+        chunk_idx, file_idx = update_chunk_file_indices(chunk_idx, file_idx, DEFAULT_CHUNK_SIZE)
+
+    # Write remaining videos if any
+    if paths_to_cat:
+        concat_video_files(paths_to_cat, new_root, video_key, chunk_idx, file_idx)
+
+    return episodes_metadata
+
+
+def generate_episode_metadata_dict(
+    episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_videos=None
+):
+    num_episodes = len(episodes_metadata)
+    episodes_legacy_metadata_vals = list(episodes_legacy_metadata.values())
+    episodes_stats_vals = list(episodes_stats.values())
+    episodes_stats_keys = list(episodes_stats.keys())
+
+    for i in range(num_episodes):
+        ep_legacy_metadata = episodes_legacy_metadata_vals[i]
+        ep_metadata = episodes_metadata[i]
+        ep_stats = episodes_stats_vals[i]
+
+        ep_ids_set = {
+            ep_legacy_metadata["episode_index"],
+            ep_metadata["episode_index"],
+            episodes_stats_keys[i],
+        }
+
+        if episodes_videos is None:
+            ep_video = {}
+        else:
+            ep_video = episodes_videos[i]
+            ep_ids_set.add(ep_video["episode_index"])
+
+        if len(ep_ids_set) != 1:
+            raise ValueError(f"Number of episodes is not the same ({ep_ids_set}).")
+
+        ep_dict = {**ep_metadata, **ep_video, **ep_legacy_metadata, **flatten_dict({"stats": ep_stats})}
+        ep_dict["meta/episodes/chunk_index"] = 0
+        ep_dict["meta/episodes/file_index"] = 0
+        yield ep_dict
+
+
+def convert_episodes_metadata(root, new_root, episodes_metadata, episodes_video_metadata=None):
+    episodes_legacy_metadata = legacy_load_episodes(root)
+    episodes_stats = legacy_load_episodes_stats(root)
+
+    num_eps_set = {len(episodes_legacy_metadata), len(episodes_metadata)}
+    if episodes_video_metadata is not None:
+        num_eps_set.add(len(episodes_video_metadata))
+
+    if len(num_eps_set) != 1:
+        raise ValueError(f"Number of episodes is not the same ({num_eps_set}).")
+
+    ds_episodes = Dataset.from_generator(
+        lambda: generate_episode_metadata_dict(
+            episodes_legacy_metadata, episodes_metadata, episodes_stats, episodes_video_metadata
+        )
+    )
+    write_episodes(ds_episodes, new_root)
+
+    stats = aggregate_stats(list(episodes_stats.values()))
+    write_stats(stats, new_root)
+
+
+def convert_info(root, new_root):
+    info = load_info(root)
+    info["codebase_version"] = "v3.0"
+    del info["total_chunks"]
+    del info["total_videos"]
+    info["data_files_size_in_mb"] = DEFAULT_DATA_FILE_SIZE_IN_MB
+    info["video_files_size_in_mb"] = DEFAULT_VIDEO_FILE_SIZE_IN_MB
+    info["data_path"] = DEFAULT_DATA_PATH
+    info["video_path"] = DEFAULT_VIDEO_PATH
+    info["fps"] = float(info["fps"])
+    for key in info["features"]:
+        if info["features"][key]["dtype"] == "video":
+            # already has fps in video_info
+            continue
+        info["features"][key]["fps"] = info["fps"]
+    write_info(info, new_root)
+
+
+def convert_dataset(
+    repo_id: str,
+    branch: str | None = None,
+    num_workers: int = 4,
+):
+    root = HF_LEROBOT_HOME / repo_id
+    old_root = HF_LEROBOT_HOME / f"{repo_id}_old"
+    new_root = HF_LEROBOT_HOME / f"{repo_id}_v30"
+
+    if old_root.is_dir() and root.is_dir():
+        shutil.rmtree(str(root))
+        shutil.move(str(old_root), str(root))
+
+    if new_root.is_dir():
+        shutil.rmtree(new_root)
+
+    snapshot_download(
+        repo_id,
+        repo_type="dataset",
+        revision=V21,
+        local_dir=root,
+    )
+
+    convert_info(root, new_root)
+    convert_tasks(root, new_root)
+    episodes_metadata = convert_data(root, new_root)
+    episodes_videos_metadata = convert_videos(root, new_root)
+    convert_episodes_metadata(root, new_root, episodes_metadata, episodes_videos_metadata)
+
+    shutil.move(str(root), str(old_root))
+    shutil.move(str(new_root), str(root))
+
+    hub_api = HfApi()
+    try:
+        hub_api.delete_tag(repo_id, tag=CODEBASE_VERSION, repo_type="dataset")
+    except HTTPError as e:
+        print(f"tag={CODEBASE_VERSION} probably doesn't exist. Skipping exception ({e})")
+        pass
+    hub_api.delete_files(
+        delete_patterns=["data/chunk*/episode_*", "meta/*.jsonl", "videos/chunk*"],
+        repo_id=repo_id,
+        revision=branch,
+        repo_type="dataset",
+    )
+    hub_api.create_tag(repo_id, tag=CODEBASE_VERSION, revision=branch, repo_type="dataset")
+
+    LeRobotDataset(repo_id).push_to_hub()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    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`).",
+    )
+    parser.add_argument(
+        "--branch",
+        type=str,
+        default=None,
+        help="Repo branch to push your dataset. Defaults to the main branch.",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=4,
+        help="Number of workers for parallelizing stats compute. Defaults to 4.",
+    )
+
+    args = parser.parse_args()
+    convert_dataset(**vars(args))

lerobot/common/datasets/video_utils.py

@@ -13,15 +13,16 @@
 # 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 glob
 import importlib
+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 av
 import pyarrow as pa
 import torch
 import torchvision
@@ -101,7 +102,7 @@ def decode_video_frames_torchvision(
     keyframes_only = False
     torchvision.set_video_backend(backend)
     if backend == "pyav":
-        keyframes_only = True  # pyav doesn't support accurate seek
+        keyframes_only = True  # pyav doesnt support accuracte seek
 
     # set a video stream reader
     # TODO(rcadene): also load audio stream at the same time
@@ -154,6 +155,7 @@ def decode_video_frames_torchvision(
     )
 
     # get closest frames to the query timestamps
+    # TODO(rcadene): remove torch.stack
     closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
     closest_ts = loaded_ts[argmin_]
 
@@ -251,83 +253,51 @@ def encode_video_frames(
     g: int | None = 2,
     crf: int | None = 30,
     fast_decode: int = 0,
-    log_level: int | None = av.logging.ERROR,
+    log_level: str | None = "quiet",
     overwrite: bool = False,
 ) -> None:
     """More info on ffmpeg arguments tuning on `benchmark/video/README.md`"""
-    # Check encoder availability
-    if vcodec not in ["h264", "hevc", "libsvtav1"]:
-        raise ValueError(f"Unsupported video codec: {vcodec}. Supported codecs are: h264, hevc, libsvtav1.")
-
     video_path = Path(video_path)
     imgs_dir = Path(imgs_dir)
+    video_path.parent.mkdir(parents=True, exist_ok=True)
 
-    video_path.parent.mkdir(parents=True, exist_ok=overwrite)
-
-    # Encoders/pixel formats incompatibility check
-    if (vcodec == "libsvtav1" or vcodec == "hevc") and pix_fmt == "yuv444p":
-        logging.warning(
-            f"Incompatible pixel format 'yuv444p' for codec {vcodec}, auto-selecting format 'yuv420p'"
-        )
-        pix_fmt = "yuv420p"
-
-    # Get input frames
-    template = "frame_" + ("[0-9]" * 6) + ".png"
-    input_list = sorted(
-        glob.glob(str(imgs_dir / template)), key=lambda x: int(x.split("_")[-1].split(".")[0])
+    ffmpeg_args = OrderedDict(
+        [
+            ("-f", "image2"),
+            ("-r", str(fps)),
+            ("-i", str(imgs_dir / "frame-%06d.png")),
+            ("-vcodec", vcodec),
+            ("-pix_fmt", pix_fmt),
+        ]
     )
 
-    # Define video output frame size (assuming all input frames are the same size)
-    if len(input_list) == 0:
-        raise FileNotFoundError(f"No images found in {imgs_dir}.")
-    dummy_image = Image.open(input_list[0])
-    width, height = dummy_image.size
-
-    # Define video codec options
-    video_options = {}
-
     if g is not None:
-        video_options["g"] = str(g)
+        ffmpeg_args["-g"] = str(g)
 
     if crf is not None:
-        video_options["crf"] = str(crf)
+        ffmpeg_args["-crf"] = str(crf)
 
     if fast_decode:
-        key = "svtav1-params" if vcodec == "libsvtav1" else "tune"
+        key = "-svtav1-params" if vcodec == "libsvtav1" else "-tune"
         value = f"fast-decode={fast_decode}" if vcodec == "libsvtav1" else "fastdecode"
-        video_options[key] = value
+        ffmpeg_args[key] = value
 
-    # Set logging level
     if log_level is not None:
-        # "While less efficient, it is generally preferable to modify logging with Python’s logging"
-        logging.getLogger("libav").setLevel(log_level)
+        ffmpeg_args["-loglevel"] = str(log_level)
 
-    # Create and open output file (overwrite by default)
-    with av.open(str(video_path), "w") as output:
-        output_stream = output.add_stream(vcodec, fps, options=video_options)
-        output_stream.pix_fmt = pix_fmt
-        output_stream.width = width
-        output_stream.height = height
+    ffmpeg_args = [item for pair in ffmpeg_args.items() for item in pair]
+    if overwrite:
+        ffmpeg_args.append("-y")
 
-        # Loop through input frames and encode them
-        for input_data in input_list:
-            input_image = Image.open(input_data).convert("RGB")
-            input_frame = av.VideoFrame.from_image(input_image)
-            packet = output_stream.encode(input_frame)
-            if packet:
-                output.mux(packet)
-
-        # Flush the encoder
-        packet = output_stream.encode()
-        if packet:
-            output.mux(packet)
-
-    # Reset logging level
-    if log_level is not None:
-        av.logging.restore_default_callback()
+    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}.")
+        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
@@ -363,68 +333,78 @@ with warnings.catch_warnings():
 
 
 def get_audio_info(video_path: Path | str) -> dict:
-    # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.ERROR)
+    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}")
 
-    # Getting audio stream information
-    audio_info = {}
-    with av.open(str(video_path), "r") as audio_file:
-        try:
-            audio_stream = audio_file.streams.audio[0]
-        except IndexError:
-            # Reset logging level
-            av.logging.restore_default_callback()
-            return {"has_audio": False}
+    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}
 
-        audio_info["audio.channels"] = audio_stream.channels
-        audio_info["audio.codec"] = audio_stream.codec.canonical_name
-        # In an ideal loseless case : bit depth x sample rate x channels = bit rate.
-        # In an actual compressed case, the bit rate is set according to the compression level : the lower the bit rate, the more compression is applied.
-        audio_info["audio.bit_rate"] = audio_stream.bit_rate
-        audio_info["audio.sample_rate"] = audio_stream.sample_rate  # Number of samples per second
-        # In an ideal loseless case : fixed number of bits per sample.
-        # In an actual compressed case : variable number of bits per sample (often reduced to match a given depth rate).
-        audio_info["audio.bit_depth"] = audio_stream.format.bits
-        audio_info["audio.channel_layout"] = audio_stream.layout.name
-        audio_info["has_audio"] = True
-
-    # Reset logging level
-    av.logging.restore_default_callback()
-
-    return audio_info
+    # 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:
-    # Set logging level
-    logging.getLogger("libav").setLevel(av.logging.ERROR)
+    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}")
 
-    # Getting video stream information
-    video_info = {}
-    with av.open(str(video_path), "r") as video_file:
-        try:
-            video_stream = video_file.streams.video[0]
-        except IndexError:
-            # Reset logging level
-            av.logging.restore_default_callback()
-            return {}
+    info = json.loads(result.stdout)
+    video_stream_info = info["streams"][0]
 
-        video_info["video.height"] = video_stream.height
-        video_info["video.width"] = video_stream.width
-        video_info["video.codec"] = video_stream.codec.canonical_name
-        video_info["video.pix_fmt"] = video_stream.pix_fmt
-        video_info["video.is_depth_map"] = False
+    # 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
 
-        # Calculate fps from r_frame_rate
-        video_info["video.fps"] = int(video_stream.base_rate)
+    pixel_channels = get_video_pixel_channels(video_stream_info["pix_fmt"])
 
-        pixel_channels = get_video_pixel_channels(video_stream.pix_fmt)
-        video_info["video.channels"] = pixel_channels
-
-    # Reset logging level
-    av.logging.restore_default_callback()
-
-    # Adding audio stream information
-    video_info.update(**get_audio_info(video_path))
+    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
 

lerobot/common/envs/configs.py

@@ -14,13 +14,10 @@
 
 import abc
 from dataclasses import dataclass, field
-from typing import Any, Dict, Optional, Tuple
 
 import draccus
 
-from lerobot.common.constants import ACTION, OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE
-from lerobot.common.robots import RobotConfig
-from lerobot.common.teleoperators.config import TeleoperatorConfig
+from lerobot.common.constants import ACTION, OBS_ENV, OBS_IMAGE, OBS_IMAGES, OBS_ROBOT
 from lerobot.configs.types import FeatureType, PolicyFeature
 
 
@@ -35,8 +32,7 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
     def type(self) -> str:
         return self.get_choice_name(self.__class__)
 
-    @property
-    @abc.abstractmethod
+    @abc.abstractproperty
     def gym_kwargs(self) -> dict:
         raise NotImplementedError()
 
@@ -57,7 +53,7 @@ class AlohaEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             "action": ACTION,
-            "agent_pos": OBS_STATE,
+            "agent_pos": OBS_ROBOT,
             "top": f"{OBS_IMAGE}.top",
             "pixels/top": f"{OBS_IMAGES}.top",
         }
@@ -98,8 +94,8 @@ class PushtEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             "action": ACTION,
-            "agent_pos": OBS_STATE,
-            "environment_state": OBS_ENV_STATE,
+            "agent_pos": OBS_ROBOT,
+            "environment_state": OBS_ENV,
             "pixels": OBS_IMAGE,
         }
     )
@@ -140,7 +136,7 @@ class XarmEnv(EnvConfig):
     features_map: dict[str, str] = field(
         default_factory=lambda: {
             "action": ACTION,
-            "agent_pos": OBS_STATE,
+            "agent_pos": OBS_ROBOT,
             "pixels": OBS_IMAGE,
         }
     )
@@ -158,125 +154,3 @@ class XarmEnv(EnvConfig):
             "visualization_height": self.visualization_height,
             "max_episode_steps": self.episode_length,
         }
-
-
-@dataclass
-class VideoRecordConfig:
-    """Configuration for video recording in ManiSkill environments."""
-
-    enabled: bool = False
-    record_dir: str = "videos"
-    trajectory_name: str = "trajectory"
-
-
-# @dataclass
-# class EEActionSpaceConfig:
-#     """Configuration parameters for end-effector action space."""
-
-#     x_step_size: float
-#     y_step_size: float
-#     z_step_size: float
-#     bounds: Dict[str, Any]  # Contains 'min' and 'max' keys with position bounds
-#     control_mode: str = "gamepad"
-
-
-@dataclass
-class EnvTransformConfig:
-    """Configuration for environment wrappers."""
-
-    # ee_action_space_params: EEActionSpaceConfig = field(default_factory=EEActionSpaceConfig)
-    control_mode: str = "gamepad"
-    display_cameras: bool = False
-    add_joint_velocity_to_observation: bool = False
-    add_current_to_observation: bool = False
-    add_ee_pose_to_observation: bool = False
-    crop_params_dict: Optional[Dict[str, Tuple[int, int, int, int]]] = None
-    resize_size: Optional[Tuple[int, int]] = None
-    control_time_s: float = 20.0
-    fixed_reset_joint_positions: Optional[Any] = None
-    reset_time_s: float = 5.0
-    use_gripper: bool = False
-    gripper_quantization_threshold: float | None = 0.8
-    gripper_penalty: float = 0.0
-    gripper_penalty_in_reward: bool = False
-
-
-@EnvConfig.register_subclass(name="gym_manipulator")
-@dataclass
-class HILSerlRobotEnvConfig(EnvConfig):
-    """Configuration for the HILSerlRobotEnv environment."""
-
-    robot: Optional[RobotConfig] = None
-    teleop: Optional[TeleoperatorConfig] = None
-    wrapper: Optional[EnvTransformConfig] = None
-    fps: int = 10
-    name: str = "real_robot"
-    mode: str = None  # Either "record", "replay", None
-    repo_id: Optional[str] = None
-    dataset_root: Optional[str] = None
-    task: str = ""
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: Optional[str] = None
-    reward_classifier_pretrained_path: Optional[str] = None
-    # For the reward classifier, to record more positive examples after a success
-    number_of_steps_after_success: int = 0
-
-    def gym_kwargs(self) -> dict:
-        return {}
-
-
-@EnvConfig.register_subclass("hil")
-@dataclass
-class HILEnvConfig(EnvConfig):
-    """Configuration for the HIL environment."""
-
-    type: str = "hil"
-    name: str = "PandaPickCube"
-    task: str = "PandaPickCubeKeyboard-v0"
-    use_viewer: bool = True
-    gripper_penalty: float = 0.0
-    use_gamepad: bool = True
-    state_dim: int = 18
-    action_dim: int = 4
-    fps: int = 100
-    episode_length: int = 100
-    video_record: VideoRecordConfig = field(default_factory=VideoRecordConfig)
-    features: dict[str, PolicyFeature] = field(
-        default_factory=lambda: {
-            "action": PolicyFeature(type=FeatureType.ACTION, shape=(4,)),
-            "observation.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 128, 128)),
-            "observation.state": PolicyFeature(type=FeatureType.STATE, shape=(18,)),
-        }
-    )
-    features_map: dict[str, str] = field(
-        default_factory=lambda: {
-            "action": ACTION,
-            "observation.image": OBS_IMAGE,
-            "observation.state": OBS_STATE,
-        }
-    )
-    ################# args from hilserlrobotenv
-    reward_classifier_pretrained_path: Optional[str] = None
-    robot_config: Optional[RobotConfig] = None
-    teleop_config: Optional[TeleoperatorConfig] = None
-    wrapper: Optional[EnvTransformConfig] = None
-    mode: str = None  # Either "record", "replay", None
-    repo_id: Optional[str] = None
-    dataset_root: Optional[str] = None
-    num_episodes: int = 10  # only for record mode
-    episode: int = 0
-    device: str = "cuda"
-    push_to_hub: bool = True
-    pretrained_policy_name_or_path: Optional[str] = None
-    ############################
-
-    @property
-    def gym_kwargs(self) -> dict:
-        return {
-            "use_viewer": self.use_viewer,
-            "use_gamepad": self.use_gamepad,
-            "gripper_penalty": self.gripper_penalty,
-        }

lerobot/common/envs/factory.py

@@ -17,7 +17,7 @@ import importlib
 
 import gymnasium as gym
 
-from lerobot.common.envs.configs import AlohaEnv, EnvConfig, HILEnvConfig, PushtEnv, XarmEnv
+from lerobot.common.envs.configs import AlohaEnv, EnvConfig, PushtEnv, XarmEnv
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -27,8 +27,6 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
         return PushtEnv(**kwargs)
     elif env_type == "xarm":
         return XarmEnv(**kwargs)
-    elif env_type == "hil":
-        return HILEnvConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{env_type}' is not available.")
 
@@ -67,8 +65,5 @@ def make_env(cfg: EnvConfig, n_envs: int = 1, use_async_envs: bool = False) -> g
     env = env_cls(
         [lambda: gym.make(gym_handle, disable_env_checker=True, **cfg.gym_kwargs) for _ in range(n_envs)]
     )
-    # TODO: add observation processor wrapper and remove preprocess_observation in the codebase
-    # https://github.com/Farama-Foundation/Gymnasium/blob/main/gymnasium/wrappers/vector/vectorize_observation.py#L19,
-    # env = ObservationProcessorWrapper(env=env)
 
     return env

lerobot/common/envs/utils.py

@@ -47,10 +47,6 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
             # TODO(aliberts, rcadene): use transforms.ToTensor()?
             img = torch.from_numpy(img)
 
-            # When preprocessing observations in a non-vectorized environment, we need to add a batch dimension.
-            # This is the case for human-in-the-loop RL where there is only one environment.
-            if img.ndim == 3:
-                img = img.unsqueeze(0)
             # 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=}"
@@ -66,18 +62,13 @@ def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Ten
             return_observations[imgkey] = img
 
     if "environment_state" in observations:
-        env_state = torch.from_numpy(observations["environment_state"]).float()
-        if env_state.dim() == 1:
-            env_state = env_state.unsqueeze(0)
-
-        return_observations["observation.environment_state"] = env_state
+        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
-    agent_pos = torch.from_numpy(observations["agent_pos"]).float()
-    if agent_pos.dim() == 1:
-        agent_pos = agent_pos.unsqueeze(0)
-    return_observations["observation.state"] = agent_pos
-
+    # requirement for "agent_pos"
+    return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
     return return_observations
 
 

lerobot/common/errors.py

@@ -1,43 +0,0 @@
-# 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.
-
-
-class DeviceNotConnectedError(ConnectionError):
-    """Exception raised when the device is not connected."""
-
-    def __init__(self, message="This device is not connected. Try calling `connect()` first."):
-        self.message = message
-        super().__init__(self.message)
-
-
-class DeviceAlreadyConnectedError(ConnectionError):
-    """Exception raised when the device is already connected."""
-
-    def __init__(
-        self,
-        message="This device is already connected. Try not calling `connect()` twice.",
-    ):
-        self.message = message
-        super().__init__(self.message)
-
-
-class InvalidActionError(ValueError):
-    """Exception raised when an action is already invalid."""
-
-    def __init__(
-        self,
-        message="The action is invalid. Check the value follows what it is expected from the action space.",
-    ):
-        self.message = message
-        super().__init__(self.message)

lerobot/common/model/kinematics.py

@@ -1,589 +0,0 @@
-# ruff: noqa: N806, N815, N803
-
-# 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 numpy as np
-from scipy.spatial.transform import Rotation
-
-
-def skew_symmetric(w):
-    """Creates the skew-symmetric matrix from a 3D vector."""
-    return np.array([[0, -w[2], w[1]], [w[2], 0, -w[0]], [-w[1], w[0], 0]])
-
-
-def rodrigues_rotation(w, theta):
-    """Computes the rotation matrix using Rodrigues' formula."""
-    w_hat = skew_symmetric(w)
-    return np.eye(3) + np.sin(theta) * w_hat + (1 - np.cos(theta)) * w_hat @ w_hat
-
-
-def screw_axis_to_transform(S, theta):
-    """Converts a screw axis to a 4x4 transformation matrix."""
-    S_w = S[:3]
-    S_v = S[3:]
-    if np.allclose(S_w, 0) and np.linalg.norm(S_v) == 1:  # Pure translation
-        T = np.eye(4)
-        T[:3, 3] = S_v * theta
-    elif np.linalg.norm(S_w) == 1:  # Rotation and translation
-        w_hat = skew_symmetric(S_w)
-        R = np.eye(3) + np.sin(theta) * w_hat + (1 - np.cos(theta)) * w_hat @ w_hat
-        t = (np.eye(3) * theta + (1 - np.cos(theta)) * w_hat + (theta - np.sin(theta)) * w_hat @ w_hat) @ S_v
-        T = np.eye(4)
-        T[:3, :3] = R
-        T[:3, 3] = t
-    else:
-        raise ValueError("Invalid screw axis parameters")
-    return T
-
-
-def pose_difference_se3(pose1, pose2):
-    """
-    Calculates the SE(3) difference between two 4x4 homogeneous transformation matrices.
-    SE(3) (Special Euclidean Group) represents rigid body transformations in 3D space, combining rotation (SO(3)) and translation.
-    Each 4x4 matrix has the following structure, a 3x3 rotation matrix in the top-left and a 3x1 translation vector in the top-right:
-
-    [R11 R12 R13 tx]
-    [R21 R22 R23 ty]
-    [R31 R32 R33 tz]
-    [ 0   0   0   1]
-
-    where Rij is the 3x3 rotation matrix and [tx,ty,tz] is the translation vector.
-
-    pose1 - pose2
-
-    Args:
-        pose1: A 4x4 numpy array representing the first pose.
-        pose2: A 4x4 numpy array representing the second pose.
-
-    Returns:
-        A tuple (translation_diff, rotation_diff) where:
-        - translation_diff is a 3x1 numpy array representing the translational difference.
-        - rotation_diff is a 3x1 numpy array representing the rotational difference in axis-angle representation.
-    """
-
-    # Extract rotation matrices from poses
-    R1 = pose1[:3, :3]
-    R2 = pose2[:3, :3]
-
-    # Calculate translational difference
-    translation_diff = pose1[:3, 3] - pose2[:3, 3]
-
-    # Calculate rotational difference using scipy's Rotation library
-    R_diff = Rotation.from_matrix(R1 @ R2.T)
-    rotation_diff = R_diff.as_rotvec()  # Convert to axis-angle representation
-
-    return np.concatenate([translation_diff, rotation_diff])
-
-
-def se3_error(target_pose, current_pose):
-    pos_error = target_pose[:3, 3] - current_pose[:3, 3]
-    R_target = target_pose[:3, :3]
-    R_current = current_pose[:3, :3]
-    R_error = R_target @ R_current.T
-    rot_error = Rotation.from_matrix(R_error).as_rotvec()
-    return np.concatenate([pos_error, rot_error])
-
-
-class RobotKinematics:
-    """Robot kinematics class supporting multiple robot models."""
-
-    # Robot measurements dictionary
-    ROBOT_MEASUREMENTS = {
-        "koch": {
-            "gripper": [0.239, -0.001, 0.024],
-            "wrist": [0.209, 0, 0.024],
-            "forearm": [0.108, 0, 0.02],
-            "humerus": [0, 0, 0.036],
-            "shoulder": [0, 0, 0],
-            "base": [0, 0, 0.02],
-        },
-        "so100": {
-            "gripper": [0.320, 0, 0.050],
-            "wrist": [0.278, 0, 0.050],
-            "forearm": [0.143, 0, 0.044],
-            "humerus": [0.031, 0, 0.072],
-            "shoulder": [0, 0, 0],
-            "base": [0, 0, 0.02],
-        },
-        "moss": {
-            "gripper": [0.246, 0.013, 0.111],
-            "wrist": [0.245, 0.002, 0.064],
-            "forearm": [0.122, 0, 0.064],
-            "humerus": [0.001, 0.001, 0.063],
-            "shoulder": [0, 0, 0],
-            "base": [0, 0, 0.02],
-        },
-        "so101": {
-            "gripper": [0.33, 0.0, 0.285],
-            "wrist": [0.30, 0.0, 0.267],
-            "forearm": [0.25, 0.0, 0.266],
-            "humerus": [0.06, 0.0, 0.264],
-            "shoulder": [0.0, 0.0, 0.238],
-            "base": [0.0, 0.0, 0.12],
-        },
-    }
-
-    def __init__(self, robot_type="so100"):
-        """Initialize kinematics for the specified robot type.
-
-        Args:
-            robot_type: String specifying the robot model ("koch", "so100", or "moss")
-        """
-        if robot_type not in self.ROBOT_MEASUREMENTS:
-            raise ValueError(
-                f"Unknown robot type: {robot_type}. Available types: {list(self.ROBOT_MEASUREMENTS.keys())}"
-            )
-
-        self.robot_type = robot_type
-        self.measurements = self.ROBOT_MEASUREMENTS[robot_type]
-
-        # Initialize all transformation matrices and screw axes
-        self._setup_transforms()
-
-    def _create_translation_matrix(self, x=0, y=0, z=0):
-        """Create a 4x4 translation matrix."""
-        return np.array([[1, 0, 0, x], [0, 1, 0, y], [0, 0, 1, z], [0, 0, 0, 1]])
-
-    def _setup_transforms(self):
-        """Setup all transformation matrices and screw axes for the robot."""
-        # Set up rotation matrices (constant across robot types)
-
-        # Gripper orientation
-        self.gripper_X0 = np.array(
-            [
-                [1, 0, 0, 0],
-                [0, 0, 1, 0],
-                [0, -1, 0, 0],
-                [0, 0, 0, 1],
-            ]
-        )
-
-        # Wrist orientation
-        self.wrist_X0 = np.array(
-            [
-                [0, -1, 0, 0],
-                [1, 0, 0, 0],
-                [0, 0, 1, 0],
-                [0, 0, 0, 1],
-            ]
-        )
-
-        # Base orientation
-        self.base_X0 = np.array(
-            [
-                [0, 0, 1, 0],
-                [1, 0, 0, 0],
-                [0, 1, 0, 0],
-                [0, 0, 0, 1],
-            ]
-        )
-
-        # Gripper
-        # Screw axis of gripper frame wrt base frame
-        self.S_BG = np.array(
-            [
-                1,
-                0,
-                0,
-                0,
-                self.measurements["gripper"][2],
-                -self.measurements["gripper"][1],
-            ]
-        )
-
-        # Gripper origin to centroid transform
-        self.X_GoGc = self._create_translation_matrix(x=0.07)
-
-        # Gripper origin to tip transform
-        self.X_GoGt = self._create_translation_matrix(x=0.12)
-
-        # 0-position gripper frame pose wrt base
-        self.X_BoGo = self._create_translation_matrix(
-            x=self.measurements["gripper"][0],
-            y=self.measurements["gripper"][1],
-            z=self.measurements["gripper"][2],
-        )
-
-        # Wrist
-        # Screw axis of wrist frame wrt base frame
-        self.S_BR = np.array([0, 1, 0, -self.measurements["wrist"][2], 0, self.measurements["wrist"][0]])
-
-        # 0-position origin to centroid transform
-        self.X_RoRc = self._create_translation_matrix(x=0.0035, y=-0.002)
-
-        # 0-position wrist frame pose wrt base
-        self.X_BR = self._create_translation_matrix(
-            x=self.measurements["wrist"][0],
-            y=self.measurements["wrist"][1],
-            z=self.measurements["wrist"][2],
-        )
-
-        # Forearm
-        # Screw axis of forearm frame wrt base frame
-        self.S_BF = np.array(
-            [
-                0,
-                1,
-                0,
-                -self.measurements["forearm"][2],
-                0,
-                self.measurements["forearm"][0],
-            ]
-        )
-
-        # Forearm origin + centroid transform
-        self.X_FoFc = self._create_translation_matrix(x=0.036)  # spellchecker:disable-line
-
-        # 0-position forearm frame pose wrt base
-        self.X_BF = self._create_translation_matrix(
-            x=self.measurements["forearm"][0],
-            y=self.measurements["forearm"][1],
-            z=self.measurements["forearm"][2],
-        )
-
-        # Humerus
-        # Screw axis of humerus frame wrt base frame
-        self.S_BH = np.array(
-            [
-                0,
-                -1,
-                0,
-                self.measurements["humerus"][2],
-                0,
-                -self.measurements["humerus"][0],
-            ]
-        )
-
-        # Humerus origin to centroid transform
-        self.X_HoHc = self._create_translation_matrix(x=0.0475)
-
-        # 0-position humerus frame pose wrt base
-        self.X_BH = self._create_translation_matrix(
-            x=self.measurements["humerus"][0],
-            y=self.measurements["humerus"][1],
-            z=self.measurements["humerus"][2],
-        )
-
-        # Shoulder
-        # Screw axis of shoulder frame wrt Base frame
-        self.S_BS = np.array([0, 0, -1, 0, 0, 0])
-
-        # Shoulder origin to centroid transform
-        self.X_SoSc = self._create_translation_matrix(x=-0.017, z=0.0235)
-
-        # 0-position shoulder frame pose wrt base
-        self.X_BS = self._create_translation_matrix(
-            x=self.measurements["shoulder"][0],
-            y=self.measurements["shoulder"][1],
-            z=self.measurements["shoulder"][2],
-        )
-
-        # Base
-        # Base origin to centroid transform
-        self.X_BoBc = self._create_translation_matrix(y=0.015)
-
-        # World to base transform
-        self.X_WoBo = self._create_translation_matrix(
-            x=self.measurements["base"][0],
-            y=self.measurements["base"][1],
-            z=self.measurements["base"][2],
-        )
-
-        # Pre-compute gripper post-multiplication matrix
-        self._fk_gripper_post = self.X_GoGc @ self.X_BoGo @ self.gripper_X0
-
-    def fk_base(self):
-        """Forward kinematics for the base frame."""
-        return self.X_WoBo @ self.X_BoBc @ self.base_X0
-
-    def fk_shoulder(self, robot_pos_deg):
-        """Forward kinematics for the shoulder frame."""
-        robot_pos_rad = robot_pos_deg / 180 * np.pi
-        return self.X_WoBo @ screw_axis_to_transform(self.S_BS, robot_pos_rad[0]) @ self.X_SoSc @ self.X_BS
-
-    def fk_humerus(self, robot_pos_deg):
-        """Forward kinematics for the humerus frame."""
-        robot_pos_rad = robot_pos_deg / 180 * np.pi
-
-        theta_shoulder_pan = robot_pos_rad[0]
-        # NOTE: Negate shoulder lift angle for all robot types
-        theta_shoulder_lift = -robot_pos_rad[1]
-
-        return (
-            self.X_WoBo
-            @ screw_axis_to_transform(self.S_BS, theta_shoulder_pan)
-            @ screw_axis_to_transform(self.S_BH, theta_shoulder_lift)
-            @ self.X_HoHc
-            @ self.X_BH
-        )
-
-    def fk_forearm(self, robot_pos_deg):
-        """Forward kinematics for the forearm frame."""
-        robot_pos_rad = robot_pos_deg / 180 * np.pi
-
-        theta_shoulder_pan = robot_pos_rad[0]
-        # NOTE: Negate shoulder lift angle for all robot types
-        theta_shoulder_lift = -robot_pos_rad[1]
-        theta_elbow_flex = robot_pos_rad[2]
-
-        return (
-            self.X_WoBo
-            @ screw_axis_to_transform(self.S_BS, theta_shoulder_pan)
-            @ screw_axis_to_transform(self.S_BH, theta_shoulder_lift)
-            @ screw_axis_to_transform(self.S_BF, theta_elbow_flex)
-            @ self.X_FoFc  # spellchecker:disable-line
-            @ self.X_BF
-        )
-
-    def fk_wrist(self, robot_pos_deg):
-        """Forward kinematics for the wrist frame."""
-        robot_pos_rad = robot_pos_deg / 180 * np.pi
-
-        theta_shoulder_pan = robot_pos_rad[0]
-        # NOTE: Negate shoulder lift angle for all robot types
-        theta_shoulder_lift = -robot_pos_rad[1]
-        theta_elbow_flex = robot_pos_rad[2]
-        theta_wrist_flex = robot_pos_rad[3]
-
-        return (
-            self.X_WoBo
-            @ screw_axis_to_transform(self.S_BS, theta_shoulder_pan)
-            @ screw_axis_to_transform(self.S_BH, theta_shoulder_lift)
-            @ screw_axis_to_transform(self.S_BF, theta_elbow_flex)
-            @ screw_axis_to_transform(self.S_BR, theta_wrist_flex)
-            @ self.X_RoRc
-            @ self.X_BR
-            @ self.wrist_X0
-        )
-
-    def fk_gripper(self, robot_pos_deg):
-        """Forward kinematics for the gripper frame."""
-        robot_pos_rad = robot_pos_deg / 180 * np.pi
-
-        theta_shoulder_pan = robot_pos_rad[0]
-        # NOTE: Negate shoulder lift angle for all robot types
-        theta_shoulder_lift = -robot_pos_rad[1]
-        theta_elbow_flex = robot_pos_rad[2]
-        theta_wrist_flex = robot_pos_rad[3]
-        theta_wrist_roll = robot_pos_rad[4]
-
-        return (
-            self.X_WoBo
-            @ screw_axis_to_transform(self.S_BS, theta_shoulder_pan)
-            @ screw_axis_to_transform(self.S_BH, theta_shoulder_lift)
-            @ screw_axis_to_transform(self.S_BF, theta_elbow_flex)
-            @ screw_axis_to_transform(self.S_BR, theta_wrist_flex)
-            @ screw_axis_to_transform(self.S_BG, theta_wrist_roll)
-            @ self._fk_gripper_post
-        )
-
-    def fk_gripper_tip(self, robot_pos_deg):
-        """Forward kinematics for the gripper tip frame."""
-        robot_pos_rad = robot_pos_deg / 180 * np.pi
-
-        theta_shoulder_pan = robot_pos_rad[0]
-        # Negate shoulder lift angle for all robot types
-        theta_shoulder_lift = -robot_pos_rad[1]
-        theta_elbow_flex = robot_pos_rad[2]
-        theta_wrist_flex = robot_pos_rad[3]
-        theta_wrist_roll = robot_pos_rad[4]
-
-        return (
-            self.X_WoBo
-            @ screw_axis_to_transform(self.S_BS, theta_shoulder_pan)
-            @ screw_axis_to_transform(self.S_BH, theta_shoulder_lift)
-            @ screw_axis_to_transform(self.S_BF, theta_elbow_flex)
-            @ screw_axis_to_transform(self.S_BR, theta_wrist_flex)
-            @ screw_axis_to_transform(self.S_BG, theta_wrist_roll)
-            @ self.X_GoGt
-            @ self.X_BoGo
-            @ self.gripper_X0
-        )
-
-    def compute_jacobian(self, robot_pos_deg, fk_func=None):
-        """Finite differences to compute the Jacobian.
-        J(i, j) represents how the ith component of the end-effector's velocity changes wrt a small change
-        in the jth joint's velocity.
-
-        Args:
-            robot_pos_deg: Current joint positions in degrees
-            fk_func: Forward kinematics function to use (defaults to fk_gripper)
-        """
-        if fk_func is None:
-            fk_func = self.fk_gripper
-
-        eps = 1e-8
-        jac = np.zeros(shape=(6, 5))
-        delta = np.zeros(len(robot_pos_deg[:-1]), dtype=np.float64)
-        for el_ix in range(len(robot_pos_deg[:-1])):
-            delta *= 0
-            delta[el_ix] = eps / 2
-            Sdot = (
-                pose_difference_se3(
-                    fk_func(robot_pos_deg[:-1] + delta),
-                    fk_func(robot_pos_deg[:-1] - delta),
-                )
-                / eps
-            )
-            jac[:, el_ix] = Sdot
-        return jac
-
-    def compute_positional_jacobian(self, robot_pos_deg, fk_func=None):
-        """Finite differences to compute the positional Jacobian.
-        J(i, j) represents how the ith component of the end-effector's position changes wrt a small change
-        in the jth joint's velocity.
-
-        Args:
-            robot_pos_deg: Current joint positions in degrees
-            fk_func: Forward kinematics function to use (defaults to fk_gripper)
-        """
-        if fk_func is None:
-            fk_func = self.fk_gripper
-
-        eps = 1e-8
-        jac = np.zeros(shape=(3, 5))
-        delta = np.zeros(len(robot_pos_deg[:-1]), dtype=np.float64)
-        for el_ix in range(len(robot_pos_deg[:-1])):
-            delta *= 0
-            delta[el_ix] = eps / 2
-            Sdot = (
-                fk_func(robot_pos_deg[:-1] + delta)[:3, 3] - fk_func(robot_pos_deg[:-1] - delta)[:3, 3]
-            ) / eps
-            jac[:, el_ix] = Sdot
-        return jac
-
-    def ik(self, current_joint_pos, desired_ee_pose, position_only=True, fk_func=None):
-        """Inverse kinematics using gradient descent.
-
-        Args:
-            current_joint_state: Initial joint positions in degrees
-            desired_ee_pose: Target end-effector pose as a 4x4 transformation matrix
-            position_only: If True, only match end-effector position, not orientation
-            fk_func: Forward kinematics function to use (defaults to fk_gripper)
-
-        Returns:
-            Joint positions in degrees that achieve the desired end-effector pose
-        """
-        if fk_func is None:
-            fk_func = self.fk_gripper
-
-        # Do gradient descent.
-        current_joint_state = current_joint_pos.copy()
-        max_iterations = 5
-        learning_rate = 1
-        for _ in range(max_iterations):
-            current_ee_pose = fk_func(current_joint_state)
-            if not position_only:
-                error = se3_error(desired_ee_pose, current_ee_pose)
-                jac = self.compute_jacobian(current_joint_state, fk_func)
-            else:
-                error = desired_ee_pose[:3, 3] - current_ee_pose[:3, 3]
-                jac = self.compute_positional_jacobian(current_joint_state, fk_func)
-            delta_angles = np.linalg.pinv(jac) @ error
-            current_joint_state[:-1] += learning_rate * delta_angles
-
-            if np.linalg.norm(error) < 5e-3:
-                return current_joint_state
-        return current_joint_state
-
-
-if __name__ == "__main__":
-    import time
-
-    def run_test(robot_type):
-        """Run test suite for a specific robot type."""
-        print(f"\n--- Testing {robot_type.upper()} Robot ---")
-
-        # Initialize kinematics for this robot
-        robot = RobotKinematics(robot_type)
-
-        # Test 1: Forward kinematics consistency
-        print("Test 1: Forward kinematics consistency")
-        test_angles = np.array([30, 45, -30, 20, 10, 0])  # Example joint angles in degrees
-
-        # Calculate FK for different joints
-        shoulder_pose = robot.fk_shoulder(test_angles)
-        humerus_pose = robot.fk_humerus(test_angles)
-        forearm_pose = robot.fk_forearm(test_angles)
-        wrist_pose = robot.fk_wrist(test_angles)
-        gripper_pose = robot.fk_gripper(test_angles)
-        gripper_tip_pose = robot.fk_gripper_tip(test_angles)
-
-        # Check that poses form a consistent kinematic chain (positions should be progressively further from origin)
-        distances = [
-            np.linalg.norm(shoulder_pose[:3, 3]),
-            np.linalg.norm(humerus_pose[:3, 3]),
-            np.linalg.norm(forearm_pose[:3, 3]),
-            np.linalg.norm(wrist_pose[:3, 3]),
-            np.linalg.norm(gripper_pose[:3, 3]),
-            np.linalg.norm(gripper_tip_pose[:3, 3]),
-        ]
-
-        # Check if distances generally increase along the chain
-        is_consistent = all(distances[i] <= distances[i + 1] for i in range(len(distances) - 1))
-        print(f"  Pose distances from origin: {[round(d, 3) for d in distances]}")
-        print(f"  Kinematic chain consistency: {'PASSED' if is_consistent else 'FAILED'}")
-
-        # Test 2: Jacobian computation
-        print("Test 2: Jacobian computation")
-        jacobian = robot.compute_jacobian(test_angles)
-        positional_jacobian = robot.compute_positional_jacobian(test_angles)
-
-        # Check shapes
-        jacobian_shape_ok = jacobian.shape == (6, 5)
-        pos_jacobian_shape_ok = positional_jacobian.shape == (3, 5)
-
-        print(f"  Jacobian shape: {'PASSED' if jacobian_shape_ok else 'FAILED'}")
-        print(f"  Positional Jacobian shape: {'PASSED' if pos_jacobian_shape_ok else 'FAILED'}")
-
-        # Test 3: Inverse kinematics
-        print("Test 3: Inverse kinematics (position only)")
-
-        # Generate target pose from known joint angles
-        original_angles = np.array([10, 20, 30, -10, 5, 0])
-        target_pose = robot.fk_gripper(original_angles)
-
-        # Start IK from a different position
-        initial_guess = np.array([0.0, 0.0, 0.0, 0.0, 0.0, 0.0])
-
-        # Measure IK performance
-        start_time = time.time()
-        computed_angles = robot.ik(initial_guess.copy(), target_pose)
-        ik_time = time.time() - start_time
-
-        # Compute resulting pose from IK solution
-        result_pose = robot.fk_gripper(computed_angles)
-
-        # Calculate position error
-        pos_error = np.linalg.norm(target_pose[:3, 3] - result_pose[:3, 3])
-        passed = pos_error < 0.01  # Accept errors less than 1cm
-
-        print(f"  IK computation time: {ik_time:.4f} seconds")
-        print(f"  Position error: {pos_error:.4f}")
-        print(f"  IK position accuracy: {'PASSED' if passed else 'FAILED'}")
-
-        return is_consistent and jacobian_shape_ok and pos_jacobian_shape_ok and passed
-
-    # Run tests for all robot types
-    results = {}
-    for robot_type in ["koch", "so100", "moss", "so101"]:
-        results[robot_type] = run_test(robot_type)
-
-    # Print overall summary
-    print("\n=== Test Summary ===")
-    all_passed = all(results.values())
-    for robot_type, passed in results.items():
-        print(f"{robot_type.upper()}: {'PASSED' if passed else 'FAILED'}")
-    print(f"\nOverall: {'ALL TESTS PASSED' if all_passed else 'SOME TESTS FAILED'}")

lerobot/common/motors/__init__.py

@@ -1 +0,0 @@
-from .motors_bus import Motor, MotorCalibration, MotorNormMode, MotorsBus

lerobot/common/motors/dynamixel/__init__.py

@@ -1,2 +0,0 @@
-from .dynamixel import DriveMode, DynamixelMotorsBus, OperatingMode, TorqueMode
-from .tables import *

lerobot/common/motors/dynamixel/dynamixel.py

@@ -1,264 +0,0 @@
-# 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.
-
-# TODO(aliberts): Should we implement FastSyncRead/Write?
-# https://github.com/ROBOTIS-GIT/DynamixelSDK/pull/643
-# https://github.com/ROBOTIS-GIT/DynamixelSDK/releases/tag/3.8.2
-# https://emanual.robotis.com/docs/en/dxl/protocol2/#fast-sync-read-0x8a
-# -> Need to check compatibility across models
-
-import logging
-from copy import deepcopy
-from enum import Enum
-
-from lerobot.common.utils.encoding_utils import decode_twos_complement, encode_twos_complement
-
-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
-from .tables import (
-    AVAILABLE_BAUDRATES,
-    MODEL_BAUDRATE_TABLE,
-    MODEL_CONTROL_TABLE,
-    MODEL_ENCODING_TABLE,
-    MODEL_NUMBER_TABLE,
-    MODEL_RESOLUTION,
-)
-
-PROTOCOL_VERSION = 2.0
-DEFAULT_BAUDRATE = 1_000_000
-DEFAULT_TIMEOUT_MS = 1000
-
-NORMALIZED_DATA = ["Goal_Position", "Present_Position"]
-CONVERT_UINT32_TO_INT32_REQUIRED = ["Goal_Position", "Present_Position"]
-
-logger = logging.getLogger(__name__)
-
-
-class OperatingMode(Enum):
-    # DYNAMIXEL only controls current(torque) regardless of speed and position. This mode is ideal for a
-    # gripper or a system that only uses current(torque) control or a system that has additional
-    # velocity/position controllers.
-    CURRENT = 0
-
-    # This mode controls velocity. This mode is identical to the Wheel Mode(endless) from existing DYNAMIXEL.
-    # This mode is ideal for wheel-type robots.
-    VELOCITY = 1
-
-    # This mode controls position. This mode is identical to the Joint Mode from existing DYNAMIXEL. Operating
-    # position range is limited by the Max Position Limit(48) and the Min Position Limit(52). This mode is
-    # ideal for articulated robots that each joint rotates less than 360 degrees.
-    POSITION = 3
-
-    # This mode controls position. This mode is identical to the Multi-turn Position Control from existing
-    # DYNAMIXEL. 512 turns are supported(-256[rev] ~ 256[rev]). This mode is ideal for multi-turn wrists or
-    # conveyer systems or a system that requires an additional reduction gear. Note that Max Position
-    # Limit(48), Min Position Limit(52) are not used on Extended Position Control Mode.
-    EXTENDED_POSITION = 4
-
-    # This mode controls both position and current(torque). Up to 512 turns are supported (-256[rev] ~
-    # 256[rev]). This mode is ideal for a system that requires both position and current control such as
-    # articulated robots or grippers.
-    CURRENT_POSITION = 5
-
-    # This mode directly controls PWM output. (Voltage Control Mode)
-    PWM = 16
-
-
-class DriveMode(Enum):
-    NON_INVERTED = 0
-    INVERTED = 1
-
-
-class TorqueMode(Enum):
-    ENABLED = 1
-    DISABLED = 0
-
-
-def _split_into_byte_chunks(value: int, length: int) -> list[int]:
-    import dynamixel_sdk as dxl
-
-    if length == 1:
-        data = [value]
-    elif length == 2:
-        data = [dxl.DXL_LOBYTE(value), dxl.DXL_HIBYTE(value)]
-    elif length == 4:
-        data = [
-            dxl.DXL_LOBYTE(dxl.DXL_LOWORD(value)),
-            dxl.DXL_HIBYTE(dxl.DXL_LOWORD(value)),
-            dxl.DXL_LOBYTE(dxl.DXL_HIWORD(value)),
-            dxl.DXL_HIBYTE(dxl.DXL_HIWORD(value)),
-        ]
-    return data
-
-
-class DynamixelMotorsBus(MotorsBus):
-    """
-    The Dynamixel implementation for a MotorsBus. It relies on the python dynamixel sdk to communicate with
-    the motors. For more info, see the Dynamixel SDK Documentation:
-    https://emanual.robotis.com/docs/en/software/dynamixel/dynamixel_sdk/sample_code/python_read_write_protocol_2_0/#python-read-write-protocol-20
-    """
-
-    apply_drive_mode = False
-    available_baudrates = deepcopy(AVAILABLE_BAUDRATES)
-    default_baudrate = DEFAULT_BAUDRATE
-    default_timeout = DEFAULT_TIMEOUT_MS
-    model_baudrate_table = deepcopy(MODEL_BAUDRATE_TABLE)
-    model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
-    model_encoding_table = deepcopy(MODEL_ENCODING_TABLE)
-    model_number_table = deepcopy(MODEL_NUMBER_TABLE)
-    model_resolution_table = deepcopy(MODEL_RESOLUTION)
-    normalized_data = deepcopy(NORMALIZED_DATA)
-
-    def __init__(
-        self,
-        port: str,
-        motors: dict[str, Motor],
-        calibration: dict[str, MotorCalibration] | None = None,
-    ):
-        super().__init__(port, motors, calibration)
-        import dynamixel_sdk as dxl
-
-        self.port_handler = dxl.PortHandler(self.port)
-        self.packet_handler = dxl.PacketHandler(PROTOCOL_VERSION)
-        self.sync_reader = dxl.GroupSyncRead(self.port_handler, self.packet_handler, 0, 0)
-        self.sync_writer = dxl.GroupSyncWrite(self.port_handler, self.packet_handler, 0, 0)
-        self._comm_success = dxl.COMM_SUCCESS
-        self._no_error = 0x00
-
-    def _assert_protocol_is_compatible(self, instruction_name: str) -> None:
-        pass
-
-    def _handshake(self) -> None:
-        self._assert_motors_exist()
-
-    def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
-        model = self.motors[motor].model
-        search_baudrates = (
-            [initial_baudrate] if initial_baudrate is not None else self.model_baudrate_table[model]
-        )
-
-        for baudrate in search_baudrates:
-            self.set_baudrate(baudrate)
-            id_model = self.broadcast_ping()
-            if id_model:
-                found_id, found_model = next(iter(id_model.items()))
-                expected_model_nb = self.model_number_table[model]
-                if found_model != expected_model_nb:
-                    raise RuntimeError(
-                        f"Found one motor on {baudrate=} with id={found_id} but it has a "
-                        f"model number '{found_model}' different than the one expected: '{expected_model_nb}'. "
-                        f"Make sure you are connected only connected to the '{motor}' motor (model '{model}')."
-                    )
-                return baudrate, found_id
-
-        raise RuntimeError(f"Motor '{motor}' (model '{model}') was not found. Make sure it is connected.")
-
-    def configure_motors(self) -> None:
-        # By default, Dynamixel motors have a 500µs delay response time (corresponding to a value of 250 on
-        # the 'Return_Delay_Time' address). We ensure this is reduced to the minimum of 2µs (value of 0).
-        for motor in self.motors:
-            self.write("Return_Delay_Time", motor, 0)
-
-    @property
-    def is_calibrated(self) -> bool:
-        return self.calibration == self.read_calibration()
-
-    def read_calibration(self) -> dict[str, MotorCalibration]:
-        offsets = self.sync_read("Homing_Offset", normalize=False)
-        mins = self.sync_read("Min_Position_Limit", normalize=False)
-        maxes = self.sync_read("Max_Position_Limit", normalize=False)
-        drive_modes = self.sync_read("Drive_Mode", normalize=False)
-
-        calibration = {}
-        for motor, m in self.motors.items():
-            calibration[motor] = MotorCalibration(
-                id=m.id,
-                drive_mode=drive_modes[motor],
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
-            )
-
-        return calibration
-
-    def write_calibration(self, calibration_dict: dict[str, MotorCalibration]) -> None:
-        for motor, calibration in calibration_dict.items():
-            self.write("Homing_Offset", motor, calibration.homing_offset)
-            self.write("Min_Position_Limit", motor, calibration.range_min)
-            self.write("Max_Position_Limit", motor, calibration.range_max)
-
-        self.calibration = calibration_dict
-
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
-        for motor in self._get_motors_list(motors):
-            self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
-
-    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
-        addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
-
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
-        for motor in self._get_motors_list(motors):
-            self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
-
-    def _encode_sign(self, data_name: str, ids_values: dict[int, int]) -> dict[int, int]:
-        for id_ in ids_values:
-            model = self._id_to_model(id_)
-            encoding_table = self.model_encoding_table.get(model)
-            if encoding_table and data_name in encoding_table:
-                n_bytes = encoding_table[data_name]
-                ids_values[id_] = encode_twos_complement(ids_values[id_], n_bytes)
-
-        return ids_values
-
-    def _decode_sign(self, data_name: str, ids_values: dict[int, int]) -> dict[int, int]:
-        for id_ in ids_values:
-            model = self._id_to_model(id_)
-            encoding_table = self.model_encoding_table.get(model)
-            if encoding_table and data_name in encoding_table:
-                n_bytes = encoding_table[data_name]
-                ids_values[id_] = decode_twos_complement(ids_values[id_], n_bytes)
-
-        return ids_values
-
-    def _get_half_turn_homings(self, positions: dict[NameOrID, Value]) -> dict[NameOrID, Value]:
-        """
-        On Dynamixel Motors:
-        Present_Position = Actual_Position + Homing_Offset
-        """
-        half_turn_homings = {}
-        for motor, pos in positions.items():
-            model = self._get_motor_model(motor)
-            max_res = self.model_resolution_table[model] - 1
-            half_turn_homings[motor] = int(max_res / 2) - pos
-
-        return half_turn_homings
-
-    def _split_into_byte_chunks(self, value: int, length: int) -> list[int]:
-        return _split_into_byte_chunks(value, length)
-
-    def broadcast_ping(self, num_retry: int = 0, raise_on_error: bool = False) -> dict[int, int] | None:
-        for n_try in range(1 + num_retry):
-            data_list, comm = self.packet_handler.broadcastPing(self.port_handler)
-            if self._is_comm_success(comm):
-                break
-            logger.debug(f"Broadcast ping failed on port '{self.port}' ({n_try=})")
-            logger.debug(self.packet_handler.getTxRxResult(comm))
-
-        if not self._is_comm_success(comm):
-            if raise_on_error:
-                raise ConnectionError(self.packet_handler.getTxRxResult(comm))
-
-            return
-
-        return {id_: data[0] for id_, data in data_list.items()}

lerobot/common/motors/dynamixel/tables.py

@@ -1,197 +0,0 @@
-# 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.
-
-# TODO(Steven): Consider doing the following:
-# from enum import Enum
-# class MyControlTableKey(Enum):
-#   ID = "ID"
-#   GOAL_SPEED = "Goal_Speed"
-#   ...
-#
-# MY_CONTROL_TABLE ={
-#   MyControlTableKey.ID.value: (5,1)
-#   MyControlTableKey.GOAL_SPEED.value: (46, 2)
-#   ...
-# }
-# This allows me do to:
-# bus.write(MyControlTableKey.GOAL_SPEED, ...)
-# Instead of:
-# bus.write("Goal_Speed", ...)
-# This is important for two reasons:
-# 1. The linter will tell me if I'm trying to use an invalid key, instead of me realizing when I get the RunTimeError
-# 2. We can change the value of the MyControlTableKey enums without impacting the client code
-
-
-# {data_name: (address, size_byte)}
-# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#control-table
-X_SERIES_CONTROL_TABLE = {
-    "Model_Number": (0, 2),
-    "Model_Information": (2, 4),
-    "Firmware_Version": (6, 1),
-    "ID": (7, 1),
-    "Baud_Rate": (8, 1),
-    "Return_Delay_Time": (9, 1),
-    "Drive_Mode": (10, 1),
-    "Operating_Mode": (11, 1),
-    "Secondary_ID": (12, 1),
-    "Protocol_Type": (13, 1),
-    "Homing_Offset": (20, 4),
-    "Moving_Threshold": (24, 4),
-    "Temperature_Limit": (31, 1),
-    "Max_Voltage_Limit": (32, 2),
-    "Min_Voltage_Limit": (34, 2),
-    "PWM_Limit": (36, 2),
-    "Current_Limit": (38, 2),
-    "Acceleration_Limit": (40, 4),
-    "Velocity_Limit": (44, 4),
-    "Max_Position_Limit": (48, 4),
-    "Min_Position_Limit": (52, 4),
-    "Shutdown": (63, 1),
-    "Torque_Enable": (64, 1),
-    "LED": (65, 1),
-    "Status_Return_Level": (68, 1),
-    "Registered_Instruction": (69, 1),
-    "Hardware_Error_Status": (70, 1),
-    "Velocity_I_Gain": (76, 2),
-    "Velocity_P_Gain": (78, 2),
-    "Position_D_Gain": (80, 2),
-    "Position_I_Gain": (82, 2),
-    "Position_P_Gain": (84, 2),
-    "Feedforward_2nd_Gain": (88, 2),
-    "Feedforward_1st_Gain": (90, 2),
-    "Bus_Watchdog": (98, 1),
-    "Goal_PWM": (100, 2),
-    "Goal_Current": (102, 2),
-    "Goal_Velocity": (104, 4),
-    "Profile_Acceleration": (108, 4),
-    "Profile_Velocity": (112, 4),
-    "Goal_Position": (116, 4),
-    "Realtime_Tick": (120, 2),
-    "Moving": (122, 1),
-    "Moving_Status": (123, 1),
-    "Present_PWM": (124, 2),
-    "Present_Current": (126, 2),
-    "Present_Velocity": (128, 4),
-    "Present_Position": (132, 4),
-    "Velocity_Trajectory": (136, 4),
-    "Position_Trajectory": (140, 4),
-    "Present_Input_Voltage": (144, 2),
-    "Present_Temperature": (146, 1),
-}
-
-# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#baud-rate8
-X_SERIES_BAUDRATE_TABLE = {
-    9_600: 0,
-    57_600: 1,
-    115_200: 2,
-    1_000_000: 3,
-    2_000_000: 4,
-    3_000_000: 5,
-    4_000_000: 6,
-}
-
-# {data_name: size_byte}
-X_SERIES_ENCODINGS_TABLE = {
-    "Homing_Offset": X_SERIES_CONTROL_TABLE["Homing_Offset"][1],
-    "Goal_PWM": X_SERIES_CONTROL_TABLE["Goal_PWM"][1],
-    "Goal_Current": X_SERIES_CONTROL_TABLE["Goal_Current"][1],
-    "Goal_Velocity": X_SERIES_CONTROL_TABLE["Goal_Velocity"][1],
-    "Present_PWM": X_SERIES_CONTROL_TABLE["Present_PWM"][1],
-    "Present_Current": X_SERIES_CONTROL_TABLE["Present_Current"][1],
-    "Present_Velocity": X_SERIES_CONTROL_TABLE["Present_Velocity"][1],
-}
-
-MODEL_ENCODING_TABLE = {
-    "x_series": X_SERIES_ENCODINGS_TABLE,
-    "xl330-m077": X_SERIES_ENCODINGS_TABLE,
-    "xl330-m288": X_SERIES_ENCODINGS_TABLE,
-    "xl430-w250": X_SERIES_ENCODINGS_TABLE,
-    "xm430-w350": X_SERIES_ENCODINGS_TABLE,
-    "xm540-w270": X_SERIES_ENCODINGS_TABLE,
-    "xc430-w150": X_SERIES_ENCODINGS_TABLE,
-}
-
-# {model: model_resolution}
-# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#specifications
-MODEL_RESOLUTION = {
-    "x_series": 4096,
-    "xl330-m077": 4096,
-    "xl330-m288": 4096,
-    "xl430-w250": 4096,
-    "xm430-w350": 4096,
-    "xm540-w270": 4096,
-    "xc430-w150": 4096,
-}
-
-# {model: model_number}
-# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#control-table-of-eeprom-area
-MODEL_NUMBER_TABLE = {
-    "xl330-m077": 1190,
-    "xl330-m288": 1200,
-    "xl430-w250": 1060,
-    "xm430-w350": 1020,
-    "xm540-w270": 1120,
-    "xc430-w150": 1070,
-}
-
-# {model: available_operating_modes}
-# https://emanual.robotis.com/docs/en/dxl/x/{MODEL}/#operating-mode11
-MODEL_OPERATING_MODES = {
-    "xl330-m077": [0, 1, 3, 4, 5, 16],
-    "xl330-m288": [0, 1, 3, 4, 5, 16],
-    "xl430-w250": [1, 3, 4, 16],
-    "xm430-w350": [0, 1, 3, 4, 5, 16],
-    "xm540-w270": [0, 1, 3, 4, 5, 16],
-    "xc430-w150": [1, 3, 4, 16],
-}
-
-MODEL_CONTROL_TABLE = {
-    "x_series": X_SERIES_CONTROL_TABLE,
-    "xl330-m077": X_SERIES_CONTROL_TABLE,
-    "xl330-m288": X_SERIES_CONTROL_TABLE,
-    "xl430-w250": X_SERIES_CONTROL_TABLE,
-    "xm430-w350": X_SERIES_CONTROL_TABLE,
-    "xm540-w270": X_SERIES_CONTROL_TABLE,
-    "xc430-w150": X_SERIES_CONTROL_TABLE,
-}
-
-MODEL_BAUDRATE_TABLE = {
-    "x_series": X_SERIES_BAUDRATE_TABLE,
-    "xl330-m077": X_SERIES_BAUDRATE_TABLE,
-    "xl330-m288": X_SERIES_BAUDRATE_TABLE,
-    "xl430-w250": X_SERIES_BAUDRATE_TABLE,
-    "xm430-w350": X_SERIES_BAUDRATE_TABLE,
-    "xm540-w270": X_SERIES_BAUDRATE_TABLE,
-    "xc430-w150": X_SERIES_BAUDRATE_TABLE,
-}
-
-AVAILABLE_BAUDRATES = [
-    9_600,
-    19_200,
-    38_400,
-    57_600,
-    115_200,
-    230_400,
-    460_800,
-    500_000,
-    576_000,
-    921_600,
-    1_000_000,
-    1_152_000,
-    2_000_000,
-    2_500_000,
-    3_000_000,
-    3_500_000,
-    4_000_000,
-]

lerobot/common/motors/feetech/__init__.py

@@ -1,2 +0,0 @@
-from .feetech import DriveMode, FeetechMotorsBus, OperatingMode, TorqueMode
-from .tables import *

lerobot/common/motors/feetech/feetech.py

@@ -1,458 +0,0 @@
-# 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
-from copy import deepcopy
-from enum import Enum
-from pprint import pformat
-
-from lerobot.common.utils.encoding_utils import decode_sign_magnitude, encode_sign_magnitude
-
-from ..motors_bus import Motor, MotorCalibration, MotorsBus, NameOrID, Value, get_address
-from .tables import (
-    FIRMWARE_MAJOR_VERSION,
-    FIRMWARE_MINOR_VERSION,
-    MODEL_BAUDRATE_TABLE,
-    MODEL_CONTROL_TABLE,
-    MODEL_ENCODING_TABLE,
-    MODEL_NUMBER,
-    MODEL_NUMBER_TABLE,
-    MODEL_PROTOCOL,
-    MODEL_RESOLUTION,
-    SCAN_BAUDRATES,
-)
-
-DEFAULT_PROTOCOL_VERSION = 0
-DEFAULT_BAUDRATE = 1_000_000
-DEFAULT_TIMEOUT_MS = 1000
-
-NORMALIZED_DATA = ["Goal_Position", "Present_Position"]
-
-logger = logging.getLogger(__name__)
-
-
-class OperatingMode(Enum):
-    # position servo mode
-    POSITION = 0
-    # The motor is in constant speed mode, which is controlled by parameter 0x2e, and the highest bit 15 is
-    # the direction bit
-    VELOCITY = 1
-    # PWM open-loop speed regulation mode, with parameter 0x2c running time parameter control, bit11 as
-    # direction bit
-    PWM = 2
-    # In step servo mode, the number of step progress is represented by parameter 0x2a, and the highest bit 15
-    # is the direction bit
-    STEP = 3
-
-
-class DriveMode(Enum):
-    NON_INVERTED = 0
-    INVERTED = 1
-
-
-class TorqueMode(Enum):
-    ENABLED = 1
-    DISABLED = 0
-
-
-def _split_into_byte_chunks(value: int, length: int) -> list[int]:
-    import scservo_sdk as scs
-
-    if length == 1:
-        data = [value]
-    elif length == 2:
-        data = [scs.SCS_LOBYTE(value), scs.SCS_HIBYTE(value)]
-    elif length == 4:
-        data = [
-            scs.SCS_LOBYTE(scs.SCS_LOWORD(value)),
-            scs.SCS_HIBYTE(scs.SCS_LOWORD(value)),
-            scs.SCS_LOBYTE(scs.SCS_HIWORD(value)),
-            scs.SCS_HIBYTE(scs.SCS_HIWORD(value)),
-        ]
-    return data
-
-
-def patch_setPacketTimeout(self, packet_length):  # noqa: N802
-    """
-    HACK: This patches the PortHandler behavior to set the correct packet timeouts.
-
-    It fixes https://gitee.com/ftservo/SCServoSDK/issues/IBY2S6
-    The bug is fixed on the official Feetech SDK repo (https://gitee.com/ftservo/FTServo_Python)
-    but because that version is not published on PyPI, we rely on the (unofficial) on that is, which needs
-    patching.
-    """
-    self.packet_start_time = self.getCurrentTime()
-    self.packet_timeout = (self.tx_time_per_byte * packet_length) + (self.tx_time_per_byte * 3.0) + 50
-
-
-class FeetechMotorsBus(MotorsBus):
-    """
-    The FeetechMotorsBus class allows to efficiently read and write to the attached motors. It relies on the
-    python feetech sdk to communicate with the motors, which is itself based on the dynamixel sdk.
-    """
-
-    apply_drive_mode = True
-    available_baudrates = deepcopy(SCAN_BAUDRATES)
-    default_baudrate = DEFAULT_BAUDRATE
-    default_timeout = DEFAULT_TIMEOUT_MS
-    model_baudrate_table = deepcopy(MODEL_BAUDRATE_TABLE)
-    model_ctrl_table = deepcopy(MODEL_CONTROL_TABLE)
-    model_encoding_table = deepcopy(MODEL_ENCODING_TABLE)
-    model_number_table = deepcopy(MODEL_NUMBER_TABLE)
-    model_resolution_table = deepcopy(MODEL_RESOLUTION)
-    normalized_data = deepcopy(NORMALIZED_DATA)
-
-    def __init__(
-        self,
-        port: str,
-        motors: dict[str, Motor],
-        calibration: dict[str, MotorCalibration] | None = None,
-        protocol_version: int = DEFAULT_PROTOCOL_VERSION,
-    ):
-        super().__init__(port, motors, calibration)
-        self.protocol_version = protocol_version
-        self._assert_same_protocol()
-        import scservo_sdk as scs
-
-        self.port_handler = scs.PortHandler(self.port)
-        # HACK: monkeypatch
-        self.port_handler.setPacketTimeout = patch_setPacketTimeout.__get__(
-            self.port_handler, scs.PortHandler
-        )
-        self.packet_handler = scs.PacketHandler(protocol_version)
-        self.sync_reader = scs.GroupSyncRead(self.port_handler, self.packet_handler, 0, 0)
-        self.sync_writer = scs.GroupSyncWrite(self.port_handler, self.packet_handler, 0, 0)
-        self._comm_success = scs.COMM_SUCCESS
-        self._no_error = 0x00
-
-        if any(MODEL_PROTOCOL[model] != self.protocol_version for model in self.models):
-            raise ValueError(f"Some motors are incompatible with protocol_version={self.protocol_version}")
-
-    def _assert_same_protocol(self) -> None:
-        if any(MODEL_PROTOCOL[model] != self.protocol_version for model in self.models):
-            raise RuntimeError("Some motors use an incompatible protocol.")
-
-    def _assert_protocol_is_compatible(self, instruction_name: str) -> None:
-        if instruction_name == "sync_read" and self.protocol_version == 1:
-            raise NotImplementedError(
-                "'Sync Read' is not available with Feetech motors using Protocol 1. Use 'Read' sequentially instead."
-            )
-        if instruction_name == "broadcast_ping" and self.protocol_version == 1:
-            raise NotImplementedError(
-                "'Broadcast Ping' is not available with Feetech motors using Protocol 1. Use 'Ping' sequentially instead."
-            )
-
-    def _assert_same_firmware(self) -> None:
-        firmware_versions = self._read_firmware_version(self.ids)
-        if len(set(firmware_versions.values())) != 1:
-            raise RuntimeError(
-                "Some Motors use different firmware versions:"
-                f"\n{pformat(firmware_versions)}\n"
-                "Update their firmware first using Feetech's software. "
-                "Visit https://www.feetechrc.com/software."
-            )
-
-    def _handshake(self) -> None:
-        self._assert_motors_exist()
-        self._assert_same_firmware()
-
-    def _find_single_motor(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
-        if self.protocol_version == 0:
-            return self._find_single_motor_p0(motor, initial_baudrate)
-        else:
-            return self._find_single_motor_p1(motor, initial_baudrate)
-
-    def _find_single_motor_p0(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
-        model = self.motors[motor].model
-        search_baudrates = (
-            [initial_baudrate] if initial_baudrate is not None else self.model_baudrate_table[model]
-        )
-        expected_model_nb = self.model_number_table[model]
-
-        for baudrate in search_baudrates:
-            self.set_baudrate(baudrate)
-            id_model = self.broadcast_ping()
-            if id_model:
-                found_id, found_model = next(iter(id_model.items()))
-                if found_model != expected_model_nb:
-                    raise RuntimeError(
-                        f"Found one motor on {baudrate=} with id={found_id} but it has a "
-                        f"model number '{found_model}' different than the one expected: '{expected_model_nb}'. "
-                        f"Make sure you are connected only connected to the '{motor}' motor (model '{model}')."
-                    )
-                return baudrate, found_id
-
-        raise RuntimeError(f"Motor '{motor}' (model '{model}') was not found. Make sure it is connected.")
-
-    def _find_single_motor_p1(self, motor: str, initial_baudrate: int | None = None) -> tuple[int, int]:
-        import scservo_sdk as scs
-
-        model = self.motors[motor].model
-        search_baudrates = (
-            [initial_baudrate] if initial_baudrate is not None else self.model_baudrate_table[model]
-        )
-        expected_model_nb = self.model_number_table[model]
-
-        for baudrate in search_baudrates:
-            self.set_baudrate(baudrate)
-            for id_ in range(scs.MAX_ID + 1):
-                found_model = self.ping(id_)
-                if found_model is not None:
-                    if found_model != expected_model_nb:
-                        raise RuntimeError(
-                            f"Found one motor on {baudrate=} with id={id_} but it has a "
-                            f"model number '{found_model}' different than the one expected: '{expected_model_nb}'. "
-                            f"Make sure you are connected only connected to the '{motor}' motor (model '{model}')."
-                        )
-                    return baudrate, id_
-
-        raise RuntimeError(f"Motor '{motor}' (model '{model}') was not found. Make sure it is connected.")
-
-    def configure_motors(self) -> None:
-        for motor in self.motors:
-            # By default, Feetech motors have a 500µs delay response time (corresponding to a value of 250 on
-            # the 'Return_Delay_Time' address). We ensure this is reduced to the minimum of 2µs (value of 0).
-            self.write("Return_Delay_Time", motor, 0)
-            # Set 'Maximum_Acceleration' to 254 to speedup acceleration and deceleration of the motors.
-            # Note: this address is not in the official STS3215 Memory Table
-            self.write("Maximum_Acceleration", motor, 254)
-            self.write("Acceleration", motor, 254)
-
-    @property
-    def is_calibrated(self) -> bool:
-        motors_calibration = self.read_calibration()
-        if set(motors_calibration) != set(self.calibration):
-            return False
-
-        same_ranges = all(
-            self.calibration[motor].range_min == cal.range_min
-            and self.calibration[motor].range_max == cal.range_max
-            for motor, cal in motors_calibration.items()
-        )
-        if self.protocol_version == 1:
-            return same_ranges
-
-        same_offsets = all(
-            self.calibration[motor].homing_offset == cal.homing_offset
-            for motor, cal in motors_calibration.items()
-        )
-        return same_ranges and same_offsets
-
-    def read_calibration(self) -> dict[str, MotorCalibration]:
-        offsets, mins, maxes = {}, {}, {}
-        drive_modes = dict.fromkeys(self.motors, 0)
-        for motor in self.motors:
-            mins[motor] = self.read("Min_Position_Limit", motor, normalize=False)
-            maxes[motor] = self.read("Max_Position_Limit", motor, normalize=False)
-            offsets[motor] = (
-                self.read("Homing_Offset", motor, normalize=False) if self.protocol_version == 0 else 0
-            )
-
-        calibration = {}
-        for motor, m in self.motors.items():
-            calibration[motor] = MotorCalibration(
-                id=m.id,
-                drive_mode=drive_modes[motor],
-                homing_offset=offsets[motor],
-                range_min=mins[motor],
-                range_max=maxes[motor],
-            )
-
-        return calibration
-
-    def write_calibration(self, calibration_dict: dict[str, MotorCalibration]) -> None:
-        for motor, calibration in calibration_dict.items():
-            if self.protocol_version == 0:
-                self.write("Homing_Offset", motor, calibration.homing_offset)
-            self.write("Min_Position_Limit", motor, calibration.range_min)
-            self.write("Max_Position_Limit", motor, calibration.range_max)
-
-        self.calibration = calibration_dict
-
-    def _get_half_turn_homings(self, positions: dict[NameOrID, Value]) -> dict[NameOrID, Value]:
-        """
-        On Feetech Motors:
-        Present_Position = Actual_Position - Homing_Offset
-        """
-        half_turn_homings = {}
-        for motor, pos in positions.items():
-            model = self._get_motor_model(motor)
-            max_res = self.model_resolution_table[model] - 1
-            half_turn_homings[motor] = pos - int(max_res / 2)
-
-        return half_turn_homings
-
-    def disable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
-        for motor in self._get_motors_list(motors):
-            self.write("Torque_Enable", motor, TorqueMode.DISABLED.value, num_retry=num_retry)
-            self.write("Lock", motor, 0, num_retry=num_retry)
-
-    def _disable_torque(self, motor_id: int, model: str, num_retry: int = 0) -> None:
-        addr, length = get_address(self.model_ctrl_table, model, "Torque_Enable")
-        self._write(addr, length, motor_id, TorqueMode.DISABLED.value, num_retry=num_retry)
-        addr, length = get_address(self.model_ctrl_table, model, "Lock")
-        self._write(addr, length, motor_id, 0, num_retry=num_retry)
-
-    def enable_torque(self, motors: str | list[str] | None = None, num_retry: int = 0) -> None:
-        for motor in self._get_motors_list(motors):
-            self.write("Torque_Enable", motor, TorqueMode.ENABLED.value, num_retry=num_retry)
-            self.write("Lock", motor, 1, num_retry=num_retry)
-
-    def _encode_sign(self, data_name: str, ids_values: dict[int, int]) -> dict[int, int]:
-        for id_ in ids_values:
-            model = self._id_to_model(id_)
-            encoding_table = self.model_encoding_table.get(model)
-            if encoding_table and data_name in encoding_table:
-                sign_bit = encoding_table[data_name]
-                ids_values[id_] = encode_sign_magnitude(ids_values[id_], sign_bit)
-
-        return ids_values
-
-    def _decode_sign(self, data_name: str, ids_values: dict[int, int]) -> dict[int, int]:
-        for id_ in ids_values:
-            model = self._id_to_model(id_)
-            encoding_table = self.model_encoding_table.get(model)
-            if encoding_table and data_name in encoding_table:
-                sign_bit = encoding_table[data_name]
-                ids_values[id_] = decode_sign_magnitude(ids_values[id_], sign_bit)
-
-        return ids_values
-
-    def _split_into_byte_chunks(self, value: int, length: int) -> list[int]:
-        return _split_into_byte_chunks(value, length)
-
-    def _broadcast_ping(self) -> tuple[dict[int, int], int]:
-        import scservo_sdk as scs
-
-        data_list = {}
-
-        status_length = 6
-
-        rx_length = 0
-        wait_length = status_length * scs.MAX_ID
-
-        txpacket = [0] * 6
-
-        tx_time_per_byte = (1000.0 / self.port_handler.getBaudRate()) * 10.0
-
-        txpacket[scs.PKT_ID] = scs.BROADCAST_ID
-        txpacket[scs.PKT_LENGTH] = 2
-        txpacket[scs.PKT_INSTRUCTION] = scs.INST_PING
-
-        result = self.packet_handler.txPacket(self.port_handler, txpacket)
-        if result != scs.COMM_SUCCESS:
-            self.port_handler.is_using = False
-            return data_list, result
-
-        # set rx timeout
-        self.port_handler.setPacketTimeoutMillis((wait_length * tx_time_per_byte) + (3.0 * scs.MAX_ID) + 16.0)
-
-        rxpacket = []
-        while True:
-            rxpacket += self.port_handler.readPort(wait_length - rx_length)
-            rx_length = len(rxpacket)
-
-            if self.port_handler.isPacketTimeout():  # or rx_length >= wait_length
-                break
-
-        self.port_handler.is_using = False
-
-        if rx_length == 0:
-            return data_list, scs.COMM_RX_TIMEOUT
-
-        while True:
-            if rx_length < status_length:
-                return data_list, scs.COMM_RX_CORRUPT
-
-            # find packet header
-            for idx in range(0, (rx_length - 1)):
-                if (rxpacket[idx] == 0xFF) and (rxpacket[idx + 1] == 0xFF):
-                    break
-
-            if idx == 0:  # found at the beginning of the packet
-                # calculate checksum
-                checksum = 0
-                for idx in range(2, status_length - 1):  # except header & checksum
-                    checksum += rxpacket[idx]
-
-                checksum = ~checksum & 0xFF
-                if rxpacket[status_length - 1] == checksum:
-                    result = scs.COMM_SUCCESS
-                    data_list[rxpacket[scs.PKT_ID]] = rxpacket[scs.PKT_ERROR]
-
-                    del rxpacket[0:status_length]
-                    rx_length = rx_length - status_length
-
-                    if rx_length == 0:
-                        return data_list, result
-                else:
-                    result = scs.COMM_RX_CORRUPT
-                    # remove header (0xFF 0xFF)
-                    del rxpacket[0:2]
-                    rx_length = rx_length - 2
-            else:
-                # remove unnecessary packets
-                del rxpacket[0:idx]
-                rx_length = rx_length - idx
-
-    def broadcast_ping(self, num_retry: int = 0, raise_on_error: bool = False) -> dict[int, int] | None:
-        self._assert_protocol_is_compatible("broadcast_ping")
-        for n_try in range(1 + num_retry):
-            ids_status, comm = self._broadcast_ping()
-            if self._is_comm_success(comm):
-                break
-            logger.debug(f"Broadcast ping failed on port '{self.port}' ({n_try=})")
-            logger.debug(self.packet_handler.getTxRxResult(comm))
-
-        if not self._is_comm_success(comm):
-            if raise_on_error:
-                raise ConnectionError(self.packet_handler.getTxRxResult(comm))
-            return
-
-        ids_errors = {id_: status for id_, status in ids_status.items() if self._is_error(status)}
-        if ids_errors:
-            display_dict = {id_: self.packet_handler.getRxPacketError(err) for id_, err in ids_errors.items()}
-            logger.error(f"Some motors found returned an error status:\n{pformat(display_dict, indent=4)}")
-
-        return self._read_model_number(list(ids_status), raise_on_error)
-
-    def _read_firmware_version(self, motor_ids: list[int], raise_on_error: bool = False) -> dict[int, str]:
-        firmware_versions = {}
-        for id_ in motor_ids:
-            firm_ver_major, comm, error = self._read(
-                *FIRMWARE_MAJOR_VERSION, id_, raise_on_error=raise_on_error
-            )
-            if not self._is_comm_success(comm) or self._is_error(error):
-                return
-
-            firm_ver_minor, comm, error = self._read(
-                *FIRMWARE_MINOR_VERSION, id_, raise_on_error=raise_on_error
-            )
-            if not self._is_comm_success(comm) or self._is_error(error):
-                return
-
-            firmware_versions[id_] = f"{firm_ver_major}.{firm_ver_minor}"
-
-        return firmware_versions
-
-    def _read_model_number(self, motor_ids: list[int], raise_on_error: bool = False) -> dict[int, int]:
-        model_numbers = {}
-        for id_ in motor_ids:
-            model_nb, comm, error = self._read(*MODEL_NUMBER, id_, raise_on_error=raise_on_error)
-            if not self._is_comm_success(comm) or self._is_error(error):
-                return
-
-            model_numbers[id_] = model_nb
-
-        return model_numbers

lerobot/common/motors/feetech/tables.py

@@ -1,252 +0,0 @@
-# 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.
-
-FIRMWARE_MAJOR_VERSION = (0, 1)
-FIRMWARE_MINOR_VERSION = (1, 1)
-MODEL_NUMBER = (3, 2)
-
-# TODO(Steven): Consider doing the following:
-# from enum import Enum
-# class MyControlTableKey(Enum):
-#   ID = "ID"
-#   GOAL_SPEED = "Goal_Speed"
-#   ...
-#
-# MY_CONTROL_TABLE ={
-#   MyControlTableKey.ID.value: (5,1)
-#   MyControlTableKey.GOAL_SPEED.value: (46, 2)
-#   ...
-# }
-# This allows me do to:
-# bus.write(MyControlTableKey.GOAL_SPEED, ...)
-# Instead of:
-# bus.write("Goal_Speed", ...)
-# This is important for two reasons:
-# 1. The linter will tell me if I'm trying to use an invalid key, instead of me realizing when I get the RunTimeError
-# 2. We can change the value of the MyControlTableKey enums without impacting the client code
-
-# data_name: (address, size_byte)
-# http://doc.feetech.cn/#/prodinfodownload?srcType=FT-SMS-STS-emanual-229f4476422d4059abfb1cb0
-STS_SMS_SERIES_CONTROL_TABLE = {
-    # EPROM
-    "Firmware_Major_Version": FIRMWARE_MAJOR_VERSION,  # read-only
-    "Firmware_Minor_Version": FIRMWARE_MINOR_VERSION,  # read-only
-    "Model_Number": MODEL_NUMBER,  # read-only
-    "ID": (5, 1),
-    "Baud_Rate": (6, 1),
-    "Return_Delay_Time": (7, 1),
-    "Response_Status_Level": (8, 1),
-    "Min_Position_Limit": (9, 2),
-    "Max_Position_Limit": (11, 2),
-    "Max_Temperature_Limit": (13, 1),
-    "Max_Voltage_Limit": (14, 1),
-    "Min_Voltage_Limit": (15, 1),
-    "Max_Torque_Limit": (16, 2),
-    "Phase": (18, 1),
-    "Unloading_Condition": (19, 1),
-    "LED_Alarm_Condition": (20, 1),
-    "P_Coefficient": (21, 1),
-    "D_Coefficient": (22, 1),
-    "I_Coefficient": (23, 1),
-    "Minimum_Startup_Force": (24, 2),
-    "CW_Dead_Zone": (26, 1),
-    "CCW_Dead_Zone": (27, 1),
-    "Protection_Current": (28, 2),
-    "Angular_Resolution": (30, 1),
-    "Homing_Offset": (31, 2),
-    "Operating_Mode": (33, 1),
-    "Protective_Torque": (34, 1),
-    "Protection_Time": (35, 1),
-    "Overload_Torque": (36, 1),
-    "Velocity_closed_loop_P_proportional_coefficient": (37, 1),
-    "Over_Current_Protection_Time": (38, 1),
-    "Velocity_closed_loop_I_integral_coefficient": (39, 1),
-    # SRAM
-    "Torque_Enable": (40, 1),
-    "Acceleration": (41, 1),
-    "Goal_Position": (42, 2),
-    "Goal_Time": (44, 2),
-    "Goal_Velocity": (46, 2),
-    "Torque_Limit": (48, 2),
-    "Lock": (55, 1),
-    "Present_Position": (56, 2),  # read-only
-    "Present_Velocity": (58, 2),  # read-only
-    "Present_Load": (60, 2),  # read-only
-    "Present_Voltage": (62, 1),  # read-only
-    "Present_Temperature": (63, 1),  # read-only
-    "Status": (65, 1),  # read-only
-    "Moving": (66, 1),  # read-only
-    "Present_Current": (69, 2),  # read-only
-    "Goal_Position_2": (71, 2),  # read-only
-    # Factory
-    "Moving_Velocity": (80, 1),
-    "Moving_Velocity_Threshold": (80, 1),
-    "DTs": (81, 1),  # (ms)
-    "Velocity_Unit_factor": (82, 1),
-    "Hts": (83, 1),  # (ns) valid for firmware >= 2.54, other versions keep 0
-    "Maximum_Velocity_Limit": (84, 1),
-    "Maximum_Acceleration": (85, 1),
-    "Acceleration_Multiplier ": (86, 1),  # Acceleration multiplier in effect when acceleration is 0
-}
-
-# http://doc.feetech.cn/#/prodinfodownload?srcType=FT-SCSCL-emanual-cbcc8ab2e3384282a01d4bf3
-SCS_SERIES_CONTROL_TABLE = {
-    # EPROM
-    "Firmware_Major_Version": FIRMWARE_MAJOR_VERSION,  # read-only
-    "Firmware_Minor_Version": FIRMWARE_MINOR_VERSION,  # read-only
-    "Model_Number": MODEL_NUMBER,  # read-only
-    "ID": (5, 1),
-    "Baud_Rate": (6, 1),
-    "Return_Delay_Time": (7, 1),
-    "Response_Status_Level": (8, 1),
-    "Min_Position_Limit": (9, 2),
-    "Max_Position_Limit": (11, 2),
-    "Max_Temperature_Limit": (13, 1),
-    "Max_Voltage_Limit": (14, 1),
-    "Min_Voltage_Limit": (15, 1),
-    "Max_Torque_Limit": (16, 2),
-    "Phase": (18, 1),
-    "Unloading_Condition": (19, 1),
-    "LED_Alarm_Condition": (20, 1),
-    "P_Coefficient": (21, 1),
-    "D_Coefficient": (22, 1),
-    "I_Coefficient": (23, 1),
-    "Minimum_Startup_Force": (24, 2),
-    "CW_Dead_Zone": (26, 1),
-    "CCW_Dead_Zone": (27, 1),
-    "Protective_Torque": (37, 1),
-    "Protection_Time": (38, 1),
-    # SRAM
-    "Torque_Enable": (40, 1),
-    "Acceleration": (41, 1),
-    "Goal_Position": (42, 2),
-    "Running_Time": (44, 2),
-    "Goal_Velocity": (46, 2),
-    "Lock": (48, 1),
-    "Present_Position": (56, 2),  # read-only
-    "Present_Velocity": (58, 2),  # read-only
-    "Present_Load": (60, 2),  # read-only
-    "Present_Voltage": (62, 1),  # read-only
-    "Present_Temperature": (63, 1),  # read-only
-    "Sync_Write_Flag": (64, 1),  # read-only
-    "Status": (65, 1),  # read-only
-    "Moving": (66, 1),  # read-only
-    # Factory
-    "PWM_Maximum_Step": (78, 1),
-    "Moving_Velocity_Threshold*50": (79, 1),
-    "DTs": (80, 1),  # (ms)
-    "Minimum_Velocity_Limit*50": (81, 1),
-    "Maximum_Velocity_Limit*50": (82, 1),
-    "Acceleration_2": (83, 1),  # don't know what that is
-}
-
-STS_SMS_SERIES_BAUDRATE_TABLE = {
-    1_000_000: 0,
-    500_000: 1,
-    250_000: 2,
-    128_000: 3,
-    115_200: 4,
-    57_600: 5,
-    38_400: 6,
-    19_200: 7,
-}
-
-SCS_SERIES_BAUDRATE_TABLE = {
-    1_000_000: 0,
-    500_000: 1,
-    250_000: 2,
-    128_000: 3,
-    115_200: 4,
-    57_600: 5,
-    38_400: 6,
-    19_200: 7,
-}
-
-MODEL_CONTROL_TABLE = {
-    "sts_series": STS_SMS_SERIES_CONTROL_TABLE,
-    "scs_series": SCS_SERIES_CONTROL_TABLE,
-    "sms_series": STS_SMS_SERIES_CONTROL_TABLE,
-    "sts3215": STS_SMS_SERIES_CONTROL_TABLE,
-    "sts3250": STS_SMS_SERIES_CONTROL_TABLE,
-    "scs0009": SCS_SERIES_CONTROL_TABLE,
-    "sm8512bl": STS_SMS_SERIES_CONTROL_TABLE,
-}
-
-MODEL_RESOLUTION = {
-    "sts_series": 4096,
-    "sms_series": 4096,
-    "scs_series": 1024,
-    "sts3215": 4096,
-    "sts3250": 4096,
-    "sm8512bl": 65536,
-    "scs0009": 1024,
-}
-
-MODEL_BAUDRATE_TABLE = {
-    "sts_series": STS_SMS_SERIES_BAUDRATE_TABLE,
-    "sms_series": STS_SMS_SERIES_BAUDRATE_TABLE,
-    "scs_series": SCS_SERIES_BAUDRATE_TABLE,
-    "sm8512bl": STS_SMS_SERIES_BAUDRATE_TABLE,
-    "sts3215": STS_SMS_SERIES_BAUDRATE_TABLE,
-    "sts3250": STS_SMS_SERIES_BAUDRATE_TABLE,
-    "scs0009": SCS_SERIES_BAUDRATE_TABLE,
-}
-
-# Sign-Magnitude encoding bits
-STS_SMS_SERIES_ENCODINGS_TABLE = {
-    "Homing_Offset": 11,
-    "Goal_Velocity": 15,
-    "Present_Velocity": 15,
-}
-
-MODEL_ENCODING_TABLE = {
-    "sts_series": STS_SMS_SERIES_ENCODINGS_TABLE,
-    "sms_series": STS_SMS_SERIES_ENCODINGS_TABLE,
-    "scs_series": {},
-    "sts3215": STS_SMS_SERIES_ENCODINGS_TABLE,
-    "sts3250": STS_SMS_SERIES_ENCODINGS_TABLE,
-    "sm8512bl": STS_SMS_SERIES_ENCODINGS_TABLE,
-    "scs0009": {},
-}
-
-SCAN_BAUDRATES = [
-    4_800,
-    9_600,
-    14_400,
-    19_200,
-    38_400,
-    57_600,
-    115_200,
-    128_000,
-    250_000,
-    500_000,
-    1_000_000,
-]
-
-MODEL_NUMBER_TABLE = {
-    "sts3215": 777,
-    "sts3250": 2825,
-    "sm8512bl": 11272,
-    "scs0009": 1284,
-}
-
-MODEL_PROTOCOL = {
-    "sts_series": 0,
-    "sms_series": 0,
-    "scs_series": 1,
-    "sts3215": 0,
-    "sts3250": 0,
-    "sm8512bl": 0,
-    "scs0009": 1,
-}

lerobot/common/optim/optimizers.py

@@ -14,9 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import abc
-from dataclasses import asdict, dataclass, field
+from dataclasses import asdict, dataclass
 from pathlib import Path
-from typing import Any
 
 import draccus
 import torch
@@ -45,16 +44,7 @@ class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
         return "adam"
 
     @abc.abstractmethod
-    def build(self) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
-        """
-        Build the optimizer. It can be a single optimizer or a dictionary of optimizers.
-        NOTE: Multiple optimizers are useful when you have different models to optimize.
-        For example, you can have one optimizer for the policy and another one for the value function
-        in reinforcement learning settings.
-
-        Returns:
-            The optimizer or a dictionary of optimizers.
-        """
+    def build(self) -> torch.optim.Optimizer:
         raise NotImplementedError
 
 
@@ -104,76 +94,7 @@ class SGDConfig(OptimizerConfig):
         return torch.optim.SGD(params, **kwargs)
 
 
-@OptimizerConfig.register_subclass("multi_adam")
-@dataclass
-class MultiAdamConfig(OptimizerConfig):
-    """Configuration for multiple Adam optimizers with different parameter groups.
-
-    This creates a dictionary of Adam optimizers, each with its own hyperparameters.
-
-    Args:
-        lr: Default learning rate (used if not specified for a group)
-        weight_decay: Default weight decay (used if not specified for a group)
-        optimizer_groups: Dictionary mapping parameter group names to their hyperparameters
-        grad_clip_norm: Gradient clipping norm
-    """
-
-    lr: float = 1e-3
-    weight_decay: float = 0.0
-    grad_clip_norm: float = 10.0
-    optimizer_groups: dict[str, dict[str, Any]] = field(default_factory=dict)
-
-    def build(self, params_dict: dict[str, list]) -> dict[str, torch.optim.Optimizer]:
-        """Build multiple Adam optimizers.
-
-        Args:
-            params_dict: Dictionary mapping parameter group names to lists of parameters
-                         The keys should match the keys in optimizer_groups
-
-        Returns:
-            Dictionary mapping parameter group names to their optimizers
-        """
-        optimizers = {}
-
-        for name, params in params_dict.items():
-            # Get group-specific hyperparameters or use defaults
-            group_config = self.optimizer_groups.get(name, {})
-
-            # Create optimizer with merged parameters (defaults + group-specific)
-            optimizer_kwargs = {
-                "lr": group_config.get("lr", self.lr),
-                "betas": group_config.get("betas", (0.9, 0.999)),
-                "eps": group_config.get("eps", 1e-5),
-                "weight_decay": group_config.get("weight_decay", self.weight_decay),
-            }
-
-            optimizers[name] = torch.optim.Adam(params, **optimizer_kwargs)
-
-        return optimizers
-
-
-def save_optimizer_state(
-    optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer], save_dir: Path
-) -> None:
-    """Save optimizer state to disk.
-
-    Args:
-        optimizer: Either a single optimizer or a dictionary of optimizers.
-        save_dir: Directory to save the optimizer state.
-    """
-    if isinstance(optimizer, dict):
-        # Handle dictionary of optimizers
-        for name, opt in optimizer.items():
-            optimizer_dir = save_dir / name
-            optimizer_dir.mkdir(exist_ok=True, parents=True)
-            _save_single_optimizer_state(opt, optimizer_dir)
-    else:
-        # Handle single optimizer
-        _save_single_optimizer_state(optimizer, save_dir)
-
-
-def _save_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
-    """Save a single optimizer's state to disk."""
+def save_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> None:
     state = optimizer.state_dict()
     param_groups = state.pop("param_groups")
     flat_state = flatten_dict(state)
@@ -181,44 +102,11 @@ def _save_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Pat
     write_json(param_groups, save_dir / OPTIMIZER_PARAM_GROUPS)
 
 
-def load_optimizer_state(
-    optimizer: torch.optim.Optimizer | dict[str, torch.optim.Optimizer], save_dir: Path
-) -> torch.optim.Optimizer | dict[str, torch.optim.Optimizer]:
-    """Load optimizer state from disk.
-
-    Args:
-        optimizer: Either a single optimizer or a dictionary of optimizers.
-        save_dir: Directory to load the optimizer state from.
-
-    Returns:
-        The updated optimizer(s) with loaded state.
-    """
-    if isinstance(optimizer, dict):
-        # Handle dictionary of optimizers
-        loaded_optimizers = {}
-        for name, opt in optimizer.items():
-            optimizer_dir = save_dir / name
-            if optimizer_dir.exists():
-                loaded_optimizers[name] = _load_single_optimizer_state(opt, optimizer_dir)
-            else:
-                loaded_optimizers[name] = opt
-        return loaded_optimizers
-    else:
-        # Handle single optimizer
-        return _load_single_optimizer_state(optimizer, save_dir)
-
-
-def _load_single_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> torch.optim.Optimizer:
-    """Load a single optimizer's state from disk."""
+def load_optimizer_state(optimizer: torch.optim.Optimizer, save_dir: Path) -> torch.optim.Optimizer:
     current_state_dict = optimizer.state_dict()
     flat_state = load_file(save_dir / OPTIMIZER_STATE)
     state = unflatten_dict(flat_state)
-
-    # Handle case where 'state' key might not exist (for newly created optimizers)
-    if "state" in state:
-        loaded_state_dict = {"state": {int(k): v for k, v in state["state"].items()}}
-    else:
-        loaded_state_dict = {"state": {}}
+    loaded_state_dict = {"state": {int(k): v for k, v in state["state"].items()}}
 
     if "param_groups" in current_state_dict:
         param_groups = deserialize_json_into_object(

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -33,7 +33,7 @@ from diffusers.schedulers.scheduling_ddim import DDIMScheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 from torch import Tensor, nn
 
-from lerobot.common.constants import OBS_ENV_STATE, OBS_STATE
+from lerobot.common.constants import OBS_ENV, OBS_ROBOT
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pretrained import PreTrainedPolicy
@@ -238,8 +238,8 @@ class DiffusionModel(nn.Module):
 
     def _prepare_global_conditioning(self, batch: dict[str, Tensor]) -> Tensor:
         """Encode image features and concatenate them all together along with the state vector."""
-        batch_size, n_obs_steps = batch[OBS_STATE].shape[:2]
-        global_cond_feats = [batch[OBS_STATE]]
+        batch_size, n_obs_steps = batch[OBS_ROBOT].shape[:2]
+        global_cond_feats = [batch[OBS_ROBOT]]
         # Extract image features.
         if self.config.image_features:
             if self.config.use_separate_rgb_encoder_per_camera:
@@ -269,7 +269,7 @@ class DiffusionModel(nn.Module):
             global_cond_feats.append(img_features)
 
         if self.config.env_state_feature:
-            global_cond_feats.append(batch[OBS_ENV_STATE])
+            global_cond_feats.append(batch[OBS_ENV])
 
         # Concatenate features then flatten to (B, global_cond_dim).
         return torch.cat(global_cond_feats, dim=-1).flatten(start_dim=1)

lerobot/common/policies/factory.py

@@ -27,7 +27,6 @@ from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionC
 from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.common.policies.pretrained import PreTrainedPolicy
-from lerobot.common.policies.reward_model.configuration_classifier import RewardClassifierConfig
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
 from lerobot.configs.policies import PreTrainedConfig
@@ -60,14 +59,6 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
         from lerobot.common.policies.pi0fast.modeling_pi0fast import PI0FASTPolicy
 
         return PI0FASTPolicy
-    elif name == "sac":
-        from lerobot.common.policies.sac.modeling_sac import SACPolicy
-
-        return SACPolicy
-    elif name == "reward_classifier":
-        from lerobot.common.policies.reward_model.modeling_classifier import Classifier
-
-        return Classifier
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -85,8 +76,6 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return PI0Config(**kwargs)
     elif policy_type == "pi0fast":
         return PI0FASTConfig(**kwargs)
-    elif policy_type == "reward_classifier":
-        return RewardClassifierConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 

lerobot/common/policies/normalize.py

@@ -151,7 +151,6 @@ class Normalize(nn.Module):
     # TODO(rcadene): should we remove torch.no_grad?
     @torch.no_grad
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        # TODO: Remove this shallow copy
         batch = dict(batch)  # shallow copy avoids mutating the input batch
         for key, ft in self.features.items():
             if key not in batch:
@@ -253,168 +252,3 @@ class Unnormalize(nn.Module):
             else:
                 raise ValueError(norm_mode)
         return batch
-
-
-# TODO (azouitine): We should replace all normalization on the policies with register_buffer normalization
-#       and remove the `Normalize` and `Unnormalize` classes.
-def _initialize_stats_buffers(
-    module: nn.Module,
-    features: dict[str, PolicyFeature],
-    norm_map: dict[str, NormalizationMode],
-    stats: dict[str, dict[str, Tensor]] | None = None,
-) -> None:
-    """Register statistics buffers (mean/std or min/max) on the given *module*.
-
-    The logic matches the previous constructors of `NormalizeBuffer` and `UnnormalizeBuffer`,
-    but is factored out so it can be reused by both classes and stay in sync.
-    """
-    for key, ft in features.items():
-        norm_mode = norm_map.get(ft.type, NormalizationMode.IDENTITY)
-        if norm_mode is NormalizationMode.IDENTITY:
-            continue
-
-        shape: tuple[int, ...] = tuple(ft.shape)
-        if ft.type is FeatureType.VISUAL:
-            # reduce spatial dimensions, keep channel dimension only
-            c, *_ = shape
-            shape = (c, 1, 1)
-
-        prefix = key.replace(".", "_")
-
-        if norm_mode is NormalizationMode.MEAN_STD:
-            mean = torch.full(shape, torch.inf, dtype=torch.float32)
-            std = torch.full(shape, torch.inf, dtype=torch.float32)
-
-            if stats and key in stats and "mean" in stats[key] and "std" in stats[key]:
-                mean_data = stats[key]["mean"]
-                std_data = stats[key]["std"]
-                if isinstance(mean_data, torch.Tensor):
-                    # Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
-                    # tensors anywhere (for example, when we use the same stats for normalization and
-                    # unnormalization). See the logic here
-                    # https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
-                    mean = mean_data.clone().to(dtype=torch.float32)
-                    std = std_data.clone().to(dtype=torch.float32)
-                else:
-                    raise ValueError(f"Unsupported stats type for key '{key}' (expected ndarray or Tensor).")
-
-            module.register_buffer(f"{prefix}_mean", mean)
-            module.register_buffer(f"{prefix}_std", std)
-            continue
-
-        if norm_mode is NormalizationMode.MIN_MAX:
-            min_val = torch.full(shape, torch.inf, dtype=torch.float32)
-            max_val = torch.full(shape, torch.inf, dtype=torch.float32)
-
-            if stats and key in stats and "min" in stats[key] and "max" in stats[key]:
-                min_data = stats[key]["min"]
-                max_data = stats[key]["max"]
-                if isinstance(min_data, torch.Tensor):
-                    min_val = min_data.clone().to(dtype=torch.float32)
-                    max_val = max_data.clone().to(dtype=torch.float32)
-                else:
-                    raise ValueError(f"Unsupported stats type for key '{key}' (expected ndarray or Tensor).")
-
-            module.register_buffer(f"{prefix}_min", min_val)
-            module.register_buffer(f"{prefix}_max", max_val)
-            continue
-
-        raise ValueError(norm_mode)
-
-
-class NormalizeBuffer(nn.Module):
-    """Same as `Normalize` but statistics are stored as registered buffers rather than parameters."""
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-
-        _initialize_stats_buffers(self, features, norm_map, stats)
-
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        batch = dict(batch)
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            prefix = key.replace(".", "_")
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = getattr(self, f"{prefix}_mean")
-                std = getattr(self, f"{prefix}_std")
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = (batch[key] - mean) / (std + 1e-8)
-                continue
-
-            if norm_mode is NormalizationMode.MIN_MAX:
-                min_val = getattr(self, f"{prefix}_min")
-                max_val = getattr(self, f"{prefix}_max")
-                assert not torch.isinf(min_val).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max_val).any(), _no_stats_error_str("max")
-                batch[key] = (batch[key] - min_val) / (max_val - min_val + 1e-8)
-                batch[key] = batch[key] * 2 - 1
-                continue
-
-            raise ValueError(norm_mode)
-
-        return batch
-
-
-class UnnormalizeBuffer(nn.Module):
-    """Inverse operation of `NormalizeBuffer`. Uses registered buffers for statistics."""
-
-    def __init__(
-        self,
-        features: dict[str, PolicyFeature],
-        norm_map: dict[str, NormalizationMode],
-        stats: dict[str, dict[str, Tensor]] | None = None,
-    ):
-        super().__init__()
-        self.features = features
-        self.norm_map = norm_map
-
-        _initialize_stats_buffers(self, features, norm_map, stats)
-
-    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        # batch = dict(batch)
-        for key, ft in self.features.items():
-            if key not in batch:
-                continue
-
-            norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
-            if norm_mode is NormalizationMode.IDENTITY:
-                continue
-
-            prefix = key.replace(".", "_")
-
-            if norm_mode is NormalizationMode.MEAN_STD:
-                mean = getattr(self, f"{prefix}_mean")
-                std = getattr(self, f"{prefix}_std")
-                assert not torch.isinf(mean).any(), _no_stats_error_str("mean")
-                assert not torch.isinf(std).any(), _no_stats_error_str("std")
-                batch[key] = batch[key] * std + mean
-                continue
-
-            if norm_mode is NormalizationMode.MIN_MAX:
-                min_val = getattr(self, f"{prefix}_min")
-                max_val = getattr(self, f"{prefix}_max")
-                assert not torch.isinf(min_val).any(), _no_stats_error_str("min")
-                assert not torch.isinf(max_val).any(), _no_stats_error_str("max")
-                batch[key] = (batch[key] + 1) / 2
-                batch[key] = batch[key] * (max_val - min_val) + min_val
-                continue
-
-            raise ValueError(norm_mode)
-
-        return batch

lerobot/common/policies/pi0/modeling_pi0.py

@@ -57,7 +57,7 @@ import torch.nn.functional as F  # noqa: N812
 from torch import Tensor, nn
 from transformers import AutoTokenizer
 
-from lerobot.common.constants import ACTION, OBS_STATE
+from lerobot.common.constants import ACTION, OBS_ROBOT
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.common.policies.pi0.paligemma_with_expert import (
@@ -271,7 +271,7 @@ class PI0Policy(PreTrainedPolicy):
         self.eval()
 
         if self.config.adapt_to_pi_aloha:
-            batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
 
         batch = self.normalize_inputs(batch)
 
@@ -303,7 +303,7 @@ class PI0Policy(PreTrainedPolicy):
     def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> tuple[Tensor, dict[str, Tensor]]:
         """Do a full training forward pass to compute the loss"""
         if self.config.adapt_to_pi_aloha:
-            batch[OBS_STATE] = self._pi_aloha_decode_state(batch[OBS_STATE])
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
             batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
 
         batch = self.normalize_inputs(batch)
@@ -357,7 +357,7 @@ class PI0Policy(PreTrainedPolicy):
             if self.config.resize_imgs_with_padding is not None:
                 img = resize_with_pad(img, *self.config.resize_imgs_with_padding, pad_value=0)
 
-            # Normalize from range [0,1] to [-1,1] as expected by siglip
+            # Normalize from range [0,1] to [-1,1] as expacted by siglip
             img = img * 2.0 - 1.0
 
             bsize = img.shape[0]
@@ -380,7 +380,7 @@ class PI0Policy(PreTrainedPolicy):
 
     def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
         """Tokenize the text input"""
-        device = batch[OBS_STATE].device
+        device = batch[OBS_ROBOT].device
         tasks = batch["task"]
 
         # PaliGemma prompt has to end with a new line
@@ -427,7 +427,7 @@ class PI0Policy(PreTrainedPolicy):
 
     def prepare_state(self, batch):
         """Pad state"""
-        state = pad_vector(batch[OBS_STATE], self.config.max_state_dim)
+        state = pad_vector(batch[OBS_ROBOT], self.config.max_state_dim)
         return state
 
     def prepare_action(self, batch):

lerobot/common/policies/pi0fast/modeling_pi0fast.py

@@ -516,7 +516,7 @@ class PI0FAST(nn.Module):
                         interpolate_like_pi=self.config.interpolate_like_pi,
                     )
 
-                # Normalize from range [0,1] to [-1,1] as expected by siglip
+                # Normalize from range [0,1] to [-1,1] as expacted by siglip
                 img = img * 2.0 - 1.0
 
                 bsize = img.shape[0]

lerobot/common/policies/reward_model/configuration_classifier.py

@@ -1,62 +0,0 @@
-from dataclasses import dataclass, field
-from typing import List
-
-from lerobot.common.optim.optimizers import AdamWConfig, OptimizerConfig
-from lerobot.common.optim.schedulers import LRSchedulerConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
-
-
-@PreTrainedConfig.register_subclass(name="reward_classifier")
-@dataclass
-class RewardClassifierConfig(PreTrainedConfig):
-    """Configuration for the Reward Classifier model."""
-
-    name: str = "reward_classifier"
-    num_classes: int = 2
-    hidden_dim: int = 256
-    latent_dim: int = 256
-    image_embedding_pooling_dim: int = 8
-    dropout_rate: float = 0.1
-    model_name: str = "helper2424/resnet10"
-    device: str = "cpu"
-    model_type: str = "cnn"  # "transformer" or "cnn"
-    num_cameras: int = 2
-    learning_rate: float = 1e-4
-    weight_decay: float = 0.01
-    grad_clip_norm: float = 1.0
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.MEAN_STD,
-        }
-    )
-
-    @property
-    def observation_delta_indices(self) -> List | None:
-        return None
-
-    @property
-    def action_delta_indices(self) -> List | None:
-        return None
-
-    @property
-    def reward_delta_indices(self) -> List | None:
-        return None
-
-    def get_optimizer_preset(self) -> OptimizerConfig:
-        return AdamWConfig(
-            lr=self.learning_rate,
-            weight_decay=self.weight_decay,
-            grad_clip_norm=self.grad_clip_norm,
-        )
-
-    def get_scheduler_preset(self) -> LRSchedulerConfig | None:
-        return None
-
-    def validate_features(self) -> None:
-        """Validate feature configurations."""
-        has_image = any(key.startswith("observation.image") for key in self.input_features)
-        if not has_image:
-            raise ValueError(
-                "You must provide an image observation (key starting with 'observation.image') in the input features"
-            )

lerobot/common/policies/reward_model/modeling_classifier.py

@@ -1,301 +0,0 @@
-import logging
-from typing import Dict, Optional, Tuple
-
-import torch
-from torch import Tensor, nn
-
-from lerobot.common.constants import OBS_IMAGE
-from lerobot.common.policies.normalize import Normalize, Unnormalize
-from lerobot.common.policies.pretrained import PreTrainedPolicy
-from lerobot.common.policies.reward_model.configuration_classifier import RewardClassifierConfig
-
-
-class ClassifierOutput:
-    """Wrapper for classifier outputs with additional metadata."""
-
-    def __init__(
-        self,
-        logits: Tensor,
-        probabilities: Optional[Tensor] = None,
-        hidden_states: Optional[Tensor] = None,
-    ):
-        self.logits = logits
-        self.probabilities = probabilities
-        self.hidden_states = hidden_states
-
-    def __repr__(self):
-        return (
-            f"ClassifierOutput(logits={self.logits}, "
-            f"probabilities={self.probabilities}, "
-            f"hidden_states={self.hidden_states})"
-        )
-
-
-class SpatialLearnedEmbeddings(nn.Module):
-    def __init__(self, height, width, channel, num_features=8):
-        """
-        PyTorch implementation of learned spatial embeddings
-
-        Args:
-            height: Spatial height of input features
-            width: Spatial width of input features
-            channel: Number of input channels
-            num_features: Number of output embedding dimensions
-        """
-        super().__init__()
-        self.height = height
-        self.width = width
-        self.channel = channel
-        self.num_features = num_features
-
-        self.kernel = nn.Parameter(torch.empty(channel, height, width, num_features))
-
-        nn.init.kaiming_normal_(self.kernel, mode="fan_in", nonlinearity="linear")
-
-    def forward(self, features):
-        """
-        Forward pass for spatial embedding
-
-        Args:
-            features: Input tensor of shape [B, H, W, C] or [H, W, C] if no batch
-        Returns:
-            Output tensor of shape [B, C*F] or [C*F] if no batch
-        """
-
-        features = features.last_hidden_state
-
-        original_shape = features.shape
-        if features.dim() == 3:
-            features = features.unsqueeze(0)  # Add batch dim
-
-        features_expanded = features.unsqueeze(-1)  # [B, H, W, C, 1]
-        kernel_expanded = self.kernel.unsqueeze(0)  # [1, H, W, C, F]
-
-        # Element-wise multiplication and spatial reduction
-        output = (features_expanded * kernel_expanded).sum(dim=(2, 3))  # Sum H,W
-
-        # Reshape to combine channel and feature dimensions
-        output = output.view(output.size(0), -1)  # [B, C*F]
-
-        # Remove batch dim
-        if len(original_shape) == 3:
-            output = output.squeeze(0)
-
-        return output
-
-
-class Classifier(PreTrainedPolicy):
-    """Image classifier built on top of a pre-trained encoder."""
-
-    name = "reward_classifier"
-    config_class = RewardClassifierConfig
-
-    def __init__(
-        self,
-        config: RewardClassifierConfig,
-        dataset_stats: Dict[str, Dict[str, Tensor]] | None = None,
-    ):
-        from transformers import AutoModel
-
-        super().__init__(config)
-        self.config = config
-
-        # Initialize normalization (standardized with the policy framework)
-        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
-        self.normalize_targets = Normalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-        self.unnormalize_outputs = Unnormalize(
-            config.output_features, config.normalization_mapping, dataset_stats
-        )
-
-        # Set up encoder
-        encoder = AutoModel.from_pretrained(self.config.model_name, trust_remote_code=True)
-        # Extract vision model if we're given a multimodal model
-        if hasattr(encoder, "vision_model"):
-            logging.info("Multimodal model detected - using vision encoder only")
-            self.encoder = encoder.vision_model
-            self.vision_config = encoder.config.vision_config
-        else:
-            self.encoder = encoder
-            self.vision_config = getattr(encoder, "config", None)
-
-        # Model type from config
-        self.is_cnn = self.config.model_type == "cnn"
-
-        # For CNNs, initialize backbone
-        if self.is_cnn:
-            self._setup_cnn_backbone()
-
-        self._freeze_encoder()
-
-        # Extract image keys from input_features
-        self.image_keys = [
-            key.replace(".", "_") for key in config.input_features if key.startswith(OBS_IMAGE)
-        ]
-
-        if self.is_cnn:
-            self.encoders = nn.ModuleDict()
-            for image_key in self.image_keys:
-                encoder = self._create_single_encoder()
-                self.encoders[image_key] = encoder
-
-        self._build_classifier_head()
-
-    def _setup_cnn_backbone(self):
-        """Set up CNN encoder"""
-        if hasattr(self.encoder, "fc"):
-            self.feature_dim = self.encoder.fc.in_features
-            self.encoder = nn.Sequential(*list(self.encoder.children())[:-1])
-        elif hasattr(self.encoder.config, "hidden_sizes"):
-            self.feature_dim = self.encoder.config.hidden_sizes[-1]  # Last channel dimension
-        else:
-            raise ValueError("Unsupported CNN architecture")
-
-    def _freeze_encoder(self) -> None:
-        """Freeze the encoder parameters."""
-        for param in self.encoder.parameters():
-            param.requires_grad = False
-
-    def _create_single_encoder(self):
-        encoder = nn.Sequential(
-            self.encoder,
-            SpatialLearnedEmbeddings(
-                height=4,
-                width=4,
-                channel=self.feature_dim,
-                num_features=self.config.image_embedding_pooling_dim,
-            ),
-            nn.Dropout(self.config.dropout_rate),
-            nn.Linear(self.feature_dim * self.config.image_embedding_pooling_dim, self.config.latent_dim),
-            nn.LayerNorm(self.config.latent_dim),
-            nn.Tanh(),
-        )
-
-        return encoder
-
-    def _build_classifier_head(self) -> None:
-        """Initialize the classifier head architecture."""
-        # Get input dimension based on model type
-        if self.is_cnn:
-            input_dim = self.config.latent_dim
-        else:  # Transformer models
-            if hasattr(self.encoder.config, "hidden_size"):
-                input_dim = self.encoder.config.hidden_size
-            else:
-                raise ValueError("Unsupported transformer architecture since hidden_size is not found")
-
-        self.classifier_head = nn.Sequential(
-            nn.Linear(input_dim * self.config.num_cameras, self.config.hidden_dim),
-            nn.Dropout(self.config.dropout_rate),
-            nn.LayerNorm(self.config.hidden_dim),
-            nn.ReLU(),
-            nn.Linear(
-                self.config.hidden_dim,
-                1 if self.config.num_classes == 2 else self.config.num_classes,
-            ),
-        )
-
-    def _get_encoder_output(self, x: torch.Tensor, image_key: str) -> torch.Tensor:
-        """Extract the appropriate output from the encoder."""
-        with torch.no_grad():
-            if self.is_cnn:
-                # The HF ResNet applies pooling internally
-                outputs = self.encoders[image_key](x)
-                return outputs
-            else:  # Transformer models
-                outputs = self.encoder(x)
-                return outputs.last_hidden_state[:, 0, :]
-
-    def extract_images_and_labels(self, batch: Dict[str, Tensor]) -> Tuple[list, Tensor]:
-        """Extract image tensors and label tensors from batch."""
-        # Check for both OBS_IMAGE and OBS_IMAGES prefixes
-        images = [batch[key] for key in self.config.input_features if key.startswith(OBS_IMAGE)]
-        labels = batch["next.reward"]
-
-        return images, labels
-
-    def predict(self, xs: list) -> ClassifierOutput:
-        """Forward pass of the classifier for inference."""
-        encoder_outputs = torch.hstack(
-            [self._get_encoder_output(x, img_key) for x, img_key in zip(xs, self.image_keys, strict=True)]
-        )
-        logits = self.classifier_head(encoder_outputs)
-
-        if self.config.num_classes == 2:
-            logits = logits.squeeze(-1)
-            probabilities = torch.sigmoid(logits)
-        else:
-            probabilities = torch.softmax(logits, dim=-1)
-
-        return ClassifierOutput(logits=logits, probabilities=probabilities, hidden_states=encoder_outputs)
-
-    def forward(self, batch: Dict[str, Tensor]) -> Tuple[Tensor, Dict[str, Tensor]]:
-        """Standard forward pass for training compatible with train.py."""
-        # Normalize inputs if needed
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
-        # Extract images and labels
-        images, labels = self.extract_images_and_labels(batch)
-
-        # Get predictions
-        outputs = self.predict(images)
-
-        # Calculate loss
-        if self.config.num_classes == 2:
-            # Binary classification
-            loss = nn.functional.binary_cross_entropy_with_logits(outputs.logits, labels)
-            predictions = (torch.sigmoid(outputs.logits) > 0.5).float()
-        else:
-            # Multi-class classification
-            loss = nn.functional.cross_entropy(outputs.logits, labels.long())
-            predictions = torch.argmax(outputs.logits, dim=1)
-
-        # Calculate accuracy for logging
-        correct = (predictions == labels).sum().item()
-        total = labels.size(0)
-        accuracy = 100 * correct / total
-
-        # Return loss and metrics for logging
-        output_dict = {
-            "accuracy": accuracy,
-            "correct": correct,
-            "total": total,
-        }
-
-        return loss, output_dict
-
-    def predict_reward(self, batch, threshold=0.5):
-        """Eval method. Returns predicted reward with the decision threshold as argument."""
-        # Check for both OBS_IMAGE and OBS_IMAGES prefixes
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
-        # Extract images from batch dict
-        images = [batch[key] for key in self.config.input_features if key.startswith(OBS_IMAGE)]
-
-        if self.config.num_classes == 2:
-            probs = self.predict(images).probabilities
-            logging.debug(f"Predicted reward images: {probs}")
-            return (probs > threshold).float()
-        else:
-            return torch.argmax(self.predict(images).probabilities, dim=1)
-
-    def get_optim_params(self):
-        """Return optimizer parameters for the policy."""
-        return self.parameters()
-
-    def select_action(self, batch: Dict[str, Tensor]) -> Tensor:
-        """
-        This method is required by PreTrainedPolicy but not used for reward classifiers.
-        The reward classifier is not an actor and does not select actions.
-        """
-        raise NotImplementedError("Reward classifiers do not select actions")
-
-    def reset(self):
-        """
-        This method is required by PreTrainedPolicy but not used for reward classifiers.
-        The reward classifier is not an actor and does not select actions.
-        """
-        pass

lerobot/common/policies/sac/configuration_sac.py

@@ -1,243 +0,0 @@
-#!/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 dataclasses import dataclass, field
-
-from lerobot.common.optim.optimizers import MultiAdamConfig
-from lerobot.configs.policies import PreTrainedConfig
-from lerobot.configs.types import NormalizationMode
-
-
-def is_image_feature(key: str) -> bool:
-    """Check if a feature key represents an image feature.
-
-    Args:
-        key: The feature key to check
-
-    Returns:
-        True if the key represents an image feature, False otherwise
-    """
-    return key.startswith("observation.image")
-
-
-@dataclass
-class ConcurrencyConfig:
-    """Configuration for the concurrency of the actor and learner.
-    Possible values are:
-    - "threads": Use threads for the actor and learner.
-    - "processes": Use processes for the actor and learner.
-    """
-
-    actor: str = "threads"
-    learner: str = "threads"
-
-
-@dataclass
-class ActorLearnerConfig:
-    learner_host: str = "127.0.0.1"
-    learner_port: int = 50051
-    policy_parameters_push_frequency: int = 4
-
-
-@dataclass
-class CriticNetworkConfig:
-    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
-    activate_final: bool = True
-    final_activation: str | None = None
-
-
-@dataclass
-class ActorNetworkConfig:
-    hidden_dims: list[int] = field(default_factory=lambda: [256, 256])
-    activate_final: bool = True
-
-
-@dataclass
-class PolicyConfig:
-    use_tanh_squash: bool = True
-    log_std_min: float = 1e-5
-    log_std_max: float = 10.0
-    init_final: float = 0.05
-
-
-@PreTrainedConfig.register_subclass("sac")
-@dataclass
-class SACConfig(PreTrainedConfig):
-    """Soft Actor-Critic (SAC) configuration.
-
-    SAC is an off-policy actor-critic deep RL algorithm based on the maximum entropy
-    reinforcement learning framework. It learns a policy and a Q-function simultaneously
-    using experience collected from the environment.
-
-    This configuration class contains all the parameters needed to define a SAC agent,
-    including network architectures, optimization settings, and algorithm-specific
-    hyperparameters.
-    """
-
-    # Mapping of feature types to normalization modes
-    normalization_mapping: dict[str, NormalizationMode] = field(
-        default_factory=lambda: {
-            "VISUAL": NormalizationMode.MEAN_STD,
-            "STATE": NormalizationMode.MIN_MAX,
-            "ENV": NormalizationMode.MIN_MAX,
-            "ACTION": NormalizationMode.MIN_MAX,
-        }
-    )
-
-    # Statistics for normalizing different types of inputs
-    dataset_stats: dict[str, dict[str, list[float]]] | None = field(
-        default_factory=lambda: {
-            "observation.image": {
-                "mean": [0.485, 0.456, 0.406],
-                "std": [0.229, 0.224, 0.225],
-            },
-            "observation.state": {
-                "min": [0.0, 0.0],
-                "max": [1.0, 1.0],
-            },
-            "action": {
-                "min": [0.0, 0.0, 0.0],
-                "max": [1.0, 1.0, 1.0],
-            },
-        }
-    )
-
-    # Architecture specifics
-    # Device to run the model on (e.g., "cuda", "cpu")
-    device: str = "cpu"
-    # Device to store the model on
-    storage_device: str = "cpu"
-    # Name of the vision encoder model (Set to "helper2424/resnet10" for hil serl resnet10)
-    vision_encoder_name: str | None = None
-    # Whether to freeze the vision encoder during training
-    freeze_vision_encoder: bool = True
-    # Hidden dimension size for the image encoder
-    image_encoder_hidden_dim: int = 32
-    # Whether to use a shared encoder for actor and critic
-    shared_encoder: bool = True
-    # Number of discrete actions, eg for gripper actions
-    num_discrete_actions: int | None = None
-    # Dimension of the image embedding pooling
-    image_embedding_pooling_dim: int = 8
-
-    # Training parameter
-    # Number of steps for online training
-    online_steps: int = 1000000
-    # Seed for the online environment
-    online_env_seed: int = 10000
-    # Capacity of the online replay buffer
-    online_buffer_capacity: int = 100000
-    # Capacity of the offline replay buffer
-    offline_buffer_capacity: int = 100000
-    # Whether to use asynchronous prefetching for the buffers
-    async_prefetch: bool = False
-    # Number of steps before learning starts
-    online_step_before_learning: int = 100
-    # Frequency of policy updates
-    policy_update_freq: int = 1
-
-    # SAC algorithm parameters
-    # Discount factor for the SAC algorithm
-    discount: float = 0.99
-    # Initial temperature value
-    temperature_init: float = 1.0
-    # Number of critics in the ensemble
-    num_critics: int = 2
-    # Number of subsampled critics for training
-    num_subsample_critics: int | None = None
-    # Learning rate for the critic network
-    critic_lr: float = 3e-4
-    # Learning rate for the actor network
-    actor_lr: float = 3e-4
-    # Learning rate for the temperature parameter
-    temperature_lr: float = 3e-4
-    # Weight for the critic target update
-    critic_target_update_weight: float = 0.005
-    # Update-to-data ratio for the UTD algorithm (If you want enable utd_ratio, you need to set it to >1)
-    utd_ratio: int = 1
-    # Hidden dimension size for the state encoder
-    state_encoder_hidden_dim: int = 256
-    # Dimension of the latent space
-    latent_dim: int = 256
-    # Target entropy for the SAC algorithm
-    target_entropy: float | None = None
-    # Whether to use backup entropy for the SAC algorithm
-    use_backup_entropy: bool = True
-    # Gradient clipping norm for the SAC algorithm
-    grad_clip_norm: float = 40.0
-
-    # Network configuration
-    # Configuration for the critic network architecture
-    critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    # Configuration for the actor network architecture
-    actor_network_kwargs: ActorNetworkConfig = field(default_factory=ActorNetworkConfig)
-    # Configuration for the policy parameters
-    policy_kwargs: PolicyConfig = field(default_factory=PolicyConfig)
-    # Configuration for the discrete critic network
-    discrete_critic_network_kwargs: CriticNetworkConfig = field(default_factory=CriticNetworkConfig)
-    # Configuration for actor-learner architecture
-    actor_learner_config: ActorLearnerConfig = field(default_factory=ActorLearnerConfig)
-    # Configuration for concurrency settings (you can use threads or processes for the actor and learner)
-    concurrency: ConcurrencyConfig = field(default_factory=ConcurrencyConfig)
-
-    # Optimizations
-    use_torch_compile: bool = True
-
-    def __post_init__(self):
-        super().__post_init__()
-        # Any validation specific to SAC configuration
-
-    def get_optimizer_preset(self) -> MultiAdamConfig:
-        return MultiAdamConfig(
-            weight_decay=0.0,
-            optimizer_groups={
-                "actor": {"lr": self.actor_lr},
-                "critic": {"lr": self.critic_lr},
-                "temperature": {"lr": self.temperature_lr},
-            },
-        )
-
-    def get_scheduler_preset(self) -> None:
-        return None
-
-    def validate_features(self) -> None:
-        has_image = any(is_image_feature(key) for key in self.input_features)
-        has_state = "observation.state" in self.input_features
-
-        if not (has_state or has_image):
-            raise ValueError(
-                "You must provide either 'observation.state' or an image observation (key starting with 'observation.image') in the input features"
-            )
-
-        if "action" not in self.output_features:
-            raise ValueError("You must provide 'action' in the output features")
-
-    @property
-    def image_features(self) -> list[str]:
-        return [key for key in self.input_features if is_image_feature(key)]
-
-    @property
-    def observation_delta_indices(self) -> list:
-        return None
-
-    @property
-    def action_delta_indices(self) -> list:
-        return None  # SAC typically predicts one action at a time
-
-    @property
-    def reward_delta_indices(self) -> None:
-        return None