small tweaks

Co-authored-by: Kashif Rasul <kashif.rasul@gmail.com>
Co-authored-by: Alexander Soare <alexander.soare159@gmail.com>
Co-authored-by: Adil Zouitine <adilzouitinegm@gmail.com>
Co-authored-by: Cadene <re.cadene@gmail.com>

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,11 +1,15 @@
-# What does this PR do?
+## What this does
+Explain what this PR does. Feel free to tag your PR with the appropriate label(s).
 
 Examples:
-- Fixes # (issue)
-- Adds new dataset
-- Optimizes something
+|  Title               | Label           |
+|----------------------|-----------------|
+| Fixes #[issue]       | (🐛 Bug)        |
+| Adds new dataset     | (🗃️ Dataset)    |
+| Optimizes something  | (⚡️ Performance) |
 
-## How was it tested?
+## How it was tested
+Explain/show how you tested your changes.
 
 Examples:
 - Added `test_something` in `tests/test_stuff.py`.
@@ -13,6 +17,7 @@ Examples:
 - Optimized `some_function`, it now runs X times faster than previously.
 
 ## How to checkout & try? (for the reviewer)
+Provide a simple way for the reviewer to try out your changes.
 
 Examples:
 ```bash
@@ -22,11 +27,8 @@ DATA_DIR=tests/data pytest -sx tests/test_stuff.py::test_something
 python lerobot/scripts/train.py --some.option=true
 ```
 
-## Before submitting
-Please read the [contributor guideline](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md#submitting-a-pull-request-pr).
-
-
-## Who can review?
-
-Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
+## SECTION TO REMOVE BEFORE SUBMITTING YOUR PR
+**Note**: Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
 members/contributors who may be interested in your PR. Try to avoid tagging more than 3 people.
+
+**Note**: Before submitting this PR, please read the [contributor guideline](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md#submitting-a-pull-request-pr).

.github/scripts/dep_build.py

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

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

@@ -14,7 +14,7 @@ env:
 
 jobs:
   latest-cpu:
-    name: "Build CPU"
+    name: CPU
     runs-on: ubuntu-latest
     steps:
       - name: Cleanup disk
@@ -36,38 +36,6 @@ jobs:
       - name: Check out code
         uses: actions/checkout@v4
 
-      # HACK(aliberts): to be removed for release
-      # -----------------------------------------
-      - name: Checkout gym-aloha
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-aloha
-          path: envs/gym-aloha
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-xarm
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-xarm
-          path: envs/gym-xarm
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-pusht
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-pusht
-          path: envs/gym-pusht
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-
-      - name: Change envs dependencies as local path
-        run: python .github/scripts/dep_build.py
-      # -----------------------------------------
-
       - name: Login to DockerHub
         uses: docker/login-action@v3
         with:
@@ -109,7 +77,7 @@ jobs:
     #       SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
 
   latest-cuda:
-    name: "Build GPU"
+    name: GPU
     runs-on: ubuntu-latest
     steps:
       - name: Cleanup disk
@@ -130,38 +98,6 @@ jobs:
       - name: Check out code
         uses: actions/checkout@v4
 
-      # HACK(aliberts): to be removed for release
-      # -----------------------------------------
-      - name: Checkout gym-aloha
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-aloha
-          path: envs/gym-aloha
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-xarm
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-xarm
-          path: envs/gym-xarm
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-pusht
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-pusht
-          path: envs/gym-pusht
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-
-      - name: Change envs dependencies as local path
-        run: python .github/scripts/dep_build.py
-      # -----------------------------------------
-
       - name: Login to DockerHub
         uses: docker/login-action@v3
         with:

.github/workflows/nightly-tests.yml

@@ -13,7 +13,7 @@ env:
 
 jobs:
   run_all_tests_cpu:
-    name: "Test CPU"
+    name: CPU
     strategy:
       fail-fast: false
     runs-on: ubuntu-latest
@@ -40,7 +40,7 @@ jobs:
 
 
   run_all_tests_single_gpu:
-    name: "Test GPU"
+    name: GPU
     strategy:
       fail-fast: false
     runs-on: [single-gpu, nvidia-gpu, t4, ci]

.github/workflows/quality.yml

@@ -1,4 +1,4 @@
-name: Style
+name: Quality
 
 on:
   workflow_dispatch:
@@ -14,7 +14,8 @@ env:
   PYTHON_VERSION: "3.10"
 
 jobs:
-  ruff_check:
+  style:
+    name: Style
     runs-on: ubuntu-latest
     steps:
       - name: Checkout Repository
@@ -34,5 +35,22 @@ jobs:
       - name: Install Ruff
         run: python -m pip install "ruff==${{ env.RUFF_VERSION }}"
 
-      - name: Run Ruff
-        run: ruff check .
+      - name: Ruff check
+        run: ruff check
+
+      - name: Ruff format
+        run: ruff format --diff
+
+
+  poetry_check:
+    name: Poetry check
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout Repository
+        uses: actions/checkout@v3
+
+      - name: Install poetry
+        run: pipx install poetry
+
+      - name: Poetry check
+        run: poetry check

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

@@ -1,6 +1,6 @@
 # Inspired by
 # https://github.com/huggingface/peft/blob/main/.github/workflows/test-docker-build.yml
-name: Test Docker builds (PR)
+name: Test Dockerfiles
 
 on:
   pull_request:
@@ -15,7 +15,7 @@ env:
 
 jobs:
   get_changed_files:
-    name: "Get all modified Dockerfiles"
+    name: Detect modified Dockerfiles
     runs-on: ubuntu-latest
     outputs:
       matrix: ${{ steps.set-matrix.outputs.matrix }}
@@ -40,7 +40,7 @@ jobs:
 
 
   build_modified_dockerfiles:
-    name: "Build all modified Docker images"
+    name: Build modified Docker images
     needs: get_changed_files
     runs-on: ubuntu-latest
     if: ${{ needs.get_changed_files.outputs.matrix }} != ''
@@ -68,38 +68,6 @@ jobs:
       - name: Check out code
         uses: actions/checkout@v4
 
-      # HACK(aliberts): to be removed for release
-      # -----------------------------------------
-      - name: Checkout gym-aloha
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-aloha
-          path: envs/gym-aloha
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-xarm
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-xarm
-          path: envs/gym-xarm
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Checkout gym-pusht
-        uses: actions/checkout@v4
-        with:
-          repository: huggingface/gym-pusht
-          path: envs/gym-pusht
-          ssh-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
-      - name: Set up Python 3.10
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.10"
-
-      - name: Change envs dependencies as local path
-        run: python .github/scripts/dep_build.py
-      # -----------------------------------------
-
       - name: Build Docker image
         uses: docker/build-push-action@v5
         with:

.github/workflows/test.yml

@@ -21,29 +21,17 @@ on:
       - "poetry.lock"
 
 jobs:
-  tests:
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        os: [ubuntu-latest, macos-latest, macos-latest-large]
+  pytest:
+    name: Pytest
+    runs-on: ubuntu-latest
     env:
       DATA_DIR: tests/data
       MUJOCO_GL: egl
     steps:
-      - name: Add SSH key for installing envs
-        uses: webfactory/ssh-agent@v0.9.0
-        with:
-          ssh-private-key: ${{ secrets.SSH_PRIVATE_KEY }}
-
       - uses: actions/checkout@v4
 
       - name: Install EGL
-        run: |
-          if [[ "${{ matrix.os }}" == 'ubuntu-latest' ]]; then
-            sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
-          elif [[ "${{ matrix.os }}" == 'macos-latest' || "${{ matrix.os }}" == 'macos-latest-large' ]]; then
-            brew install mesa
-          fi
+        run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
 
       - name: Install poetry
         run: |
@@ -68,6 +56,34 @@ jobs:
             -W ignore::UserWarning:gymnasium.utils.env_checker:247 \
             && rm -rf tests/outputs outputs
 
+
+  end-to-end:
+    name: End-to-end
+    runs-on: ubuntu-latest
+    env:
+      DATA_DIR: tests/data
+      MUJOCO_GL: egl
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install EGL
+        run: sudo apt-get update && sudo apt-get install -y libegl1-mesa-dev
+
+      - name: Install poetry
+        run: |
+          pipx install poetry && poetry config virtualenvs.in-project true
+          echo "${{ github.workspace }}/.venv/bin" >> $GITHUB_PATH
+
+      - name: Set up Python 3.10
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+          cache: "poetry"
+
+      - name: Install poetry dependencies
+        run: |
+          poetry install --all-extras
+
       - name: Test end-to-end
         run: |
           make test-end-to-end \

.gitignore

@@ -6,6 +6,7 @@ data
 outputs
 .vscode
 rl
+.DS_Store
 
 # HPC
 nautilus/*.yaml

.pre-commit-config.yaml

@@ -18,7 +18,7 @@ repos:
     hooks:
     -   id: pyupgrade
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.4.2
+    rev: v0.4.3
     hooks:
       - id: ruff
         args: [--fix]

CONTRIBUTING.md

@@ -139,11 +139,13 @@ Follow these steps to start contributing:
 
    To develop on 🤗 LeRobot, you will at least need to install the `dev` and `test` extras dependencies along with the core library:
    ```bash
+   pip install poetry
    poetry install --sync --extras "dev test"
    ```
 
    You can also install the project with all its dependencies (including environments):
    ```bash
+   pip install poetry
    poetry install --sync --all-extras
    ```
 

Makefile

@@ -22,74 +22,90 @@ test-end-to-end:
 	${MAKE} test-act-ete-eval
 	${MAKE} test-diffusion-ete-train
 	${MAKE} test-diffusion-ete-eval
-	${MAKE} test-tdmpc-ete-train
-	${MAKE} test-tdmpc-ete-eval
+	# TODO(rcadene, alexander-soare): enable end-to-end tests for tdmpc
+	# ${MAKE} test-tdmpc-ete-train
+	# ${MAKE} test-tdmpc-ete-eval
+	${MAKE} test-default-ete-eval
 
 test-act-ete-train:
 	python lerobot/scripts/train.py \
 		policy=act \
 		env=aloha \
 		wandb.enable=False \
-		offline_steps=2 \
-		online_steps=0 \
-		eval_episodes=1 \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		device=cpu \
-		save_model=true \
-		save_freq=2 \
+		training.save_model=true \
+		training.save_freq=2 \
 		policy.n_action_steps=20 \
 		policy.chunk_size=20 \
-		policy.batch_size=2 \
+		training.batch_size=2 \
 		hydra.run.dir=tests/outputs/act/
 
 test-act-ete-eval:
 	python lerobot/scripts/eval.py \
-		--config tests/outputs/act/.hydra/config.yaml \
-		eval_episodes=1 \
+		-p tests/outputs/act/checkpoints/000002 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		env.episode_length=8 \
 		device=cpu \
-		policy.pretrained_model_path=tests/outputs/act/models/2.pt
 
 test-diffusion-ete-train:
 	python lerobot/scripts/train.py \
 		policy=diffusion \
 		env=pusht \
 		wandb.enable=False \
-		offline_steps=2 \
-		online_steps=0 \
-		eval_episodes=1 \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		device=cpu \
-		save_model=true \
-		save_freq=2 \
-		policy.batch_size=2 \
+		training.save_model=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
 		hydra.run.dir=tests/outputs/diffusion/
 
 test-diffusion-ete-eval:
 	python lerobot/scripts/eval.py \
-		--config tests/outputs/diffusion/.hydra/config.yaml \
-		eval_episodes=1 \
+		-p tests/outputs/diffusion/checkpoints/000002 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		env.episode_length=8 \
 		device=cpu \
-		policy.pretrained_model_path=tests/outputs/diffusion/models/2.pt
 
 test-tdmpc-ete-train:
 	python lerobot/scripts/train.py \
 		policy=tdmpc \
 		env=xarm \
+		env.task=XarmLift-v0 \
+		dataset_repo_id=lerobot/xarm_lift_medium_replay \
 		wandb.enable=False \
-		offline_steps=1 \
-		online_steps=2 \
-		eval_episodes=1 \
+		training.offline_steps=2 \
+		training.online_steps=2 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		env.episode_length=2 \
 		device=cpu \
-		save_model=true \
-		save_freq=2 \
-		policy.batch_size=2 \
+		training.save_model=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
 		hydra.run.dir=tests/outputs/tdmpc/
 
 test-tdmpc-ete-eval:
 	python lerobot/scripts/eval.py \
-		--config tests/outputs/tdmpc/.hydra/config.yaml \
-		eval_episodes=1 \
+		-p tests/outputs/tdmpc/checkpoints/000002 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=cpu \
+
+
+test-default-ete-eval:
+	python lerobot/scripts/eval.py \
+		--config lerobot/configs/default.yaml \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
 		env.episode_length=8 \
 		device=cpu \
-		policy.pretrained_model_path=tests/outputs/tdmpc/models/2.pt

README.md

@@ -29,15 +29,15 @@
 ---
 
 
-🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier for entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
+🤗 LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch. The goal is to lower the barrier to entry to robotics so that everyone can contribute and benefit from sharing datasets and pretrained models.
 
 🤗 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real-world with a focus on imitation learning and reinforcement learning.
 
-🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulated environments so that everyone can get started. In the coming weeks, the plan is to add more and more support for real-world robotics on the most affordable and capable robots out there.
+🤗 LeRobot already provides a set of pretrained models, datasets with human collected demonstrations, and simulation environments to get started without assembling a robot. In the coming weeks, the plan is to add more and more support for real-world robotics on the most affordable and capable robots out there.
 
-🤗 LeRobot hosts pretrained models and datasets on this HuggingFace community page: [huggingface.co/lerobot](https://huggingface.co/lerobot)
+🤗 LeRobot hosts pretrained models and datasets on this Hugging Face community page: [huggingface.co/lerobot](https://huggingface.co/lerobot)
 
-#### Examples of pretrained models and environments
+#### Examples of pretrained models on simulation environments
 
 <table>
   <tr>
@@ -54,10 +54,12 @@
 
 ### Acknowledgment
 
-- ACT policy and ALOHA environment are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha/)
-- Diffusion policy and Pusht environment are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu/)
-- TDMPC policy and Simxarm environment are adapted from [FOWM](https://www.yunhaifeng.com/FOWM/)
-- Abstractions and utilities for Reinforcement Learning come from [TorchRL](https://github.com/pytorch/rl)
+- Thanks to Tony Zaho, Zipeng Fu and colleagues for open sourcing ACT policy, ALOHA environments and datasets. Ours are adapted from [ALOHA](https://tonyzhaozh.github.io/aloha) and [Mobile ALOHA](https://mobile-aloha.github.io).
+- Thanks to Cheng Chi, Zhenjia Xu and colleagues for open sourcing Diffusion policy, Pusht environment and datasets, as well as UMI datasets. Ours are adapted from [Diffusion Policy](https://diffusion-policy.cs.columbia.edu) and [UMI Gripper](https://umi-gripper.github.io).
+- Thanks to Nicklas Hansen, Yunhai Feng and colleagues for open sourcing TDMPC policy, Simxarm environments and datasets. Ours are adapted from [TDMPC](https://github.com/nicklashansen/tdmpc) and [FOWM](https://www.yunhaifeng.com/FOWM).
+- Thanks to Vincent Moens and colleagues for open sourcing [TorchRL](https://github.com/pytorch/rl). It allowed for quick experimentations on the design of `LeRobot`.
+- Thanks to Antonio Loquercio and Ashish Kumar for their early support.
+
 
 ## Installation
 
@@ -86,7 +88,7 @@ For instance, to install 🤗 LeRobot with aloha and pusht, use:
 pip install ".[aloha, pusht]"
 ```
 
-To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiments tracking, log in with
+To use [Weights and Biases](https://docs.wandb.ai/quickstart) for experiment tracking, log in with
 ```bash
 wandb login
 ```
@@ -95,6 +97,7 @@ wandb login
 
 ```
 .
+├── examples             # contains demonstration examples, start here to learn about LeRobot
 ├── lerobot
 |   ├── configs          # contains hydra yaml files with all options that you can override in the command line
 |   |   ├── default.yaml   # selected by default, it loads pusht environment and diffusion policy
@@ -103,69 +106,84 @@ wandb login
 |   ├── common           # contains classes and utilities
 |   |   ├── datasets       # various datasets of human demonstrations: aloha, pusht, xarm
 |   |   ├── envs           # various sim environments: aloha, pusht, xarm
-|   |   └── policies       # various policies: act, diffusion, tdmpc
-|   └── scripts                  # contains functions to execute via command line
-|       ├── visualize_dataset.py  # load a dataset and render its demonstrations
-|       ├── eval.py               # load policy and evaluate it on an environment
-|       └── train.py              # train a policy via imitation learning and/or reinforcement learning
+|   |   ├── policies       # various policies: act, diffusion, tdmpc
+|   |   └── utils          # various utilities
+|   └── scripts          # contains functions to execute via command line
+|       ├── eval.py                 # load policy and evaluate it on an environment
+|       ├── train.py                # train a policy via imitation learning and/or reinforcement learning
+|       ├── push_dataset_to_hub.py  # convert your dataset into LeRobot dataset format and upload it to the Hugging Face hub
+|       └── visualize_dataset.py    # load a dataset and render its demonstrations
 ├── outputs               # contains results of scripts execution: logs, videos, model checkpoints
-├── .github
-|   └── workflows
-|       └── test.yml      # defines install settings for continuous integration and specifies end-to-end tests
 └── tests                 # contains pytest utilities for continuous integration
-
 ```
 
 ### Visualize datasets
 
-Check out [examples](./examples) to see how you can import our dataset class, download the data from the HuggingFace hub and use our rendering utilities.
+Check out [example 1](./examples/1_load_lerobot_dataset.py) that illustrates how to use our dataset class which automatically download data from the Hugging Face hub.
 
-Or you can achieve the same result by executing our script from the command line:
+You can also locally visualize episodes from a dataset by executing our script from the command line:
 ```bash
 python lerobot/scripts/visualize_dataset.py \
-env=pusht \
-hydra.run.dir=outputs/visualize_dataset/example
-# >>> ['outputs/visualize_dataset/example/episode_0.mp4']
+    --repo-id lerobot/pusht \
+    --episode-index 0
 ```
 
+It will open `rerun.io` and display the camera streams, robot states and actions, like this:
+
+https://github-production-user-asset-6210df.s3.amazonaws.com/4681518/328035972-fd46b787-b532-47e2-bb6f-fd536a55a7ed.mov?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAVCODYLSA53PQK4ZA%2F20240505%2Fus-east-1%2Fs3%2Faws4_request&X-Amz-Date=20240505T172924Z&X-Amz-Expires=300&X-Amz-Signature=d680b26c532eeaf80740f08af3320d22ad0b8a4e4da1bcc4f33142c15b509eda&X-Amz-SignedHeaders=host&actor_id=24889239&key_id=0&repo_id=748713144
+
+
+Our script can also visualize datasets stored on a distant server. See `python lerobot/scripts/visualize_dataset.py --help` for more instructions.
+
 ### Evaluate a pretrained policy
 
-Check out [examples](./examples) to see how you can load a pretrained policy from HuggingFace hub, load up the corresponding environment and model, and run an evaluation.
+Check out [example 2](./examples/2_evaluate_pretrained_policy.py) that illustrates how to download a pretrained policy from Hugging Face hub, and run an evaluation on its corresponding environment.
 
-Or you can achieve the same result by executing our script from the command line:
+We also provide a more capable script to parallelize the evaluation over multiple environments during the same rollout. Here is an example with a pretrained model hosted on [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht):
 ```bash
 python lerobot/scripts/eval.py \
---hub-id lerobot/diffusion_policy_pusht_image \
-eval_episodes=10 \
-hydra.run.dir=outputs/eval/example_hub
+    -p lerobot/diffusion_pusht \
+    eval.n_episodes=10 \
+    eval.batch_size=10
 ```
 
-After training your own policy, you can also re-evaluate the checkpoints with:
+Note: After training your own policy, you can re-evaluate the checkpoints with:
 ```bash
 python lerobot/scripts/eval.py \
---config PATH/TO/FOLDER/config.yaml \
-policy.pretrained_model_path=PATH/TO/FOLDER/weights.pth \
-eval_episodes=10 \
-hydra.run.dir=outputs/eval/example_dir
+    -p PATH/TO/TRAIN/OUTPUT/FOLDER
 ```
 
 See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-Check out [examples](./examples) to see how you can start training a model on a dataset, which will be automatically downloaded if needed.
+Check out [example 3](./examples/3_train_policy.py) that illustrates how to start training a model.
 
-In general, you can use our training script to easily train any policy on any environment:
+In general, you can use our training script to easily train any policy. To use wandb for logging training and evaluation curves, make sure you ran `wandb login`. Here is an example of training the ACT policy on trajectories collected by humans on the Aloha simulation environment for the insertion task:
 ```bash
 python lerobot/scripts/train.py \
-env=aloha \
-task=sim_insertion \
-repo_id=lerobot/aloha_sim_insertion_scripted \
-policy=act \
-hydra.run.dir=outputs/train/aloha_act
+    policy=act \
+    env=aloha \
+    env.task=AlohaInsertion-v0 \
+    dataset_repo_id=lerobot/aloha_sim_insertion_human
 ```
 
-After training, you may want to revisit model evaluation to change the evaluation settings. In fact, during training every checkpoint is already evaluated but on a low number of episodes for efficiency. Check out [example](./examples) to evaluate any model checkpoint on more episodes to increase statistical significance.
+The experiment directory is automatically generated and will show up in yellow in your terminal. It looks like `outputs/train/2024-05-05/20-21-12_aloha_act_default`. You can manually specify an experiment directory by adding this argument to the `train.py` python command:
+```bash
+    hydra.run.dir=your/new/experiment/dir
+```
+
+A link to the wandb logs for the run will also show up in yellow in your terminal. Here is an example of logs from wandb:
+![](media/wandb.png)
+
+You can deactivate wandb by adding these arguments to the `train.py` python command:
+```bash
+    wandb.disable_artifact=true \
+    wandb.enable=false
+```
+
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. After training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
+
 
 ## Contribute
 
@@ -173,105 +191,40 @@ If you would like to contribute to 🤗 LeRobot, please check out our [contribut
 
 ### Add a new dataset
 
-```python
-# TODO(rcadene, AdilZouitine): rewrite this section
-```
-
-To add a dataset to the hub, first login and use a token generated from [huggingface settings](https://huggingface.co/settings/tokens) with write access:
+To add a dataset to the hub, you need to login using a write-access token, which can be generated from the [Hugging Face settings](https://huggingface.co/settings/tokens):
 ```bash
 huggingface-cli login --token ${HUGGINGFACE_TOKEN} --add-to-git-credential
 ```
 
-Then you can upload it to the hub with:
+Then move your dataset folder in `data` directory (e.g. `data/aloha_ping_pong`), and push your dataset to the hub with:
 ```bash
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload $HF_USER/$DATASET data/$DATASET \
---repo-type dataset  \
---revision v1.0
+python lerobot/scripts/push_dataset_to_hub.py \
+--data-dir data \
+--dataset-id aloha_ping_ping \
+--raw-format aloha_hdf5 \
+--community-id lerobot
 ```
 
-You will need to set the corresponding version as a default argument in your dataset class:
-```python
-  version: str | None = "v1.1",
-```
-See: [`lerobot/common/datasets/pusht.py`](https://github.com/Cadene/lerobot/blob/main/lerobot/common/datasets/pusht.py)
+See `python lerobot/scripts/push_dataset_to_hub.py --help` for more instructions.
 
-For instance, for [lerobot/pusht](https://huggingface.co/datasets/lerobot/pusht), we used:
-```bash
-HF_USER=lerobot
-DATASET=pusht
-```
+If your dataset format is not supported, implement your own in `lerobot/common/datasets/push_dataset_to_hub/${raw_format}_format.py` by copying examples like [pusht_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py), [umi_zarr](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py), [aloha_hdf5](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py), or [xarm_pkl](https://github.com/huggingface/lerobot/blob/main/lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py).
 
-If you want to improve an existing dataset, you can download it locally with:
-```bash
-mkdir -p data/$DATASET
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli download ${HF_USER}/$DATASET \
---repo-type dataset \
---local-dir data/$DATASET \
---local-dir-use-symlinks=False \
---revision v1.0
-```
-
-Iterate on your code and dataset with:
-```bash
-DATA_DIR=data python train.py
-```
-
-Upload a new version (v2.0 or v1.1 if the changes are respectively more or less significant):
-```bash
-HF_HUB_ENABLE_HF_TRANSFER=1 huggingface-cli upload $HF_USER/$DATASET data/$DATASET \
---repo-type dataset \
---revision v1.1 \
---delete "*"
-```
-
-Then you will need to set the corresponding version as a default argument in your dataset class:
-```python
-  version: str | None = "v1.1",
-```
-See: [`lerobot/common/datasets/pusht.py`](https://github.com/Cadene/lerobot/blob/main/lerobot/common/datasets/pusht.py)
-
-
-Finally, you might want to mock the dataset if you need to update the unit tests as well:
-```bash
-python tests/scripts/mock_dataset.py --in-data-dir data/$DATASET --out-data-dir tests/data/$DATASET
-```
 
 ### Add a pretrained policy
 
-```python
-# TODO(rcadene, alexander-soare): rewrite this section
-```
+Once you have trained a policy you may upload it to the Hugging Face hub using a hub id that looks like `${hf_user}/${repo_name}` (e.g. [lerobot/diffusion_pusht](https://huggingface.co/lerobot/diffusion_pusht)).
 
-Once you have trained a policy you may upload it to the HuggingFace hub.
-
-Firstly, make sure you have a model repository set up on the hub. The hub ID looks like HF_USER/REPO_NAME.
-
-Secondly, assuming you have trained a policy, you need:
-
-- `config.yaml` which you can get from the `.hydra` directory of your training output folder.
-- `model.pt` which should be one of the saved models in the `models` directory of your training output folder (they won't be named `model.pt` but you will need to choose one).
-
-To upload these to the hub, prepare a folder with the following structure (you can use symlinks rather than copying):
-
-```
-to_upload
-    ├── config.yaml
-    └── model.pt
-```
-
-With the folder prepared, run the following with a desired revision ID.
+You first need to find the checkpoint located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). It should contain:
+- `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
+- `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
+- `config.yaml`: A consolidated Hydra training configuration containing the policy, environment, and dataset configs. The policy configuration should match `config.json` exactly. The environment config is useful for anyone who wants to evaluate your policy. The dataset config just serves as a paper trail for reproducibility.
 
+To upload these to the hub, run the following:
 ```bash
-huggingface-cli upload $HUB_ID to_upload --revision $REVISION_ID
+huggingface-cli upload ${hf_user}/${repo_name} path/to/checkpoint/dir
 ```
 
-If you want this to be the default revision also run the following (don't worry, it won't upload the files again; it will just adjust the file pointers):
-
-```bash
-huggingface-cli upload $HUB_ID to_upload
-```
-
-See `eval.py` for an example of how a user may use your policy.
+See [eval.py](https://github.com/huggingface/lerobot/blob/main/lerobot/scripts/eval.py) for an example of how other people may use your policy.
 
 
 ### Improve your code with profiling
@@ -298,9 +251,14 @@ with profile(
             # insert code to profile, potentially whole body of eval_policy function
 ```
 
-```bash
-python lerobot/scripts/eval.py \
---config outputs/pusht/.hydra/config.yaml \
-pretrained_model_path=outputs/pusht/model.pt \
-eval_episodes=7
+## Citation
+
+If you want, you can cite this work with:
+```
+@misc{cadene2024lerobot,
+    author = {Cadene, Remi and Alibert, Simon and Soare, Alexander and Gallouedec, Quentin and Zouitine, Adil and Wolf, Thomas},
+    title = {LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch},
+    howpublished = "\url{https://github.com/huggingface/lerobot}",
+    year = {2024}
+}
 ```

examples/1_load_lerobot_dataset.py

@@ -14,6 +14,7 @@ The script ends with examples of how to batch process data using PyTorch's DataL
 """
 
 from pathlib import Path
+from pprint import pprint
 
 import imageio
 import torch
@@ -21,48 +22,47 @@ import torch
 import lerobot
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
-print("List of available datasets", lerobot.available_datasets)
-# # >>> ['lerobot/aloha_sim_insertion_human', 'lerobot/aloha_sim_insertion_scripted',
-# #     'lerobot/aloha_sim_transfer_cube_human', 'lerobot/aloha_sim_transfer_cube_scripted',
-# #     'lerobot/pusht', 'lerobot/xarm_lift_medium']
+print("List of available datasets:")
+pprint(lerobot.available_datasets)
 
+# Let's take one for this example
 repo_id = "lerobot/pusht"
 
-# You can easily load a dataset from a Hugging Face repositery
+# You can easily load a dataset from a Hugging Face repository
 dataset = LeRobotDataset(repo_id)
 
-# LeRobotDataset is actually a thin wrapper around an underlying Hugging Face dataset  (see https://huggingface.co/docs/datasets/index for more information).
-# TODO(rcadene): update to make the print pretty
-print(f"{dataset=}")
-print(f"{dataset.hf_dataset=}")
+# LeRobotDataset is actually a thin wrapper around an underlying Hugging Face dataset
+# (see https://huggingface.co/docs/datasets/index for more information).
+print(dataset)
+print(dataset.hf_dataset)
 
-# and provides additional utilities for robotics and compatibility with pytorch
-print(f"number of samples/frames: {dataset.num_samples=}")
-print(f"number of episodes: {dataset.num_episodes=}")
-print(f"average number of frames per episode: {dataset.num_samples / dataset.num_episodes:.3f}")
+# And provides additional utilities for robotics and compatibility with Pytorch
+print(f"\naverage number of frames per episode: {dataset.num_samples / dataset.num_episodes:.3f}")
 print(f"frames per second used during data collection: {dataset.fps=}")
-print(f"keys to access images from cameras: {dataset.image_keys=}")
+print(f"keys to access images from cameras: {dataset.camera_keys=}\n")
 
-# While the LeRobotDataset adds helpers for working within our library, we still expose the underling Hugging Face dataset.
-# It may be freely replaced or modified in place. Here we use the filtering to keep only frames from episode 5.
-# TODO(rcadene): remove this example of accessing hf_dataset
-dataset.hf_dataset = dataset.hf_dataset.filter(lambda frame: frame["episode_index"] == 5)
+# Access frame indexes associated to first episode
+episode_index = 0
+from_idx = dataset.episode_data_index["from"][episode_index].item()
+to_idx = dataset.episode_data_index["to"][episode_index].item()
 
-# LeRobot datsets actually subclass PyTorch datasets. So you can do everything you know and love from working with the latter, for example: iterating through the dataset. Here we grab all the image frames.
-frames = [sample["observation.image"] for sample in dataset]
+# LeRobot datasets actually subclass PyTorch datasets so you can do everything you know and love from working
+# with the latter, like iterating through the dataset. Here we grab all the image frames.
+frames = [dataset[idx]["observation.image"] for idx in range(from_idx, to_idx)]
 
-# but frames are now float32 range [0,1] channel first (c,h,w) to follow pytorch convention,
-# to view them, we convert to uint8 range [0,255]
+# Video frames are now float32 in range [0,1] channel first (c,h,w) to follow pytorch convention. To visualize
+# them, we convert to uint8 in range [0,255]
 frames = [(frame * 255).type(torch.uint8) for frame in frames]
-# and to channel last (h,w,c)
+# and to channel last (h,w,c).
 frames = [frame.permute((1, 2, 0)).numpy() for frame in frames]
 
-# and finally save them to a mp4 video
+# Finally, we save the frames to a mp4 video for visualization.
 Path("outputs/examples/1_load_lerobot_dataset").mkdir(parents=True, exist_ok=True)
-imageio.mimsave("outputs/examples/1_load_lerobot_dataset/episode_5.mp4", frames, fps=dataset.fps)
+imageio.mimsave("outputs/examples/1_load_lerobot_dataset/episode_0.mp4", frames, fps=dataset.fps)
 
-# For many machine learning applications we need to load histories of past observations, or trajectorys of future actions. Our datasets can load previous and future frames for each key/modality,
-# using timestamps differences with the current loaded frame. For instance:
+# For many machine learning applications we need to load the history of past observations or trajectories of
+# future actions. Our datasets can load previous and future frames for each key/modality, using timestamps
+# differences with the current loaded frame. For instance:
 delta_timestamps = {
     # loads 4 images: 1 second before current frame, 500 ms before, 200 ms before, and current frame
     "observation.image": [-1, -0.5, -0.20, 0],
@@ -72,12 +72,12 @@ delta_timestamps = {
     "action": [t / dataset.fps for t in range(64)],
 }
 dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
-print(f"{dataset[0]['observation.image'].shape=}")  # (4,c,h,w)
+print(f"\n{dataset[0]['observation.image'].shape=}")  # (4,c,h,w)
 print(f"{dataset[0]['observation.state'].shape=}")  # (8,c)
-print(f"{dataset[0]['action'].shape=}")  # (64,c)
+print(f"{dataset[0]['action'].shape=}\n")  # (64,c)
 
-# Finally, our datasets are fully compatible with PyTorch dataloaders and samplers
-# because they are just PyTorch datasets.
+# Finally, our datasets are fully compatible with PyTorch dataloaders and samplers because they are just
+# PyTorch datasets.
 dataloader = torch.utils.data.DataLoader(
     dataset,
     num_workers=0,

examples/2_evaluate_pretrained_policy.py

@@ -5,34 +5,108 @@ training outputs directory. In the latter case, you might want to run examples/3
 
 from pathlib import Path
 
+import gym_pusht  # noqa: F401
+import gymnasium as gym
+import imageio
+import numpy
+import torch
 from huggingface_hub import snapshot_download
 
-from lerobot.common.utils.utils import init_hydra_config
-from lerobot.scripts.eval import eval
+from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 
-# Get a pretrained policy from the hub.
-# TODO(alexander-soare): This no longer works until we upload a new model that uses the current configs.
-hub_id = "lerobot/diffusion_policy_pusht_image"
-folder = Path(snapshot_download(hub_id))
+# Create a directory to store the video of the evaluation
+output_directory = Path("outputs/eval/example_pusht_diffusion")
+output_directory.mkdir(parents=True, exist_ok=True)
+
+device = torch.device("cuda")
+
+# Download the diffusion policy for pusht environment
+pretrained_policy_path = Path(snapshot_download("lerobot/diffusion_pusht"))
 # OR uncomment the following to evaluate a policy from the local outputs/train folder.
-# folder = Path("outputs/train/example_pusht_diffusion")
+# pretrained_policy_path = Path("outputs/train/example_pusht_diffusion")
 
-config_path = folder / "config.yaml"
-weights_path = folder / "model.pt"
+policy = DiffusionPolicy.from_pretrained(pretrained_policy_path)
+policy.eval()
+policy.to(device)
 
-# Override some config parameters to do with evaluation.
-overrides = [
-    f"policy.pretrained_model_path={weights_path}",
-    "eval_episodes=10",
-    "rollout_batch_size=10",
-    "device=cuda",
-]
-
-# Create a Hydra config.
-cfg = init_hydra_config(config_path, overrides)
-
-# Evaluate the policy and save the outputs including metrics and videos.
-eval(
-    cfg,
-    out_dir=f"outputs/eval/example_{cfg.env.name}_{cfg.policy.name}",
+# Initialize evaluation environment to render two observation types:
+# an image of the scene and state/position of the agent. The environment
+# also automatically stops running after 300 interactions/steps.
+env = gym.make(
+    "gym_pusht/PushT-v0",
+    obs_type="pixels_agent_pos",
+    max_episode_steps=300,
 )
+
+# Reset the policy and environmens to prepare for rollout
+policy.reset()
+numpy_observation, info = env.reset(seed=42)
+
+# Prepare to collect every rewards and all the frames of the episode,
+# from initial state to final state.
+rewards = []
+frames = []
+
+# Render frame of the initial state
+frames.append(env.render())
+
+step = 0
+done = False
+while not done:
+    # Prepare observation for the policy running in Pytorch
+    state = torch.from_numpy(numpy_observation["agent_pos"])
+    image = torch.from_numpy(numpy_observation["pixels"])
+
+    # Convert to float32 with image from channel first in [0,255]
+    # to channel last in [0,1]
+    state = state.to(torch.float32)
+    image = image.to(torch.float32) / 255
+    image = image.permute(2, 0, 1)
+
+    # Send data tensors from CPU to GPU
+    state = state.to(device, non_blocking=True)
+    image = image.to(device, non_blocking=True)
+
+    # Add extra (empty) batch dimension, required to forward the policy
+    state = state.unsqueeze(0)
+    image = image.unsqueeze(0)
+
+    # Create the policy input dictionary
+    observation = {
+        "observation.state": state,
+        "observation.image": image,
+    }
+
+    # Predict the next action with respect to the current observation
+    with torch.inference_mode():
+        action = policy.select_action(observation)
+
+    # Prepare the action for the environment
+    numpy_action = action.squeeze(0).to("cpu").numpy()
+
+    # Step through the environment and receive a new observation
+    numpy_observation, reward, terminated, truncated, info = env.step(numpy_action)
+    print(f"{step=} {reward=} {terminated=}")
+
+    # Keep track of all the rewards and frames
+    rewards.append(reward)
+    frames.append(env.render())
+
+    # The rollout is considered done when the success state is reach (i.e. terminated is True),
+    # or the maximum number of iterations is reached (i.e. truncated is True)
+    done = terminated | truncated | done
+    step += 1
+
+if terminated:
+    print("Success!")
+else:
+    print("Failure!")
+
+# Get the speed of environment (i.e. its number of frames per second).
+fps = env.metadata["render_fps"]
+
+# Encode all frames into a mp4 video.
+video_path = output_directory / "rollout.mp4"
+imageio.mimsave(str(video_path), numpy.stack(frames), fps=fps)
+
+print(f"Video of the evaluation is available in '{video_path}'.")

examples/3_train_policy.py

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

lerobot/__init__.py

@@ -9,6 +9,7 @@ Example:
         print(lerobot.available_tasks_per_env)
         print(lerobot.available_datasets)
         print(lerobot.available_datasets_per_env)
+        print(lerobot.available_real_world_datasets)
         print(lerobot.available_policies)
         print(lerobot.available_policies_per_env)
     ```
@@ -55,17 +56,33 @@ available_datasets_per_env = {
     ],
 }
 
-available_datasets_without_env = ["lerobot/umi_cup_in_the_wild"]
+available_real_world_datasets = [
+    "lerobot/aloha_mobile_cabinet",
+    "lerobot/aloha_mobile_chair",
+    "lerobot/aloha_mobile_elevator",
+    "lerobot/aloha_mobile_shrimp",
+    "lerobot/aloha_mobile_wash_pan",
+    "lerobot/aloha_mobile_wipe_wine",
+    "lerobot/aloha_static_battery",
+    "lerobot/aloha_static_candy",
+    "lerobot/aloha_static_coffee",
+    "lerobot/aloha_static_coffee_new",
+    "lerobot/aloha_static_cups_open",
+    "lerobot/aloha_static_fork_pick_up",
+    "lerobot/aloha_static_pingpong_test",
+    "lerobot/aloha_static_pro_pencil",
+    "lerobot/aloha_static_screw_driver",
+    "lerobot/aloha_static_tape",
+    "lerobot/aloha_static_thread_velcro",
+    "lerobot/aloha_static_towel",
+    "lerobot/aloha_static_vinh_cup",
+    "lerobot/aloha_static_vinh_cup_left",
+    "lerobot/aloha_static_ziploc_slide",
+    "lerobot/umi_cup_in_the_wild",
+]
 
 available_datasets = list(
-    itertools.chain(*available_datasets_per_env.values(), available_datasets_without_env)
-)
-
-# TODO(rcadene, aliberts, alexander-soare): Add real-world env with a gym API
-available_datasets_without_env = ["lerobot/umi_cup_in_the_wild"]
-
-available_datasets = list(
-    itertools.chain(*available_datasets_per_env.values(), available_datasets_without_env)
+    itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets)
 )
 
 available_policies = [

lerobot/common/datasets/_video_benchmark/README.md

@@ -0,0 +1,334 @@
+# Video benchmark
+
+
+## Questions
+
+What is the optimal trade-off between:
+- maximizing loading time with random access,
+- minimizing memory space on disk,
+- maximizing success rate of policies?
+
+How to encode videos?
+- How much compression (`-crf`)? Low compression with `0`, normal compression with `20` or extreme with `56`?
+- What pixel format to use (`-pix_fmt`)? `yuv444p` or `yuv420p`?
+- How many key frames (`-g`)? A key frame every `10` frames?
+
+How to decode videos?
+- Which `decoder`? `torchvision`, `torchaudio`, `ffmpegio`, `decord`, or `nvc`?
+
+## Metrics
+
+**Percentage of data compression (higher is better)**
+`compression_factor` is the ratio of the memory space on disk taken by the original images to encode, to the memory space taken by the encoded video. For instance, `compression_factor=4` means that the video takes 4 times less memory space on disk compared to the original images.
+
+**Percentage of loading time (higher is better)**
+`load_time_factor` is the ratio of the time it takes to load original images at given timestamps, to the time it takes to decode the exact same frames from the video. Higher is better. For instance, `load_time_factor=0.5` means that decoding from video is 2 times slower than loading the original images.
+
+**Average L2 error per pixel (lower is better)**
+`avg_per_pixel_l2_error` is the average L2 error between each decoded frame and its corresponding original image over all requested timestamps, and also divided by the number of pixels in the image to be comparable when switching to different image sizes.
+
+**Loss of a pretrained policy (higher is better)** (not available)
+`loss_pretrained` is the result of evaluating with the selected encoding/decoding settings a policy pretrained on original images. It is easier to understand than `avg_l2_error`.
+
+**Success rate after retraining (higher is better)** (not available)
+`success_rate` is the result of training and evaluating a policy with the selected encoding/decoding settings. It is the most difficult metric to get but also the very best.
+
+
+## Variables
+
+**Image content**
+We don't expect the same optimal settings for a dataset of images from a simulation, or from real-world in an appartment, or in a factory, or outdoor, etc. Hence, we run this benchmark on two datasets: `pusht` (simulation) and `umi` (real-world outdoor).
+
+**Requested timestamps**
+In this benchmark, we focus on the loading time of random access, so we are not interested in sequentially loading all frames of a video like in a movie. However, the number of consecutive timestamps requested and their spacing can greatly affect the `load_time_factor`. In fact, it is expected to get faster loading time by decoding a large number of consecutive frames from a video, than to load the same data from individual images. To reflect our robotics use case, we consider a few settings:
+- `single_frame`: 1 frame,
+- `2_frames`: 2 consecutive frames (e.g. `[t, t + 1 / fps]`),
+- `2_frames_4_space`: 2 consecutive frames with 4 frames of spacing (e.g `[t, t + 4 / fps]`),
+
+**Data augmentations**
+We might revisit this benchmark and find better settings if we train our policies with various data augmentations to make them more robust (e.g. robust to color changes, compression, etc.).
+
+
+## Results
+
+**`decoder`**
+| repo_id | decoder | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | <span style="color: #32CD32;">torchvision</span> | 0.166 | 0.0000119 |
+| lerobot/pusht | ffmpegio | 0.009 | 0.0001182 |
+| lerobot/pusht | torchaudio | 0.138 | 0.0000359 |
+| lerobot/umi_cup_in_the_wild | <span style="color: #32CD32;">torchvision</span> | 0.174 | 0.0000174 |
+| lerobot/umi_cup_in_the_wild | ffmpegio | 0.010 | 0.0000735 |
+| lerobot/umi_cup_in_the_wild | torchaudio | 0.154 | 0.0000340 |
+
+### `1_frame`
+
+**`pix_fmt`**
+| repo_id | pix_fmt | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | yuv420p | 3.788 | 0.224 | 0.0000760 |
+| lerobot/pusht | yuv444p | 3.646 | 0.185 | 0.0000443 |
+| lerobot/umi_cup_in_the_wild | yuv420p | 14.391 | 0.388 | 0.0000469 |
+| lerobot/umi_cup_in_the_wild | yuv444p | 14.932 | 0.329 | 0.0000397 |
+
+**`g`**
+| repo_id | g | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 1 | 2.543 | 0.204 | 0.0000556 |
+| lerobot/pusht | 2 | 3.646 | 0.182 | 0.0000443 |
+| lerobot/pusht | 3 | 4.431 | 0.174 | 0.0000450 |
+| lerobot/pusht | 4 | 5.103 | 0.163 | 0.0000448 |
+| lerobot/pusht | 5 | 5.625 | 0.163 | 0.0000436 |
+| lerobot/pusht | 6 | 5.974 | 0.155 | 0.0000427 |
+| lerobot/pusht | 10 | 6.814 | 0.130 | 0.0000410 |
+| lerobot/pusht | 15 | 7.431 | 0.105 | 0.0000406 |
+| lerobot/pusht | 20 | 7.662 | 0.097 | 0.0000400 |
+| lerobot/pusht | 40 | 8.163 | 0.061 | 0.0000405 |
+| lerobot/pusht | 100 | 8.761 | 0.039 | 0.0000422 |
+| lerobot/pusht | None | 8.909 | 0.024 | 0.0000431 |
+| lerobot/umi_cup_in_the_wild | 1 | 14.411 | 0.444 | 0.0000601 |
+| lerobot/umi_cup_in_the_wild | 2 | 14.932 | 0.345 | 0.0000397 |
+| lerobot/umi_cup_in_the_wild | 3 | 20.174 | 0.282 | 0.0000416 |
+| lerobot/umi_cup_in_the_wild | 4 | 24.889 | 0.271 | 0.0000415 |
+| lerobot/umi_cup_in_the_wild | 5 | 28.825 | 0.260 | 0.0000415 |
+| lerobot/umi_cup_in_the_wild | 6 | 31.635 | 0.249 | 0.0000415 |
+| lerobot/umi_cup_in_the_wild | 10 | 39.418 | 0.195 | 0.0000399 |
+| lerobot/umi_cup_in_the_wild | 15 | 44.577 | 0.169 | 0.0000394 |
+| lerobot/umi_cup_in_the_wild | 20 | 47.907 | 0.140 | 0.0000390 |
+| lerobot/umi_cup_in_the_wild | 40 | 52.554 | 0.096 | 0.0000384 |
+| lerobot/umi_cup_in_the_wild | 100 | 58.241 | 0.046 | 0.0000390 |
+| lerobot/umi_cup_in_the_wild | None | 60.530 | 0.022 | 0.0000400 |
+
+**`crf`**
+| repo_id | crf | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 0 | 1.699 | 0.175 | 0.0000035 |
+| lerobot/pusht | 5 | 1.409 | 0.181 | 0.0000080 |
+| lerobot/pusht | 10 | 1.842 | 0.172 | 0.0000123 |
+| lerobot/pusht | 15 | 2.322 | 0.187 | 0.0000211 |
+| lerobot/pusht | 20 | 3.050 | 0.181 | 0.0000346 |
+| lerobot/pusht | None | 3.646 | 0.189 | 0.0000443 |
+| lerobot/pusht | 25 | 3.969 | 0.186 | 0.0000521 |
+| lerobot/pusht | 30 | 5.687 | 0.184 | 0.0000850 |
+| lerobot/pusht | 40 | 10.818 | 0.193 | 0.0001726 |
+| lerobot/pusht | 50 | 18.185 | 0.183 | 0.0002606 |
+| lerobot/umi_cup_in_the_wild | 0 | 1.918 | 0.165 | 0.0000056 |
+| lerobot/umi_cup_in_the_wild | 5 | 3.207 | 0.171 | 0.0000111 |
+| lerobot/umi_cup_in_the_wild | 10 | 4.818 | 0.212 | 0.0000153 |
+| lerobot/umi_cup_in_the_wild | 15 | 7.329 | 0.261 | 0.0000218 |
+| lerobot/umi_cup_in_the_wild | 20 | 11.361 | 0.312 | 0.0000317 |
+| lerobot/umi_cup_in_the_wild | None | 14.932 | 0.339 | 0.0000397 |
+| lerobot/umi_cup_in_the_wild | 25 | 17.741 | 0.297 | 0.0000452 |
+| lerobot/umi_cup_in_the_wild | 30 | 27.983 | 0.406 | 0.0000629 |
+| lerobot/umi_cup_in_the_wild | 40 | 82.449 | 0.468 | 0.0001184 |
+| lerobot/umi_cup_in_the_wild | 50 | 186.145 | 0.515 | 0.0001879 |
+
+**best**
+| repo_id | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | 3.646 | 0.188 | 0.0000443 |
+| lerobot/umi_cup_in_the_wild | 14.932 | 0.339 | 0.0000397 |
+
+### `2_frames`
+
+**`pix_fmt`**
+| repo_id | pix_fmt | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | yuv420p | 3.788 | 0.314 | 0.0000799 |
+| lerobot/pusht | yuv444p | 3.646 | 0.303 | 0.0000496 |
+| lerobot/umi_cup_in_the_wild | yuv420p | 14.391 | 0.642 | 0.0000503 |
+| lerobot/umi_cup_in_the_wild | yuv444p | 14.932 | 0.529 | 0.0000436 |
+
+**`g`**
+| repo_id | g | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 1 | 2.543 | 0.308 | 0.0000599 |
+| lerobot/pusht | 2 | 3.646 | 0.279 | 0.0000496 |
+| lerobot/pusht | 3 | 4.431 | 0.259 | 0.0000498 |
+| lerobot/pusht | 4 | 5.103 | 0.243 | 0.0000501 |
+| lerobot/pusht | 5 | 5.625 | 0.235 | 0.0000492 |
+| lerobot/pusht | 6 | 5.974 | 0.230 | 0.0000481 |
+| lerobot/pusht | 10 | 6.814 | 0.194 | 0.0000468 |
+| lerobot/pusht | 15 | 7.431 | 0.152 | 0.0000460 |
+| lerobot/pusht | 20 | 7.662 | 0.151 | 0.0000455 |
+| lerobot/pusht | 40 | 8.163 | 0.095 | 0.0000454 |
+| lerobot/pusht | 100 | 8.761 | 0.062 | 0.0000472 |
+| lerobot/pusht | None | 8.909 | 0.037 | 0.0000479 |
+| lerobot/umi_cup_in_the_wild | 1 | 14.411 | 0.638 | 0.0000625 |
+| lerobot/umi_cup_in_the_wild | 2 | 14.932 | 0.537 | 0.0000436 |
+| lerobot/umi_cup_in_the_wild | 3 | 20.174 | 0.493 | 0.0000437 |
+| lerobot/umi_cup_in_the_wild | 4 | 24.889 | 0.458 | 0.0000446 |
+| lerobot/umi_cup_in_the_wild | 5 | 28.825 | 0.438 | 0.0000445 |
+| lerobot/umi_cup_in_the_wild | 6 | 31.635 | 0.424 | 0.0000444 |
+| lerobot/umi_cup_in_the_wild | 10 | 39.418 | 0.345 | 0.0000435 |
+| lerobot/umi_cup_in_the_wild | 15 | 44.577 | 0.313 | 0.0000417 |
+| lerobot/umi_cup_in_the_wild | 20 | 47.907 | 0.264 | 0.0000421 |
+| lerobot/umi_cup_in_the_wild | 40 | 52.554 | 0.185 | 0.0000414 |
+| lerobot/umi_cup_in_the_wild | 100 | 58.241 | 0.090 | 0.0000420 |
+| lerobot/umi_cup_in_the_wild | None | 60.530 | 0.042 | 0.0000424 |
+
+**`crf`**
+| repo_id | crf | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 0 | 1.699 | 0.302 | 0.0000097 |
+| lerobot/pusht | 5 | 1.409 | 0.287 | 0.0000142 |
+| lerobot/pusht | 10 | 1.842 | 0.283 | 0.0000184 |
+| lerobot/pusht | 15 | 2.322 | 0.305 | 0.0000268 |
+| lerobot/pusht | 20 | 3.050 | 0.285 | 0.0000402 |
+| lerobot/pusht | None | 3.646 | 0.285 | 0.0000496 |
+| lerobot/pusht | 25 | 3.969 | 0.293 | 0.0000572 |
+| lerobot/pusht | 30 | 5.687 | 0.293 | 0.0000893 |
+| lerobot/pusht | 40 | 10.818 | 0.319 | 0.0001762 |
+| lerobot/pusht | 50 | 18.185 | 0.304 | 0.0002626 |
+| lerobot/umi_cup_in_the_wild | 0 | 1.918 | 0.235 | 0.0000112 |
+| lerobot/umi_cup_in_the_wild | 5 | 3.207 | 0.261 | 0.0000166 |
+| lerobot/umi_cup_in_the_wild | 10 | 4.818 | 0.333 | 0.0000207 |
+| lerobot/umi_cup_in_the_wild | 15 | 7.329 | 0.406 | 0.0000267 |
+| lerobot/umi_cup_in_the_wild | 20 | 11.361 | 0.489 | 0.0000361 |
+| lerobot/umi_cup_in_the_wild | None | 14.932 | 0.537 | 0.0000436 |
+| lerobot/umi_cup_in_the_wild | 25 | 17.741 | 0.578 | 0.0000487 |
+| lerobot/umi_cup_in_the_wild | 30 | 27.983 | 0.453 | 0.0000655 |
+| lerobot/umi_cup_in_the_wild | 40 | 82.449 | 0.767 | 0.0001192 |
+| lerobot/umi_cup_in_the_wild | 50 | 186.145 | 0.816 | 0.0001881 |
+
+**best**
+| repo_id | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | 3.646 | 0.283 | 0.0000496 |
+| lerobot/umi_cup_in_the_wild | 14.932 | 0.543 | 0.0000436 |
+
+### `2_frames_4_space`
+
+**`pix_fmt`**
+| repo_id | pix_fmt | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | yuv420p | 3.788 | 0.257 | 0.0000855 |
+| lerobot/pusht | yuv444p | 3.646 | 0.261 | 0.0000556 |
+| lerobot/umi_cup_in_the_wild | yuv420p | 14.391 | 0.493 | 0.0000476 |
+| lerobot/umi_cup_in_the_wild | yuv444p | 14.932 | 0.371 | 0.0000404 |
+
+**`g`**
+| repo_id | g | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 1 | 2.543 | 0.226 | 0.0000670 |
+| lerobot/pusht | 2 | 3.646 | 0.222 | 0.0000556 |
+| lerobot/pusht | 3 | 4.431 | 0.217 | 0.0000567 |
+| lerobot/pusht | 4 | 5.103 | 0.204 | 0.0000555 |
+| lerobot/pusht | 5 | 5.625 | 0.179 | 0.0000556 |
+| lerobot/pusht | 6 | 5.974 | 0.188 | 0.0000544 |
+| lerobot/pusht | 10 | 6.814 | 0.160 | 0.0000531 |
+| lerobot/pusht | 15 | 7.431 | 0.150 | 0.0000521 |
+| lerobot/pusht | 20 | 7.662 | 0.123 | 0.0000519 |
+| lerobot/pusht | 40 | 8.163 | 0.092 | 0.0000519 |
+| lerobot/pusht | 100 | 8.761 | 0.053 | 0.0000533 |
+| lerobot/pusht | None | 8.909 | 0.034 | 0.0000541 |
+| lerobot/umi_cup_in_the_wild | 1 | 14.411 | 0.409 | 0.0000607 |
+| lerobot/umi_cup_in_the_wild | 2 | 14.932 | 0.381 | 0.0000404 |
+| lerobot/umi_cup_in_the_wild | 3 | 20.174 | 0.355 | 0.0000418 |
+| lerobot/umi_cup_in_the_wild | 4 | 24.889 | 0.346 | 0.0000425 |
+| lerobot/umi_cup_in_the_wild | 5 | 28.825 | 0.354 | 0.0000419 |
+| lerobot/umi_cup_in_the_wild | 6 | 31.635 | 0.336 | 0.0000419 |
+| lerobot/umi_cup_in_the_wild | 10 | 39.418 | 0.314 | 0.0000402 |
+| lerobot/umi_cup_in_the_wild | 15 | 44.577 | 0.269 | 0.0000397 |
+| lerobot/umi_cup_in_the_wild | 20 | 47.907 | 0.246 | 0.0000395 |
+| lerobot/umi_cup_in_the_wild | 40 | 52.554 | 0.171 | 0.0000390 |
+| lerobot/umi_cup_in_the_wild | 100 | 58.241 | 0.091 | 0.0000399 |
+| lerobot/umi_cup_in_the_wild | None | 60.530 | 0.043 | 0.0000409 |
+
+**`crf`**
+| repo_id | crf | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 0 | 1.699 | 0.212 | 0.0000193 |
+| lerobot/pusht | 5 | 1.409 | 0.211 | 0.0000232 |
+| lerobot/pusht | 10 | 1.842 | 0.199 | 0.0000270 |
+| lerobot/pusht | 15 | 2.322 | 0.198 | 0.0000347 |
+| lerobot/pusht | 20 | 3.050 | 0.211 | 0.0000469 |
+| lerobot/pusht | None | 3.646 | 0.206 | 0.0000556 |
+| lerobot/pusht | 25 | 3.969 | 0.210 | 0.0000626 |
+| lerobot/pusht | 30 | 5.687 | 0.223 | 0.0000927 |
+| lerobot/pusht | 40 | 10.818 | 0.227 | 0.0001763 |
+| lerobot/pusht | 50 | 18.185 | 0.223 | 0.0002625 |
+| lerobot/umi_cup_in_the_wild | 0 | 1.918 | 0.147 | 0.0000071 |
+| lerobot/umi_cup_in_the_wild | 5 | 3.207 | 0.182 | 0.0000125 |
+| lerobot/umi_cup_in_the_wild | 10 | 4.818 | 0.222 | 0.0000166 |
+| lerobot/umi_cup_in_the_wild | 15 | 7.329 | 0.270 | 0.0000229 |
+| lerobot/umi_cup_in_the_wild | 20 | 11.361 | 0.325 | 0.0000326 |
+| lerobot/umi_cup_in_the_wild | None | 14.932 | 0.362 | 0.0000404 |
+| lerobot/umi_cup_in_the_wild | 25 | 17.741 | 0.390 | 0.0000459 |
+| lerobot/umi_cup_in_the_wild | 30 | 27.983 | 0.437 | 0.0000633 |
+| lerobot/umi_cup_in_the_wild | 40 | 82.449 | 0.499 | 0.0001186 |
+| lerobot/umi_cup_in_the_wild | 50 | 186.145 | 0.564 | 0.0001879 |
+
+**best**
+| repo_id | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | 3.646 | 0.224 | 0.0000556 |
+| lerobot/umi_cup_in_the_wild | 14.932 | 0.368 | 0.0000404 |
+
+### `6_frames`
+
+**`pix_fmt`**
+| repo_id | pix_fmt | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | yuv420p | 3.788 | 0.660 | 0.0000839 |
+| lerobot/pusht | yuv444p | 3.646 | 0.546 | 0.0000542 |
+| lerobot/umi_cup_in_the_wild | yuv420p | 14.391 | 1.225 | 0.0000497 |
+| lerobot/umi_cup_in_the_wild | yuv444p | 14.932 | 0.908 | 0.0000428 |
+
+**`g`**
+| repo_id | g | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 1 | 2.543 | 0.552 | 0.0000646 |
+| lerobot/pusht | 2 | 3.646 | 0.534 | 0.0000542 |
+| lerobot/pusht | 3 | 4.431 | 0.563 | 0.0000546 |
+| lerobot/pusht | 4 | 5.103 | 0.537 | 0.0000545 |
+| lerobot/pusht | 5 | 5.625 | 0.477 | 0.0000532 |
+| lerobot/pusht | 6 | 5.974 | 0.515 | 0.0000530 |
+| lerobot/pusht | 10 | 6.814 | 0.410 | 0.0000512 |
+| lerobot/pusht | 15 | 7.431 | 0.405 | 0.0000503 |
+| lerobot/pusht | 20 | 7.662 | 0.345 | 0.0000500 |
+| lerobot/pusht | 40 | 8.163 | 0.247 | 0.0000496 |
+| lerobot/pusht | 100 | 8.761 | 0.147 | 0.0000510 |
+| lerobot/pusht | None | 8.909 | 0.100 | 0.0000519 |
+| lerobot/umi_cup_in_the_wild | 1 | 14.411 | 0.997 | 0.0000620 |
+| lerobot/umi_cup_in_the_wild | 2 | 14.932 | 0.911 | 0.0000428 |
+| lerobot/umi_cup_in_the_wild | 3 | 20.174 | 0.869 | 0.0000433 |
+| lerobot/umi_cup_in_the_wild | 4 | 24.889 | 0.874 | 0.0000438 |
+| lerobot/umi_cup_in_the_wild | 5 | 28.825 | 0.864 | 0.0000439 |
+| lerobot/umi_cup_in_the_wild | 6 | 31.635 | 0.834 | 0.0000440 |
+| lerobot/umi_cup_in_the_wild | 10 | 39.418 | 0.781 | 0.0000421 |
+| lerobot/umi_cup_in_the_wild | 15 | 44.577 | 0.679 | 0.0000411 |
+| lerobot/umi_cup_in_the_wild | 20 | 47.907 | 0.652 | 0.0000410 |
+| lerobot/umi_cup_in_the_wild | 40 | 52.554 | 0.465 | 0.0000404 |
+| lerobot/umi_cup_in_the_wild | 100 | 58.241 | 0.245 | 0.0000413 |
+| lerobot/umi_cup_in_the_wild | None | 60.530 | 0.116 | 0.0000417 |
+
+**`crf`**
+| repo_id | crf | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- | --- |
+| lerobot/pusht | 0 | 1.699 | 0.534 | 0.0000163 |
+| lerobot/pusht | 5 | 1.409 | 0.524 | 0.0000205 |
+| lerobot/pusht | 10 | 1.842 | 0.510 | 0.0000245 |
+| lerobot/pusht | 15 | 2.322 | 0.512 | 0.0000324 |
+| lerobot/pusht | 20 | 3.050 | 0.508 | 0.0000452 |
+| lerobot/pusht | None | 3.646 | 0.518 | 0.0000542 |
+| lerobot/pusht | 25 | 3.969 | 0.534 | 0.0000616 |
+| lerobot/pusht | 30 | 5.687 | 0.530 | 0.0000927 |
+| lerobot/pusht | 40 | 10.818 | 0.552 | 0.0001777 |
+| lerobot/pusht | 50 | 18.185 | 0.564 | 0.0002644 |
+| lerobot/umi_cup_in_the_wild | 0 | 1.918 | 0.401 | 0.0000101 |
+| lerobot/umi_cup_in_the_wild | 5 | 3.207 | 0.499 | 0.0000156 |
+| lerobot/umi_cup_in_the_wild | 10 | 4.818 | 0.599 | 0.0000197 |
+| lerobot/umi_cup_in_the_wild | 15 | 7.329 | 0.704 | 0.0000258 |
+| lerobot/umi_cup_in_the_wild | 20 | 11.361 | 0.834 | 0.0000352 |
+| lerobot/umi_cup_in_the_wild | None | 14.932 | 0.925 | 0.0000428 |
+| lerobot/umi_cup_in_the_wild | 25 | 17.741 | 0.978 | 0.0000480 |
+| lerobot/umi_cup_in_the_wild | 30 | 27.983 | 1.088 | 0.0000648 |
+| lerobot/umi_cup_in_the_wild | 40 | 82.449 | 1.324 | 0.0001190 |
+| lerobot/umi_cup_in_the_wild | 50 | 186.145 | 1.436 | 0.0001880 |
+
+**best**
+| repo_id | compression_factor | load_time_factor | avg_per_pixel_l2_error |
+| --- | --- | --- | --- |
+| lerobot/pusht | 3.646 | 0.546 | 0.0000542 |
+| lerobot/umi_cup_in_the_wild | 14.932 | 0.934 | 0.0000428 |

lerobot/common/datasets/_video_benchmark/run_video_benchmark.py

@@ -0,0 +1,357 @@
+import json
+import random
+import shutil
+import subprocess
+import time
+from pathlib import Path
+
+import einops
+import numpy
+import PIL
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.video_utils import (
+    decode_video_frames_torchvision,
+)
+
+
+def get_directory_size(directory):
+    total_size = 0
+    # Iterate over all files and subdirectories recursively
+    for item in directory.rglob("*"):
+        if item.is_file():
+            # Add the file size to the total
+            total_size += item.stat().st_size
+    return total_size
+
+
+def run_video_benchmark(
+    output_dir,
+    cfg,
+    timestamps_mode,
+    seed=1337,
+):
+    output_dir = Path(output_dir)
+    if output_dir.exists():
+        shutil.rmtree(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    repo_id = cfg["repo_id"]
+
+    # TODO(rcadene): rewrite with hardcoding of original images and episodes
+    dataset = LeRobotDataset(repo_id)
+
+    # Get fps
+    fps = dataset.fps
+
+    # we only load first episode
+    ep_num_images = dataset.episode_data_index["to"][0].item()
+
+    # Save/Load image directory for the first episode
+    imgs_dir = Path(f"tmp/data/images/{repo_id}/observation.image_episode_000000")
+    if not imgs_dir.exists():
+        imgs_dir.mkdir(parents=True, exist_ok=True)
+        hf_dataset = dataset.hf_dataset.with_format(None)
+        imgs_dataset = hf_dataset.select_columns("observation.image")
+
+        for i, item in enumerate(imgs_dataset):
+            img = item["observation.image"]
+            img.save(str(imgs_dir / f"frame_{i:06d}.png"), quality=100)
+
+            if i >= ep_num_images - 1:
+                break
+
+    sum_original_frames_size_bytes = get_directory_size(imgs_dir)
+
+    # Encode images into video
+    video_path = output_dir / "episode_0.mp4"
+
+    g = cfg.get("g")
+    crf = cfg.get("crf")
+    pix_fmt = cfg["pix_fmt"]
+
+    cmd = f"ffmpeg -r {fps} "
+    cmd += "-f image2 "
+    cmd += "-loglevel error "
+    cmd += f"-i {str(imgs_dir / 'frame_%06d.png')} "
+    cmd += "-vcodec libx264 "
+    if g is not None:
+        cmd += f"-g {g} "  # ensures at least 1 keyframe every 10 frames
+    # cmd += "-keyint_min 10 " set a minimum of 10 frames between 2 key frames
+    # cmd += "-sc_threshold 0 " disable scene change detection to lower the number of key frames
+    if crf is not None:
+        cmd += f"-crf {crf} "
+    cmd += f"-pix_fmt {pix_fmt} "
+    cmd += f"{str(video_path)}"
+    subprocess.run(cmd.split(" "), check=True)
+
+    video_size_bytes = video_path.stat().st_size
+
+    # Set decoder
+
+    decoder = cfg["decoder"]
+    decoder_kwgs = cfg["decoder_kwgs"]
+    device = cfg["device"]
+
+    if decoder == "torchvision":
+        decode_frames_fn = decode_video_frames_torchvision
+    else:
+        raise ValueError(decoder)
+
+    # Estimate average loading time
+
+    def load_original_frames(imgs_dir, timestamps):
+        frames = []
+        for ts in timestamps:
+            idx = int(ts * fps)
+            frame = PIL.Image.open(imgs_dir / f"frame_{idx:06d}.png")
+            frame = torch.from_numpy(numpy.array(frame))
+            frame = frame.type(torch.float32) / 255
+            frame = einops.rearrange(frame, "h w c -> c h w")
+            frames.append(frame)
+        return frames
+
+    list_avg_load_time = []
+    list_avg_load_time_from_images = []
+    per_pixel_l2_errors = []
+
+    random.seed(seed)
+
+    for t in range(50):
+        # test loading 2 frames that are 4 frames appart, which might be a common setting
+        ts = random.randint(fps, ep_num_images - fps) / fps
+
+        if timestamps_mode == "1_frame":
+            timestamps = [ts]
+        elif timestamps_mode == "2_frames":
+            timestamps = [ts - 1 / fps, ts]
+        elif timestamps_mode == "2_frames_4_space":
+            timestamps = [ts - 4 / fps, ts]
+        elif timestamps_mode == "6_frames":
+            timestamps = [ts - i / fps for i in range(6)][::-1]
+        else:
+            raise ValueError(timestamps_mode)
+
+        num_frames = len(timestamps)
+
+        start_time_s = time.monotonic()
+        frames = decode_frames_fn(
+            video_path, timestamps=timestamps, tolerance_s=1e-4, device=device, **decoder_kwgs
+        )
+        avg_load_time = (time.monotonic() - start_time_s) / num_frames
+        list_avg_load_time.append(avg_load_time)
+
+        start_time_s = time.monotonic()
+        original_frames = load_original_frames(imgs_dir, timestamps)
+        avg_load_time_from_images = (time.monotonic() - start_time_s) / num_frames
+        list_avg_load_time_from_images.append(avg_load_time_from_images)
+
+        # Estimate average L2 error between original frames and decoded frames
+        for i, ts in enumerate(timestamps):
+            # are_close = torch.allclose(frames[i], original_frames[i], atol=0.02)
+            num_pixels = original_frames[i].numel()
+            per_pixel_l2_error = torch.norm(frames[i] - original_frames[i], p=2).item() / num_pixels
+
+            # save decoded frames
+            if t == 0:
+                frame_hwc = (frames[i].permute((1, 2, 0)) * 255).type(torch.uint8).cpu().numpy()
+                PIL.Image.fromarray(frame_hwc).save(output_dir / f"frame_{i:06d}.png")
+
+            # save original_frames
+            idx = int(ts * fps)
+            if t == 0:
+                original_frame = PIL.Image.open(imgs_dir / f"frame_{idx:06d}.png")
+                original_frame.save(output_dir / f"original_frame_{i:06d}.png")
+
+            per_pixel_l2_errors.append(per_pixel_l2_error)
+
+    avg_load_time = float(numpy.array(list_avg_load_time).mean())
+    avg_load_time_from_images = float(numpy.array(list_avg_load_time_from_images).mean())
+    avg_per_pixel_l2_error = float(numpy.array(per_pixel_l2_errors).mean())
+
+    # Save benchmark info
+
+    info = {
+        "sum_original_frames_size_bytes": sum_original_frames_size_bytes,
+        "video_size_bytes": video_size_bytes,
+        "avg_load_time_from_images": avg_load_time_from_images,
+        "avg_load_time": avg_load_time,
+        "compression_factor": sum_original_frames_size_bytes / video_size_bytes,
+        "load_time_factor": avg_load_time_from_images / avg_load_time,
+        "avg_per_pixel_l2_error": avg_per_pixel_l2_error,
+    }
+
+    with open(output_dir / "info.json", "w") as f:
+        json.dump(info, f)
+
+    return info
+
+
+def display_markdown_table(headers, rows):
+    for i, row in enumerate(rows):
+        new_row = []
+        for col in row:
+            if col is None:
+                new_col = "None"
+            elif isinstance(col, float):
+                new_col = f"{col:.3f}"
+                if new_col == "0.000":
+                    new_col = f"{col:.7f}"
+            elif isinstance(col, int):
+                new_col = f"{col}"
+            else:
+                new_col = col
+            new_row.append(new_col)
+        rows[i] = new_row
+
+    header_line = "| " + " | ".join(headers) + " |"
+    separator_line = "| " + " | ".join(["---" for _ in headers]) + " |"
+    body_lines = ["| " + " | ".join(row) + " |" for row in rows]
+    markdown_table = "\n".join([header_line, separator_line] + body_lines)
+    print(markdown_table)
+    print()
+
+
+def load_info(out_dir):
+    with open(out_dir / "info.json") as f:
+        info = json.load(f)
+    return info
+
+
+def main():
+    out_dir = Path("tmp/run_video_benchmark")
+    dry_run = False
+    repo_ids = ["lerobot/pusht", "lerobot/umi_cup_in_the_wild"]
+    timestamps_modes = [
+        "1_frame",
+        "2_frames",
+        "2_frames_4_space",
+        "6_frames",
+    ]
+    for timestamps_mode in timestamps_modes:
+        bench_dir = out_dir / timestamps_mode
+
+        print(f"### `{timestamps_mode}`")
+        print()
+
+        print("**`pix_fmt`**")
+        headers = ["repo_id", "pix_fmt", "compression_factor", "load_time_factor", "avg_per_pixel_l2_error"]
+        rows = []
+        for repo_id in repo_ids:
+            for pix_fmt in ["yuv420p", "yuv444p"]:
+                cfg = {
+                    "repo_id": repo_id,
+                    # video encoding
+                    "g": 2,
+                    "crf": None,
+                    "pix_fmt": pix_fmt,
+                    # video decoding
+                    "device": "cpu",
+                    "decoder": "torchvision",
+                    "decoder_kwgs": {},
+                }
+                if not dry_run:
+                    run_video_benchmark(bench_dir / repo_id / f"torchvision_{pix_fmt}", cfg, timestamps_mode)
+                info = load_info(bench_dir / repo_id / f"torchvision_{pix_fmt}")
+                rows.append(
+                    [
+                        repo_id,
+                        pix_fmt,
+                        info["compression_factor"],
+                        info["load_time_factor"],
+                        info["avg_per_pixel_l2_error"],
+                    ]
+                )
+        display_markdown_table(headers, rows)
+
+        print("**`g`**")
+        headers = ["repo_id", "g", "compression_factor", "load_time_factor", "avg_per_pixel_l2_error"]
+        rows = []
+        for repo_id in repo_ids:
+            for g in [1, 2, 3, 4, 5, 6, 10, 15, 20, 40, 100, None]:
+                cfg = {
+                    "repo_id": repo_id,
+                    # video encoding
+                    "g": g,
+                    "pix_fmt": "yuv444p",
+                    # video decoding
+                    "device": "cpu",
+                    "decoder": "torchvision",
+                    "decoder_kwgs": {},
+                }
+                if not dry_run:
+                    run_video_benchmark(bench_dir / repo_id / f"torchvision_g_{g}", cfg, timestamps_mode)
+                info = load_info(bench_dir / repo_id / f"torchvision_g_{g}")
+                rows.append(
+                    [
+                        repo_id,
+                        g,
+                        info["compression_factor"],
+                        info["load_time_factor"],
+                        info["avg_per_pixel_l2_error"],
+                    ]
+                )
+        display_markdown_table(headers, rows)
+
+        print("**`crf`**")
+        headers = ["repo_id", "crf", "compression_factor", "load_time_factor", "avg_per_pixel_l2_error"]
+        rows = []
+        for repo_id in repo_ids:
+            for crf in [0, 5, 10, 15, 20, None, 25, 30, 40, 50]:
+                cfg = {
+                    "repo_id": repo_id,
+                    # video encoding
+                    "g": 2,
+                    "crf": crf,
+                    "pix_fmt": "yuv444p",
+                    # video decoding
+                    "device": "cpu",
+                    "decoder": "torchvision",
+                    "decoder_kwgs": {},
+                }
+                if not dry_run:
+                    run_video_benchmark(bench_dir / repo_id / f"torchvision_crf_{crf}", cfg, timestamps_mode)
+                info = load_info(bench_dir / repo_id / f"torchvision_crf_{crf}")
+                rows.append(
+                    [
+                        repo_id,
+                        crf,
+                        info["compression_factor"],
+                        info["load_time_factor"],
+                        info["avg_per_pixel_l2_error"],
+                    ]
+                )
+        display_markdown_table(headers, rows)
+
+        print("**best**")
+        headers = ["repo_id", "compression_factor", "load_time_factor", "avg_per_pixel_l2_error"]
+        rows = []
+        for repo_id in repo_ids:
+            cfg = {
+                "repo_id": repo_id,
+                # video encoding
+                "g": 2,
+                "crf": None,
+                "pix_fmt": "yuv444p",
+                # video decoding
+                "device": "cpu",
+                "decoder": "torchvision",
+                "decoder_kwgs": {},
+            }
+            if not dry_run:
+                run_video_benchmark(bench_dir / repo_id / "torchvision_best", cfg, timestamps_mode)
+            info = load_info(bench_dir / repo_id / "torchvision_best")
+            rows.append(
+                [
+                    repo_id,
+                    info["compression_factor"],
+                    info["load_time_factor"],
+                    info["avg_per_pixel_l2_error"],
+                ]
+            )
+        display_markdown_table(headers, rows)
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/datasets/factory.py

@@ -1,25 +1,22 @@
 import logging
-import os
-from pathlib import Path
 
 import torch
 from omegaconf import OmegaConf
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
-DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
-
 
 def make_dataset(
     cfg,
     split="train",
 ):
-    if cfg.env.name not in cfg.dataset.repo_id:
+    if cfg.env.name not in cfg.dataset_repo_id:
         logging.warning(
-            f"There might be a mismatch between your training dataset ({cfg.dataset.repo_id=}) and your environment ({cfg.env.name=})."
+            f"There might be a mismatch between your training dataset ({cfg.dataset_repo_id=}) and your "
+            f"environment ({cfg.env.name=})."
         )
 
-    delta_timestamps = cfg.policy.get("delta_timestamps")
+    delta_timestamps = cfg.training.get("delta_timestamps")
     if delta_timestamps is not None:
         for key in delta_timestamps:
             if isinstance(delta_timestamps[key], str):
@@ -28,9 +25,8 @@ def make_dataset(
     # TODO(rcadene): add data augmentations
 
     dataset = LeRobotDataset(
-        cfg.dataset.repo_id,
+        cfg.dataset_repo_id,
         split=split,
-        root=DATA_DIR,
         delta_timestamps=delta_timestamps,
     )
 

lerobot/common/datasets/lerobot_dataset.py

@@ -1,3 +1,4 @@
+import os
 from pathlib import Path
 
 import datasets
@@ -9,15 +10,20 @@ from lerobot.common.datasets.utils import (
     load_info,
     load_previous_and_future_frames,
     load_stats,
+    load_videos,
 )
+from lerobot.common.datasets.video_utils import VideoFrame, load_from_videos
+
+DATA_DIR = Path(os.environ["DATA_DIR"]) if "DATA_DIR" in os.environ else None
+CODEBASE_VERSION = "v1.3"
 
 
 class LeRobotDataset(torch.utils.data.Dataset):
     def __init__(
         self,
         repo_id: str,
-        version: str | None = "v1.1",
-        root: Path | None = None,
+        version: str | None = CODEBASE_VERSION,
+        root: Path | None = DATA_DIR,
         split: str = "train",
         transform: callable = None,
         delta_timestamps: dict[list[float]] | None = None,
@@ -30,27 +36,73 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.transform = transform
         self.delta_timestamps = delta_timestamps
         # load data from hub or locally when root is provided
+        # TODO(rcadene, aliberts): implement faster transfer
+        # https://huggingface.co/docs/huggingface_hub/en/guides/download#faster-downloads
         self.hf_dataset = load_hf_dataset(repo_id, version, root, split)
         self.episode_data_index = load_episode_data_index(repo_id, version, root)
         self.stats = load_stats(repo_id, version, root)
         self.info = load_info(repo_id, version, root)
+        if self.video:
+            self.videos_dir = load_videos(repo_id, version, root)
 
     @property
     def fps(self) -> int:
+        """Frames per second used during data collection."""
         return self.info["fps"]
 
     @property
-    def image_keys(self) -> list[str]:
-        return [key for key, feats in self.hf_dataset.features.items() if isinstance(feats, datasets.Image)]
+    def video(self) -> bool:
+        """Returns True if this dataset loads video frames from mp4 files.
+        Returns False if it only loads images from png files.
+        """
+        return self.info.get("video", False)
+
+    @property
+    def features(self) -> datasets.Features:
+        return self.hf_dataset.features
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """Keys to access image and video stream from cameras."""
+        keys = []
+        for key, feats in self.hf_dataset.features.items():
+            if isinstance(feats, (datasets.Image, VideoFrame)):
+                keys.append(key)
+        return keys
+
+    @property
+    def video_frame_keys(self) -> list[str]:
+        """Keys to access video frames that requires to be decoded into images.
+
+        Note: It is empty if the dataset contains images only,
+        or equal to `self.cameras` if the dataset contains videos only,
+        or can even be a subset of `self.cameras` in a case of a mixed image/video dataset.
+        """
+        video_frame_keys = []
+        for key, feats in self.hf_dataset.features.items():
+            if isinstance(feats, VideoFrame):
+                video_frame_keys.append(key)
+        return video_frame_keys
 
     @property
     def num_samples(self) -> int:
+        """Number of samples/frames."""
         return len(self.hf_dataset)
 
     @property
     def num_episodes(self) -> int:
+        """Number of episodes."""
         return len(self.hf_dataset.unique("episode_index"))
 
+    @property
+    def tolerance_s(self) -> float:
+        """Tolerance in seconds used to discard loaded frames when their timestamps
+        are not close enough from the requested frames. It is only used when `delta_timestamps`
+        is provided or when loading video frames from mp4 files.
+        """
+        # 1e-4 to account for possible numerical error
+        return 1 / self.fps - 1e-4
+
     def __len__(self):
         return self.num_samples
 
@@ -63,10 +115,65 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 self.hf_dataset,
                 self.episode_data_index,
                 self.delta_timestamps,
-                tol=1 / self.fps - 1e-4,  # 1e-4 to account for possible numerical error
+                self.tolerance_s,
+            )
+
+        if self.video:
+            item = load_from_videos(
+                item,
+                self.video_frame_keys,
+                self.videos_dir,
+                self.tolerance_s,
             )
 
         if self.transform is not None:
             item = self.transform(item)
 
         return item
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}(\n"
+            f"  Repository ID: '{self.repo_id}',\n"
+            f"  Version: '{self.version}',\n"
+            f"  Split: '{self.split}',\n"
+            f"  Number of Samples: {self.num_samples},\n"
+            f"  Number of Episodes: {self.num_episodes},\n"
+            f"  Type: {'video (.mp4)' if self.video else 'image (.png)'},\n"
+            f"  Recorded Frames per Second: {self.fps},\n"
+            f"  Camera Keys: {self.camera_keys},\n"
+            f"  Video Frame Keys: {self.video_frame_keys if self.video else 'N/A'},\n"
+            f"  Transformations: {self.transform},\n"
+            f")"
+        )
+
+    @classmethod
+    def from_preloaded(
+        cls,
+        repo_id: str,
+        version: str | None = CODEBASE_VERSION,
+        root: Path | None = None,
+        split: str = "train",
+        transform: callable = None,
+        delta_timestamps: dict[list[float]] | None = None,
+        # additional preloaded attributes
+        hf_dataset=None,
+        episode_data_index=None,
+        stats=None,
+        info=None,
+        videos_dir=None,
+    ):
+        # create an empty object of type LeRobotDataset
+        obj = cls.__new__(cls)
+        obj.repo_id = repo_id
+        obj.version = version
+        obj.root = root
+        obj.split = split
+        obj.transform = transform
+        obj.delta_timestamps = delta_timestamps
+        obj.hf_dataset = hf_dataset
+        obj.episode_data_index = episode_data_index
+        obj.stats = stats
+        obj.info = info
+        obj.videos_dir = videos_dir
+        return obj

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

lerobot/common/datasets/push_dataset_to_hub/_download_raw.py

@@ -4,20 +4,24 @@ useless dependencies when using datasets.
 """
 
 import io
+import logging
+import shutil
 from pathlib import Path
 
 import tqdm
 
+ALOHA_RAW_URLS_DIR = "lerobot/common/datasets/push_dataset_to_hub/_aloha_raw_urls"
 
-def download_raw(root, dataset_id) -> Path:
+
+def download_raw(raw_dir, dataset_id):
     if "pusht" in dataset_id:
-        return download_pusht(root=root, dataset_id=dataset_id)
+        download_pusht(raw_dir)
     elif "xarm" in dataset_id:
-        return download_xarm(root=root, dataset_id=dataset_id)
+        download_xarm(raw_dir)
     elif "aloha" in dataset_id:
-        return download_aloha(root=root, dataset_id=dataset_id)
+        download_aloha(raw_dir, dataset_id)
     elif "umi" in dataset_id:
-        return download_umi(root=root, dataset_id=dataset_id)
+        download_umi(raw_dir)
     else:
         raise ValueError(dataset_id)
 
@@ -45,135 +49,122 @@ def download_and_extract_zip(url: str, destination_folder: Path) -> bool:
 
         with zipfile.ZipFile(zip_file, "r") as zip_ref:
             zip_ref.extractall(destination_folder)
-        return True
-    else:
-        return False
 
 
-def download_pusht(root: str, dataset_id: str = "pusht", fps: int = 10) -> Path:
+def download_pusht(raw_dir: str):
     pusht_url = "https://diffusion-policy.cs.columbia.edu/data/training/pusht.zip"
-    pusht_zarr = Path("pusht/pusht_cchi_v7_replay.zarr")
 
-    root = Path(root)
-    raw_dir: Path = root / f"{dataset_id}_raw"
-    zarr_path: Path = (raw_dir / pusht_zarr).resolve()
-    if not zarr_path.is_dir():
-        raw_dir.mkdir(parents=True, exist_ok=True)
-        download_and_extract_zip(pusht_url, raw_dir)
-    return zarr_path
+    raw_dir = Path(raw_dir)
+    raw_dir.mkdir(parents=True, exist_ok=True)
+    download_and_extract_zip(pusht_url, raw_dir)
+    # file is created inside a useful "pusht" directory, so we move it out and delete the dir
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    shutil.move(raw_dir / "pusht" / "pusht_cchi_v7_replay.zarr", zarr_path)
+    shutil.rmtree(raw_dir / "pusht")
 
 
-def download_xarm(root: str, dataset_id: str, fps: int = 15) -> Path:
-    root = Path(root)
-    raw_dir: Path = root / "xarm_datasets_raw"
-    if not raw_dir.exists():
-        import zipfile
+def download_xarm(raw_dir: Path):
+    """Download all xarm datasets at once"""
+    import zipfile
 
-        import gdown
+    import gdown
 
-        raw_dir.mkdir(parents=True, exist_ok=True)
-        # from https://github.com/fyhMer/fowm/blob/main/scripts/download_datasets.py
-        url = "https://drive.google.com/uc?id=1nhxpykGtPDhmQKm-_B8zBSywVRdgeVya"
-        zip_path = raw_dir / "data.zip"
-        gdown.download(url, str(zip_path), quiet=False)
-        print("Extracting...")
-        with zipfile.ZipFile(str(zip_path), "r") as zip_f:
-            for member in zip_f.namelist():
-                if member.startswith("data/xarm") and member.endswith(".pkl"):
-                    print(member)
-                    zip_f.extract(member=member)
-        zip_path.unlink()
-
-    dataset_path: Path = root / f"{dataset_id}"
-    return dataset_path
+    raw_dir = Path(raw_dir)
+    raw_dir.mkdir(parents=True, exist_ok=True)
+    # from https://github.com/fyhMer/fowm/blob/main/scripts/download_datasets.py
+    url = "https://drive.google.com/uc?id=1nhxpykGtPDhmQKm-_B8zBSywVRdgeVya"
+    zip_path = raw_dir / "data.zip"
+    gdown.download(url, str(zip_path), quiet=False)
+    print("Extracting...")
+    with zipfile.ZipFile(str(zip_path), "r") as zip_f:
+        for pkl_path in zip_f.namelist():
+            if pkl_path.startswith("data/xarm") and pkl_path.endswith(".pkl"):
+                zip_f.extract(member=pkl_path)
+                # move to corresponding raw directory
+                extract_dir = pkl_path.replace("/buffer.pkl", "")
+                raw_pkl_path = raw_dir / "buffer.pkl"
+                shutil.move(pkl_path, raw_pkl_path)
+                shutil.rmtree(extract_dir)
+    zip_path.unlink()
 
 
-def download_aloha(root: str, dataset_id: str) -> Path:
-    folder_urls = {
-        "aloha_sim_insertion_human": "https://drive.google.com/drive/folders/1RgyD0JgTX30H4IM5XZn8I3zSV_mr8pyF",
-        "aloha_sim_insertion_scripted": "https://drive.google.com/drive/folders/1TsojQQSXtHEoGnqgJ3gmpPQR2DPLtS2N",
-        "aloha_sim_transfer_cube_human": "https://drive.google.com/drive/folders/1sc-E4QYW7A0o23m1u2VWNGVq5smAsfCo",
-        "aloha_sim_transfer_cube_scripted": "https://drive.google.com/drive/folders/1aRyoOhQwxhyt1J8XgEig4s6kzaw__LXj",
-    }
+def download_aloha(raw_dir: Path, dataset_id: str):
+    import gdown
 
-    ep48_urls = {
-        "aloha_sim_insertion_human": "https://drive.google.com/file/d/18Cudl6nikDtgRolea7je8iF_gGKzynOP/view?usp=drive_link",
-        "aloha_sim_insertion_scripted": "https://drive.google.com/file/d/1wfMSZ24oOh5KR_0aaP3Cnu_c4ZCveduB/view?usp=drive_link",
-        "aloha_sim_transfer_cube_human": "https://drive.google.com/file/d/18smMymtr8tIxaNUQ61gW6dG50pt3MvGq/view?usp=drive_link",
-        "aloha_sim_transfer_cube_scripted": "https://drive.google.com/file/d/1pnGIOd-E4-rhz2P3VxpknMKRZCoKt6eI/view?usp=drive_link",
-    }
+    subset_id = dataset_id.replace("aloha_", "")
+    urls_path = Path(ALOHA_RAW_URLS_DIR) / f"{subset_id}.txt"
+    assert urls_path.exists(), f"{subset_id}.txt not found in '{ALOHA_RAW_URLS_DIR}' directory."
 
-    ep49_urls = {
-        "aloha_sim_insertion_human": "https://drive.google.com/file/d/1C1kZYyROzs-PrLc0SkDgUgMi4-L3lauE/view?usp=drive_link",
-        "aloha_sim_insertion_scripted": "https://drive.google.com/file/d/17EuCUWS6uCCr6yyNzpXdcdE-_TTNCKtf/view?usp=drive_link",
-        "aloha_sim_transfer_cube_human": "https://drive.google.com/file/d/1Nk7l53d9sJoGDBKAOnNrExX5nLacATc6/view?usp=drive_link",
-        "aloha_sim_transfer_cube_scripted": "https://drive.google.com/file/d/1GKReZHrXU73NMiC5zKCq_UtqPVtYq8eo/view?usp=drive_link",
-    }
-    num_episodes = {  # noqa: F841 # we keep this for reference
-        "aloha_sim_insertion_human": 50,
-        "aloha_sim_insertion_scripted": 50,
-        "aloha_sim_transfer_cube_human": 50,
-        "aloha_sim_transfer_cube_scripted": 50,
-    }
+    with open(urls_path) as f:
+        # strip lines and ignore empty lines
+        urls = [url.strip() for url in f if url.strip()]
 
-    episode_len = {  # noqa: F841 # we keep this for reference
-        "aloha_sim_insertion_human": 500,
-        "aloha_sim_insertion_scripted": 400,
-        "aloha_sim_transfer_cube_human": 400,
-        "aloha_sim_transfer_cube_scripted": 400,
-    }
+    # sanity check
+    for url in urls:
+        assert (
+            "drive.google.com/drive/folders" in url or "drive.google.com/file" in url
+        ), f"Wrong url provided '{url}' in file '{urls_path}'."
 
-    cameras = {  # noqa: F841 # we keep this for reference
-        "aloha_sim_insertion_human": ["top"],
-        "aloha_sim_insertion_scripted": ["top"],
-        "aloha_sim_transfer_cube_human": ["top"],
-        "aloha_sim_transfer_cube_scripted": ["top"],
-    }
-    root = Path(root)
-    raw_dir: Path = root / f"{dataset_id}_raw"
-    if not raw_dir.is_dir():
-        import gdown
+    raw_dir = Path(raw_dir)
+    raw_dir.mkdir(parents=True, exist_ok=True)
 
-        assert dataset_id in folder_urls
-        assert dataset_id in ep48_urls
-        assert dataset_id in ep49_urls
+    logging.info(f"Start downloading from google drive for {dataset_id}")
+    for url in urls:
+        if "drive.google.com/drive/folders" in url:
+            # when a folder url is given, download up to 50 files from the folder
+            gdown.download_folder(url, output=str(raw_dir), remaining_ok=True)
 
-        raw_dir.mkdir(parents=True, exist_ok=True)
+        elif "drive.google.com/file" in url:
+            # because of the 50 files limit per folder, we download the remaining files (file by file)
+            gdown.download(url, output=str(raw_dir), fuzzy=True)
 
-        gdown.download_folder(folder_urls[dataset_id], output=str(raw_dir))
-
-        # because of the 50 files limit per directory, two files episode 48 and 49 were missing
-        gdown.download(ep48_urls[dataset_id], output=str(raw_dir / "episode_48.hdf5"), fuzzy=True)
-        gdown.download(ep49_urls[dataset_id], output=str(raw_dir / "episode_49.hdf5"), fuzzy=True)
-    return raw_dir
+    logging.info(f"End downloading from google drive for {dataset_id}")
 
 
-def download_umi(root: str, dataset_id: str) -> Path:
+def download_umi(raw_dir: Path):
     url_cup_in_the_wild = "https://real.stanford.edu/umi/data/zarr_datasets/cup_in_the_wild.zarr.zip"
-    cup_in_the_wild_zarr = Path("umi/cup_in_the_wild/cup_in_the_wild.zarr")
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
 
-    root = Path(root)
-    raw_dir: Path = root / f"{dataset_id}_raw"
-    zarr_path: Path = (raw_dir / cup_in_the_wild_zarr).resolve()
-    if not zarr_path.is_dir():
-        raw_dir.mkdir(parents=True, exist_ok=True)
-        download_and_extract_zip(url_cup_in_the_wild, zarr_path)
-    return zarr_path
+    raw_dir = Path(raw_dir)
+    raw_dir.mkdir(parents=True, exist_ok=True)
+    download_and_extract_zip(url_cup_in_the_wild, zarr_path)
 
 
 if __name__ == "__main__":
-    root = "data"
+    data_dir = Path("data")
     dataset_ids = [
         "pusht",
         "xarm_lift_medium",
         "xarm_lift_medium_replay",
         "xarm_push_medium",
         "xarm_push_medium_replay",
+        "aloha_mobile_cabinet",
+        "aloha_mobile_chair",
+        "aloha_mobile_elevator",
+        "aloha_mobile_shrimp",
+        "aloha_mobile_wash_pan",
+        "aloha_mobile_wipe_wine",
         "aloha_sim_insertion_human",
         "aloha_sim_insertion_scripted",
         "aloha_sim_transfer_cube_human",
         "aloha_sim_transfer_cube_scripted",
+        "aloha_static_battery",
+        "aloha_static_candy",
+        "aloha_static_coffee",
+        "aloha_static_coffee_new",
+        "aloha_static_cups_open",
+        "aloha_static_fork_pick_up",
+        "aloha_static_pingpong_test",
+        "aloha_static_pro_pencil",
+        "aloha_static_screw_driver",
+        "aloha_static_tape",
+        "aloha_static_thread_velcro",
+        "aloha_static_towel",
+        "aloha_static_vinh_cup",
+        "aloha_static_vinh_cup_left",
+        "aloha_static_ziploc_slide",
         "umi_cup_in_the_wild",
     ]
     for dataset_id in dataset_ids:
-        download_raw(root=root, dataset_id=dataset_id)
+        raw_dir = data_dir / f"{dataset_id}_raw"
+        download_raw(raw_dir, dataset_id)

lerobot/common/datasets/push_dataset_to_hub/aloha_hdf5_format.py

@@ -0,0 +1,183 @@
+"""
+Contains utilities to process raw data format of HDF5 files like in: https://github.com/tonyzhaozh/act
+"""
+
+import re
+import shutil
+from pathlib import Path
+
+import h5py
+import numpy as np
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def get_cameras(hdf5_data):
+    # ignore depth channel, not currently handled
+    # TODO(rcadene): add depth
+    rgb_cameras = [key for key in hdf5_data["/observations/images"].keys() if "depth" not in key]  # noqa: SIM118
+    return rgb_cameras
+
+
+def check_format(raw_dir) -> bool:
+    # only frames from simulation are uncompressed
+    compressed_images = "sim" not in raw_dir.name
+
+    hdf5_paths = list(raw_dir.glob("episode_*.hdf5"))
+    assert len(hdf5_paths) != 0
+    for hdf5_path in hdf5_paths:
+        with h5py.File(hdf5_path, "r") as data:
+            assert "/action" in data
+            assert "/observations/qpos" in data
+
+            assert data["/action"].ndim == 2
+            assert data["/observations/qpos"].ndim == 2
+
+            num_frames = data["/action"].shape[0]
+            assert num_frames == data["/observations/qpos"].shape[0]
+
+            for camera in get_cameras(data):
+                assert num_frames == data[f"/observations/images/{camera}"].shape[0]
+
+                if compressed_images:
+                    assert data[f"/observations/images/{camera}"].ndim == 2
+                else:
+                    assert data[f"/observations/images/{camera}"].ndim == 4
+                    b, h, w, c = data[f"/observations/images/{camera}"].shape
+                    assert c < h and c < w, f"Expect (h,w,c) image format but ({h=},{w=},{c=}) provided."
+
+
+def load_from_raw(raw_dir, out_dir, fps, video, debug):
+    # only frames from simulation are uncompressed
+    compressed_images = "sim" not in raw_dir.name
+
+    hdf5_files = list(raw_dir.glob("*.hdf5"))
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+
+    for ep_path in tqdm.tqdm(hdf5_files, total=len(hdf5_files)):
+        with h5py.File(ep_path, "r") as ep:
+            ep_idx = int(re.search(r"episode_(\d+)", ep_path.name).group(1))
+            num_frames = ep["/action"].shape[0]
+
+            # last step of demonstration is considered done
+            done = torch.zeros(num_frames, dtype=torch.bool)
+            done[-1] = True
+
+            state = torch.from_numpy(ep["/observations/qpos"][:])
+            action = torch.from_numpy(ep["/action"][:])
+
+            ep_dict = {}
+
+            for camera in get_cameras(ep):
+                img_key = f"observation.images.{camera}"
+
+                if compressed_images:
+                    import cv2
+
+                    # load one compressed image after the other in RAM and uncompress
+                    imgs_array = []
+                    for data in ep[f"/observations/images/{camera}"]:
+                        imgs_array.append(cv2.imdecode(data, 1))
+                    imgs_array = np.array(imgs_array)
+
+                else:
+                    # load all images in RAM
+                    imgs_array = ep[f"/observations/images/{camera}"][:]
+
+                if video:
+                    # save png images in temporary directory
+                    tmp_imgs_dir = out_dir / "tmp_images"
+                    save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+                    # encode images to a mp4 video
+                    fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                    video_path = out_dir / "videos" / fname
+                    encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+                    # clean temporary images directory
+                    shutil.rmtree(tmp_imgs_dir)
+
+                    # store the reference to the video frame
+                    ep_dict[img_key] = [
+                        {"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)
+                    ]
+                else:
+                    ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+            ep_dict["observation.state"] = state
+            ep_dict["action"] = action
+            ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+            ep_dict["next.done"] = done
+            # TODO(rcadene): add reward and success by computing them in sim
+
+            assert isinstance(ep_idx, int)
+            ep_dicts.append(ep_dict)
+
+            episode_data_index["from"].append(id_from)
+            episode_data_index["to"].append(id_from + num_frames)
+
+        id_from += num_frames
+
+        # process first episode only
+        if debug:
+            break
+
+    data_dict = concatenate_episodes(ep_dicts)
+    return data_dict, episode_data_index
+
+
+def to_hf_dataset(data_dict, video) -> Dataset:
+    features = {}
+
+    keys = [key for key in data_dict if "observation.images." in key]
+    for key in keys:
+        if video:
+            features[key] = VideoFrame()
+        else:
+            features[key] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 50
+
+    data_dir, episode_data_index = load_from_raw(raw_dir, out_dir, fps, video, debug)
+    hf_dataset = to_hf_dataset(data_dir, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/aloha_processor.py

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

lerobot/common/datasets/push_dataset_to_hub/compute_stats.py

@@ -0,0 +1,146 @@
+from copy import deepcopy
+from math import ceil
+
+import datasets
+import einops
+import torch
+import tqdm
+from datasets import Image
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+def get_stats_einops_patterns(dataset: LeRobotDataset | datasets.Dataset, num_workers=0):
+    """These einops patterns will be used to aggregate batches and compute statistics.
+
+    Note: We assume the images are in channel first format
+    """
+
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=num_workers,
+        batch_size=2,
+        shuffle=False,
+    )
+    batch = next(iter(dataloader))
+
+    stats_patterns = {}
+    for key, feats_type in dataset.features.items():
+        # sanity check that tensors are not float64
+        assert batch[key].dtype != torch.float64
+
+        if isinstance(feats_type, (VideoFrame, Image)):
+            # sanity check that images are channel first
+            _, c, h, w = batch[key].shape
+            assert c < h and c < w, f"expect channel first images, but instead {batch[key].shape}"
+
+            # sanity check that images are float32 in range [0,1]
+            assert batch[key].dtype == torch.float32, f"expect torch.float32, but instead {batch[key].dtype=}"
+            assert batch[key].max() <= 1, f"expect pixels lower than 1, but instead {batch[key].max()=}"
+            assert batch[key].min() >= 0, f"expect pixels greater than 1, but instead {batch[key].min()=}"
+
+            stats_patterns[key] = "b c h w -> c 1 1"
+        elif batch[key].ndim == 2:
+            stats_patterns[key] = "b c -> c "
+        elif batch[key].ndim == 1:
+            stats_patterns[key] = "b -> 1"
+        else:
+            raise ValueError(f"{key}, {feats_type}, {batch[key].shape}")
+
+    return stats_patterns
+
+
+def compute_stats(
+    dataset: LeRobotDataset | datasets.Dataset, batch_size=32, num_workers=16, max_num_samples=None
+):
+    if max_num_samples is None:
+        max_num_samples = len(dataset)
+
+    # for more info on why we need to set the same number of workers, see `load_from_videos`
+    stats_patterns = get_stats_einops_patterns(dataset, num_workers)
+
+    # mean and std will be computed incrementally while max and min will track the running value.
+    mean, std, max, min = {}, {}, {}, {}
+    for key in stats_patterns:
+        mean[key] = torch.tensor(0.0).float()
+        std[key] = torch.tensor(0.0).float()
+        max[key] = torch.tensor(-float("inf")).float()
+        min[key] = torch.tensor(float("inf")).float()
+
+    def create_seeded_dataloader(dataset, batch_size, seed):
+        generator = torch.Generator()
+        generator.manual_seed(seed)
+        dataloader = torch.utils.data.DataLoader(
+            dataset,
+            num_workers=num_workers,
+            batch_size=batch_size,
+            shuffle=True,
+            drop_last=False,
+            generator=generator,
+        )
+        return dataloader
+
+    # Note: Due to be refactored soon. The point of storing `first_batch` is to make sure we don't get
+    # surprises when rerunning the sampler.
+    first_batch = None
+    running_item_count = 0  # for online mean computation
+    dataloader = create_seeded_dataloader(dataset, batch_size, seed=1337)
+    for i, batch in enumerate(
+        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute mean, min, max")
+    ):
+        this_batch_size = len(batch["index"])
+        running_item_count += this_batch_size
+        if first_batch is None:
+            first_batch = deepcopy(batch)
+        for key, pattern in stats_patterns.items():
+            batch[key] = batch[key].float()
+            # Numerically stable update step for mean computation.
+            batch_mean = einops.reduce(batch[key], pattern, "mean")
+            # Hint: to update the mean we need x̄ₙ = (Nₙ₋₁x̄ₙ₋₁ + Bₙxₙ) / Nₙ, where the subscript represents
+            # the update step, N is the running item count, B is this batch size, x̄ is the running mean,
+            # and x is the current batch mean. Some rearrangement is then required to avoid risking
+            # numerical overflow. Another hint: Nₙ₋₁ = Nₙ - Bₙ. Rearrangement yields
+            # x̄ₙ = x̄ₙ₋₁ + Bₙ * (xₙ - x̄ₙ₋₁) / Nₙ
+            mean[key] = mean[key] + this_batch_size * (batch_mean - mean[key]) / running_item_count
+            max[key] = torch.maximum(max[key], einops.reduce(batch[key], pattern, "max"))
+            min[key] = torch.minimum(min[key], einops.reduce(batch[key], pattern, "min"))
+
+        if i == ceil(max_num_samples / batch_size) - 1:
+            break
+
+    first_batch_ = None
+    running_item_count = 0  # for online std computation
+    dataloader = create_seeded_dataloader(dataset, batch_size, seed=1337)
+    for i, batch in enumerate(
+        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute std")
+    ):
+        this_batch_size = len(batch["index"])
+        running_item_count += this_batch_size
+        # Sanity check to make sure the batches are still in the same order as before.
+        if first_batch_ is None:
+            first_batch_ = deepcopy(batch)
+            for key in stats_patterns:
+                assert torch.equal(first_batch_[key], first_batch[key])
+        for key, pattern in stats_patterns.items():
+            batch[key] = batch[key].float()
+            # Numerically stable update step for mean computation (where the mean is over squared
+            # residuals).See notes in the mean computation loop above.
+            batch_std = einops.reduce((batch[key] - mean[key]) ** 2, pattern, "mean")
+            std[key] = std[key] + this_batch_size * (batch_std - std[key]) / running_item_count
+
+        if i == ceil(max_num_samples / batch_size) - 1:
+            break
+
+    for key in stats_patterns:
+        std[key] = torch.sqrt(std[key])
+
+    stats = {}
+    for key in stats_patterns:
+        stats[key] = {
+            "mean": mean[key],
+            "std": std[key],
+            "max": max[key],
+            "min": min[key],
+        }
+    return stats

lerobot/common/datasets/push_dataset_to_hub/pusht_processor.py

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

lerobot/common/datasets/push_dataset_to_hub/pusht_zarr_format.py

@@ -0,0 +1,214 @@
+"""Process zarr files formatted like in: https://github.com/real-stanford/diffusion_policy"""
+
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir):
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    required_datasets = {
+        "data/action",
+        "data/img",
+        "data/keypoint",
+        "data/n_contacts",
+        "data/state",
+        "meta/episode_ends",
+    }
+    for dataset in required_datasets:
+        assert dataset in zarr_data
+    nb_frames = zarr_data["data/img"].shape[0]
+
+    required_datasets.remove("meta/episode_ends")
+
+    assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
+
+
+def load_from_raw(raw_dir, out_dir, fps, video, debug):
+    try:
+        import pymunk
+        from gym_pusht.envs.pusht import PushTEnv, pymunk_to_shapely
+
+        from lerobot.common.datasets.push_dataset_to_hub._diffusion_policy_replay_buffer import (
+            ReplayBuffer as DiffusionPolicyReplayBuffer,
+        )
+    except ModuleNotFoundError as e:
+        print("`gym_pusht` is not installed. Please install it with `pip install 'lerobot[gym_pusht]'`")
+        raise e
+    # as define in gmy-pusht env: https://github.com/huggingface/gym-pusht/blob/e0684ff988d223808c0a9dcfaba9dc4991791370/gym_pusht/envs/pusht.py#L174
+    success_threshold = 0.95  # 95% coverage,
+
+    zarr_path = raw_dir / "pusht_cchi_v7_replay.zarr"
+    zarr_data = DiffusionPolicyReplayBuffer.copy_from_path(zarr_path)
+
+    episode_ids = torch.from_numpy(zarr_data.get_episode_idxs())
+    num_episodes = zarr_data.meta["episode_ends"].shape[0]
+    assert len(
+        {zarr_data[key].shape[0] for key in zarr_data.keys()}  # noqa: SIM118
+    ), "Some data type dont have the same number of total frames."
+
+    # TODO(rcadene): verify that goal pose is expected to be fixed
+    goal_pos_angle = np.array([256, 256, np.pi / 4])  # x, y, theta (in radians)
+    goal_body = PushTEnv.get_goal_pose_body(goal_pos_angle)
+
+    imgs = torch.from_numpy(zarr_data["img"])  # b h w c
+    states = torch.from_numpy(zarr_data["state"])
+    actions = torch.from_numpy(zarr_data["action"])
+
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+    for ep_idx in tqdm.tqdm(range(num_episodes)):
+        id_to = zarr_data.meta["episode_ends"][ep_idx]
+        num_frames = id_to - id_from
+
+        # sanity check
+        assert (episode_ids[id_from:id_to] == ep_idx).all()
+
+        # get image
+        image = imgs[id_from:id_to]
+        assert image.min() >= 0.0
+        assert image.max() <= 255.0
+        image = image.type(torch.uint8)
+
+        # get state
+        state = states[id_from:id_to]
+        agent_pos = state[:, :2]
+        block_pos = state[:, 2:4]
+        block_angle = state[:, 4]
+
+        # get reward, success, done
+        reward = torch.zeros(num_frames)
+        success = torch.zeros(num_frames, dtype=torch.bool)
+        done = torch.zeros(num_frames, dtype=torch.bool)
+        for i in range(num_frames):
+            space = pymunk.Space()
+            space.gravity = 0, 0
+            space.damping = 0
+
+            # Add walls.
+            walls = [
+                PushTEnv.add_segment(space, (5, 506), (5, 5), 2),
+                PushTEnv.add_segment(space, (5, 5), (506, 5), 2),
+                PushTEnv.add_segment(space, (506, 5), (506, 506), 2),
+                PushTEnv.add_segment(space, (5, 506), (506, 506), 2),
+            ]
+            space.add(*walls)
+
+            block_body = PushTEnv.add_tee(space, block_pos[i].tolist(), block_angle[i].item())
+            goal_geom = pymunk_to_shapely(goal_body, block_body.shapes)
+            block_geom = pymunk_to_shapely(block_body, block_body.shapes)
+            intersection_area = goal_geom.intersection(block_geom).area
+            goal_area = goal_geom.area
+            coverage = intersection_area / goal_area
+            reward[i] = np.clip(coverage / success_threshold, 0, 1)
+            success[i] = coverage > success_threshold
+
+        # last step of demonstration is considered done
+        done[-1] = True
+
+        ep_dict = {}
+
+        imgs_array = [x.numpy() for x in image]
+        img_key = "observation.image"
+        if video:
+            # save png images in temporary directory
+            tmp_imgs_dir = out_dir / "tmp_images"
+            save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+            # encode images to a mp4 video
+            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+            video_path = out_dir / "videos" / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+            # clean temporary images directory
+            shutil.rmtree(tmp_imgs_dir)
+
+            # store the reference to the video frame
+            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+        else:
+            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = agent_pos
+        ep_dict["action"] = actions[id_from:id_to]
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        # ep_dict["next.observation.image"] = image[1:],
+        # ep_dict["next.observation.state"] = agent_pos[1:],
+        # TODO(rcadene)] = verify that reward and done are aligned with image and agent_pos
+        ep_dict["next.reward"] = torch.cat([reward[1:], reward[[-1]]])
+        ep_dict["next.done"] = torch.cat([done[1:], done[[-1]]])
+        ep_dict["next.success"] = torch.cat([success[1:], success[[-1]]])
+        ep_dicts.append(ep_dict)
+
+        episode_data_index["from"].append(id_from)
+        episode_data_index["to"].append(id_from + num_frames)
+
+        id_from += num_frames
+
+        # process first episode only
+        if debug:
+            break
+
+    data_dict = concatenate_episodes(ep_dicts)
+    return data_dict, episode_data_index
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["next.success"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 10
+
+    data_dict, episode_data_index = load_from_raw(raw_dir, out_dir, fps, video, debug)
+    hf_dataset = to_hf_dataset(data_dict, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/umi_processor.py

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

lerobot/common/datasets/push_dataset_to_hub/umi_zarr_format.py

@@ -0,0 +1,207 @@
+"""Process UMI (Universal Manipulation Interface) data stored in Zarr format like in: https://github.com/real-stanford/universal_manipulation_interface"""
+
+import logging
+import shutil
+from pathlib import Path
+
+import numpy as np
+import torch
+import tqdm
+import zarr
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.push_dataset_to_hub._umi_imagecodecs_numcodecs import register_codecs
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir) -> bool:
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    required_datasets = {
+        "data/robot0_demo_end_pose",
+        "data/robot0_demo_start_pose",
+        "data/robot0_eef_pos",
+        "data/robot0_eef_rot_axis_angle",
+        "data/robot0_gripper_width",
+        "meta/episode_ends",
+        "data/camera0_rgb",
+    }
+    for dataset in required_datasets:
+        if dataset not in zarr_data:
+            return False
+
+    # mandatory to access zarr_data
+    register_codecs()
+    nb_frames = zarr_data["data/camera0_rgb"].shape[0]
+
+    required_datasets.remove("meta/episode_ends")
+    assert all(nb_frames == zarr_data[dataset].shape[0] for dataset in required_datasets)
+
+
+def get_episode_idxs(episode_ends: np.ndarray) -> np.ndarray:
+    # Optimized and simplified version of this function: https://github.com/real-stanford/universal_manipulation_interface/blob/298776ce251f33b6b3185a98d6e7d1f9ad49168b/diffusion_policy/common/replay_buffer.py#L374
+    from numba import jit
+
+    @jit(nopython=True)
+    def _get_episode_idxs(episode_ends):
+        result = np.zeros((episode_ends[-1],), dtype=np.int64)
+        start_idx = 0
+        for episode_number, end_idx in enumerate(episode_ends):
+            result[start_idx:end_idx] = episode_number
+            start_idx = end_idx
+        return result
+
+    return _get_episode_idxs(episode_ends)
+
+
+def load_from_raw(raw_dir, out_dir, fps, video, debug):
+    zarr_path = raw_dir / "cup_in_the_wild.zarr"
+    zarr_data = zarr.open(zarr_path, mode="r")
+
+    # We process the image data separately because it is too large to fit in memory
+    end_pose = torch.from_numpy(zarr_data["data/robot0_demo_end_pose"][:])
+    start_pos = torch.from_numpy(zarr_data["data/robot0_demo_start_pose"][:])
+    eff_pos = torch.from_numpy(zarr_data["data/robot0_eef_pos"][:])
+    eff_rot_axis_angle = torch.from_numpy(zarr_data["data/robot0_eef_rot_axis_angle"][:])
+    gripper_width = torch.from_numpy(zarr_data["data/robot0_gripper_width"][:])
+
+    states_pos = torch.cat([eff_pos, eff_rot_axis_angle], dim=1)
+    states = torch.cat([states_pos, gripper_width], dim=1)
+
+    episode_ends = zarr_data["meta/episode_ends"][:]
+    num_episodes = episode_ends.shape[0]
+
+    episode_ids = torch.from_numpy(get_episode_idxs(episode_ends))
+
+    # We convert it in torch tensor later because the jit function does not support torch tensors
+    episode_ends = torch.from_numpy(episode_ends)
+
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+    for ep_idx in tqdm.tqdm(range(num_episodes)):
+        id_to = episode_ends[ep_idx]
+        num_frames = id_to - id_from
+
+        # sanity heck
+        assert (episode_ids[id_from:id_to] == ep_idx).all()
+
+        # TODO(rcadene): save temporary images of the episode?
+
+        state = states[id_from:id_to]
+
+        ep_dict = {}
+
+        # load 57MB of images in RAM (400x224x224x3 uint8)
+        imgs_array = zarr_data["data/camera0_rgb"][id_from:id_to]
+        img_key = "observation.image"
+        if video:
+            # save png images in temporary directory
+            tmp_imgs_dir = out_dir / "tmp_images"
+            save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+            # encode images to a mp4 video
+            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+            video_path = out_dir / "videos" / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+            # clean temporary images directory
+            shutil.rmtree(tmp_imgs_dir)
+
+            # store the reference to the video frame
+            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+        else:
+            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = state
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        ep_dict["episode_data_index_from"] = torch.tensor([id_from] * num_frames)
+        ep_dict["episode_data_index_to"] = torch.tensor([id_from + num_frames] * num_frames)
+        ep_dict["end_pose"] = end_pose[id_from:id_to]
+        ep_dict["start_pos"] = start_pos[id_from:id_to]
+        ep_dict["gripper_width"] = gripper_width[id_from:id_to]
+        ep_dicts.append(ep_dict)
+
+        episode_data_index["from"].append(id_from)
+        episode_data_index["to"].append(id_from + num_frames)
+        id_from += num_frames
+
+        # process first episode only
+        if debug:
+            break
+
+    data_dict = concatenate_episodes(ep_dicts)
+
+    total_frames = id_from
+    data_dict["index"] = torch.arange(0, total_frames, 1)
+
+    return data_dict, episode_data_index
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+    features["episode_data_index_from"] = Value(dtype="int64", id=None)
+    features["episode_data_index_to"] = Value(dtype="int64", id=None)
+    # `start_pos` and `end_pos` respectively represent the positions of the end-effector
+    # at the beginning and the end of the episode.
+    # `gripper_width` indicates the distance between the grippers, and this value is included
+    # in the state vector, which comprises the concatenation of the end-effector position
+    # and gripper width.
+    features["end_pose"] = Sequence(
+        length=data_dict["end_pose"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["start_pos"] = Sequence(
+        length=data_dict["start_pos"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["gripper_width"] = Sequence(
+        length=data_dict["gripper_width"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        # For umi cup in the wild: https://arxiv.org/pdf/2402.10329#table.caption.16
+        fps = 10
+
+    if not video:
+        logging.warning(
+            "Generating UMI dataset without `video=True` creates ~150GB on disk and requires ~80GB in RAM."
+        )
+
+    data_dict, episode_data_index = load_from_raw(raw_dir, out_dir, fps, video, debug)
+    hf_dataset = to_hf_dataset(data_dict, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/utils.py

@@ -1,3 +1,8 @@
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+import numpy
+import PIL
 import torch
 
 
@@ -18,3 +23,16 @@ def concatenate_episodes(ep_dicts):
     total_frames = data_dict["frame_index"].shape[0]
     data_dict["index"] = torch.arange(0, total_frames, 1)
     return data_dict
+
+
+def save_images_concurrently(imgs_array: numpy.array, out_dir: Path, max_workers: int = 4):
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    def save_image(img_array, i, out_dir):
+        img = PIL.Image.fromarray(img_array)
+        img.save(str(out_dir / f"frame_{i:06d}.png"), quality=100)
+
+    num_images = len(imgs_array)
+    with ThreadPoolExecutor(max_workers=max_workers) as executor:
+        [executor.submit(save_image, imgs_array[i], i, out_dir) for i in range(num_images)]

lerobot/common/datasets/push_dataset_to_hub/xarm_pkl_format.py

@@ -0,0 +1,163 @@
+"""Process pickle files formatted like in: https://github.com/fyhMer/fowm"""
+
+import pickle
+import shutil
+from pathlib import Path
+
+import einops
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+from PIL import Image as PILImage
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes, save_images_concurrently
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame, encode_video_frames
+
+
+def check_format(raw_dir):
+    keys = {"actions", "rewards", "dones"}
+    nested_keys = {"observations": {"rgb", "state"}, "next_observations": {"rgb", "state"}}
+
+    xarm_files = list(raw_dir.glob("*.pkl"))
+    assert len(xarm_files) > 0
+
+    with open(xarm_files[0], "rb") as f:
+        dataset_dict = pickle.load(f)
+
+    assert isinstance(dataset_dict, dict)
+    assert all(k in dataset_dict for k in keys)
+
+    # Check for consistent lengths in nested keys
+    expected_len = len(dataset_dict["actions"])
+    assert all(len(dataset_dict[key]) == expected_len for key in keys if key in dataset_dict)
+
+    for key, subkeys in nested_keys.items():
+        nested_dict = dataset_dict.get(key, {})
+        assert all(len(nested_dict[subkey]) == expected_len for subkey in subkeys if subkey in nested_dict)
+
+
+def load_from_raw(raw_dir, out_dir, fps, video, debug):
+    pkl_path = raw_dir / "buffer.pkl"
+
+    with open(pkl_path, "rb") as f:
+        pkl_data = pickle.load(f)
+
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+    id_to = 0
+    ep_idx = 0
+    total_frames = pkl_data["actions"].shape[0]
+    for i in tqdm.tqdm(range(total_frames)):
+        id_to += 1
+
+        if not pkl_data["dones"][i]:
+            continue
+
+        num_frames = id_to - id_from
+
+        image = torch.tensor(pkl_data["observations"]["rgb"][id_from:id_to])
+        image = einops.rearrange(image, "b c h w -> b h w c")
+        state = torch.tensor(pkl_data["observations"]["state"][id_from:id_to])
+        action = torch.tensor(pkl_data["actions"][id_from:id_to])
+        # TODO(rcadene): we have a missing last frame which is the observation when the env is done
+        # it is critical to have this frame for tdmpc to predict a "done observation/state"
+        # next_image = torch.tensor(pkl_data["next_observations"]["rgb"][id_from:id_to])
+        # next_state = torch.tensor(pkl_data["next_observations"]["state"][id_from:id_to])
+        next_reward = torch.tensor(pkl_data["rewards"][id_from:id_to])
+        next_done = torch.tensor(pkl_data["dones"][id_from:id_to])
+
+        ep_dict = {}
+
+        imgs_array = [x.numpy() for x in image]
+        img_key = "observation.image"
+        if video:
+            # save png images in temporary directory
+            tmp_imgs_dir = out_dir / "tmp_images"
+            save_images_concurrently(imgs_array, tmp_imgs_dir)
+
+            # encode images to a mp4 video
+            fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+            video_path = out_dir / "videos" / fname
+            encode_video_frames(tmp_imgs_dir, video_path, fps)
+
+            # clean temporary images directory
+            shutil.rmtree(tmp_imgs_dir)
+
+            # store the reference to the video frame
+            ep_dict[img_key] = [{"path": f"videos/{fname}", "timestamp": i / fps} for i in range(num_frames)]
+        else:
+            ep_dict[img_key] = [PILImage.fromarray(x) for x in imgs_array]
+
+        ep_dict["observation.state"] = state
+        ep_dict["action"] = action
+        ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames, dtype=torch.int64)
+        ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+        ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+        # ep_dict["next.observation.image"] = next_image
+        # ep_dict["next.observation.state"] = next_state
+        ep_dict["next.reward"] = next_reward
+        ep_dict["next.done"] = next_done
+        ep_dicts.append(ep_dict)
+
+        episode_data_index["from"].append(id_from)
+        episode_data_index["to"].append(id_from + num_frames)
+
+        id_from = id_to
+        ep_idx += 1
+
+        # process first episode only
+        if debug:
+            break
+
+    data_dict = concatenate_episodes(ep_dicts)
+    return data_dict, episode_data_index
+
+
+def to_hf_dataset(data_dict, video):
+    features = {}
+
+    if video:
+        features["observation.image"] = VideoFrame()
+    else:
+        features["observation.image"] = Image()
+
+    features["observation.state"] = Sequence(
+        length=data_dict["observation.state"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.reward"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+    # TODO(rcadene): add success
+    # features["next.success"] = Value(dtype='bool', id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 15
+
+    data_dict, episode_data_index = load_from_raw(raw_dir, out_dir, fps, video, debug)
+    hf_dataset = to_hf_dataset(data_dict, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

lerobot/common/datasets/push_dataset_to_hub/xarm_processor.py

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

lerobot/common/datasets/utils.py

@@ -1,14 +1,10 @@
 import json
-from copy import deepcopy
-from math import ceil
 from pathlib import Path
 
 import datasets
-import einops
 import torch
-import tqdm
-from datasets import Image, load_dataset, load_from_disk
-from huggingface_hub import hf_hub_download
+from datasets import load_dataset, load_from_disk
+from huggingface_hub import hf_hub_download, snapshot_download
 from PIL import Image as PILImage
 from safetensors.torch import load_file
 from torchvision import transforms
@@ -57,6 +53,9 @@ def hf_transform_to_torch(items_dict):
         if isinstance(first_item, PILImage.Image):
             to_tensor = transforms.ToTensor()
             items_dict[key] = [to_tensor(img) for img in items_dict[key]]
+        elif isinstance(first_item, dict) and "path" in first_item and "timestamp" in first_item:
+            # video frame will be processed downstream
+            pass
         else:
             items_dict[key] = [torch.tensor(x) for x in items_dict[key]]
     return items_dict
@@ -127,17 +126,29 @@ def load_info(repo_id, version, root) -> dict:
     return info
 
 
+def load_videos(repo_id, version, root) -> Path:
+    if root is not None:
+        path = Path(root) / repo_id / "videos"
+    else:
+        # TODO(rcadene): we download the whole repo here. see if we can avoid this
+        repo_dir = snapshot_download(repo_id, repo_type="dataset", revision=version)
+        path = Path(repo_dir) / "videos"
+
+    return path
+
+
 def load_previous_and_future_frames(
     item: dict[str, torch.Tensor],
     hf_dataset: datasets.Dataset,
     episode_data_index: dict[str, torch.Tensor],
     delta_timestamps: dict[str, list[float]],
-    tol: float,
+    tolerance_s: float,
 ) -> dict[torch.Tensor]:
     """
     Given a current item in the dataset containing a timestamp (e.g. 0.6 seconds), and a list of time differences of
     some modalities (e.g. delta_timestamps={"observation.image": [-0.8, -0.2, 0, 0.2]}), this function computes for each
-    given modality a list of query timestamps (e.g. [-0.2, 0.4, 0.6, 0.8]) and loads the closest frames in the dataset.
+    given modality (e.g. "observation.image") a list of query timestamps (e.g. [-0.2, 0.4, 0.6, 0.8]) and loads the closest
+    frames in the dataset.
 
     Importantly, when no frame can be found around a query timestamp within a specified tolerance window, this function
     raises an AssertionError. When a timestamp is queried before the first available timestamp of the episode or after
@@ -156,7 +167,7 @@ def load_previous_and_future_frames(
       They indicate the start index and end index of each episode in the dataset.
     - delta_timestamps (dict): A dictionary containing lists of delta timestamps for each possible modality to be
       retrieved. These deltas are added to the item timestamp to form the query timestamps.
-    - tol (float, optional): The tolerance level used to determine if a data point is close enough to the query
+    - tolerance_s (float, optional): The tolerance level (in seconds) used to determine if a data point is close enough to the query
       timestamp by asserting `tol > difference`. It is suggested to set `tol` to a smaller value than the
       smallest expected inter-frame period, but large enough to account for jitter.
 
@@ -194,11 +205,11 @@ def load_previous_and_future_frames(
 
         # TODO(rcadene): synchronize timestamps + interpolation if needed
 
-        is_pad = min_ > tol
+        is_pad = min_ > tolerance_s
 
         # check violated query timestamps are all outside the episode range
         assert ((query_ts[is_pad] < ep_first_ts) | (ep_last_ts < query_ts[is_pad])).all(), (
-            f"One or several timestamps unexpectedly violate the tolerance ({min_} > {tol=}) inside episode range."
+            f"One or several timestamps unexpectedly violate the tolerance ({min_} > {tolerance_s=}) inside episode range."
             "This might be due to synchronization issues with timestamps during data collection."
         )
 
@@ -207,144 +218,18 @@ def load_previous_and_future_frames(
 
         # load frames modality
         item[key] = hf_dataset.select_columns(key)[data_ids][key]
-        item[key] = torch.stack(item[key])
+
+        if isinstance(item[key][0], dict) and "path" in item[key][0]:
+            # video mode where frame are expressed as dict of path and timestamp
+            item[key] = item[key]
+        else:
+            item[key] = torch.stack(item[key])
+
         item[f"{key}_is_pad"] = is_pad
 
     return item
 
 
-def get_stats_einops_patterns(hf_dataset):
-    """These einops patterns will be used to aggregate batches and compute statistics.
-
-    Note: We assume the images of `hf_dataset` are in channel first format
-    """
-
-    dataloader = torch.utils.data.DataLoader(
-        hf_dataset,
-        num_workers=0,
-        batch_size=2,
-        shuffle=False,
-    )
-    batch = next(iter(dataloader))
-
-    stats_patterns = {}
-    for key, feats_type in hf_dataset.features.items():
-        # sanity check that tensors are not float64
-        assert batch[key].dtype != torch.float64
-
-        if isinstance(feats_type, Image):
-            # sanity check that images are channel first
-            _, c, h, w = batch[key].shape
-            assert c < h and c < w, f"expect channel first images, but instead {batch[key].shape}"
-
-            # sanity check that images are float32 in range [0,1]
-            assert batch[key].dtype == torch.float32, f"expect torch.float32, but instead {batch[key].dtype=}"
-            assert batch[key].max() <= 1, f"expect pixels lower than 1, but instead {batch[key].max()=}"
-            assert batch[key].min() >= 0, f"expect pixels greater than 1, but instead {batch[key].min()=}"
-
-            stats_patterns[key] = "b c h w -> c 1 1"
-        elif batch[key].ndim == 2:
-            stats_patterns[key] = "b c -> c "
-        elif batch[key].ndim == 1:
-            stats_patterns[key] = "b -> 1"
-        else:
-            raise ValueError(f"{key}, {feats_type}, {batch[key].shape}")
-
-    return stats_patterns
-
-
-def compute_stats(hf_dataset, batch_size=32, max_num_samples=None):
-    if max_num_samples is None:
-        max_num_samples = len(hf_dataset)
-
-    stats_patterns = get_stats_einops_patterns(hf_dataset)
-
-    # mean and std will be computed incrementally while max and min will track the running value.
-    mean, std, max, min = {}, {}, {}, {}
-    for key in stats_patterns:
-        mean[key] = torch.tensor(0.0).float()
-        std[key] = torch.tensor(0.0).float()
-        max[key] = torch.tensor(-float("inf")).float()
-        min[key] = torch.tensor(float("inf")).float()
-
-    def create_seeded_dataloader(hf_dataset, batch_size, seed):
-        generator = torch.Generator()
-        generator.manual_seed(seed)
-        dataloader = torch.utils.data.DataLoader(
-            hf_dataset,
-            num_workers=4,
-            batch_size=batch_size,
-            shuffle=True,
-            drop_last=False,
-            generator=generator,
-        )
-        return dataloader
-
-    # Note: Due to be refactored soon. The point of storing `first_batch` is to make sure we don't get
-    # surprises when rerunning the sampler.
-    first_batch = None
-    running_item_count = 0  # for online mean computation
-    dataloader = create_seeded_dataloader(hf_dataset, batch_size, seed=1337)
-    for i, batch in enumerate(
-        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute mean, min, max")
-    ):
-        this_batch_size = len(batch["index"])
-        running_item_count += this_batch_size
-        if first_batch is None:
-            first_batch = deepcopy(batch)
-        for key, pattern in stats_patterns.items():
-            batch[key] = batch[key].float()
-            # Numerically stable update step for mean computation.
-            batch_mean = einops.reduce(batch[key], pattern, "mean")
-            # Hint: to update the mean we need x̄ₙ = (Nₙ₋₁x̄ₙ₋₁ + Bₙxₙ) / Nₙ, where the subscript represents
-            # the update step, N is the running item count, B is this batch size, x̄ is the running mean,
-            # and x is the current batch mean. Some rearrangement is then required to avoid risking
-            # numerical overflow. Another hint: Nₙ₋₁ = Nₙ - Bₙ. Rearrangement yields
-            # x̄ₙ = x̄ₙ₋₁ + Bₙ * (xₙ - x̄ₙ₋₁) / Nₙ
-            mean[key] = mean[key] + this_batch_size * (batch_mean - mean[key]) / running_item_count
-            max[key] = torch.maximum(max[key], einops.reduce(batch[key], pattern, "max"))
-            min[key] = torch.minimum(min[key], einops.reduce(batch[key], pattern, "min"))
-
-        if i == ceil(max_num_samples / batch_size) - 1:
-            break
-
-    first_batch_ = None
-    running_item_count = 0  # for online std computation
-    dataloader = create_seeded_dataloader(hf_dataset, batch_size, seed=1337)
-    for i, batch in enumerate(
-        tqdm.tqdm(dataloader, total=ceil(max_num_samples / batch_size), desc="Compute std")
-    ):
-        this_batch_size = len(batch["index"])
-        running_item_count += this_batch_size
-        # Sanity check to make sure the batches are still in the same order as before.
-        if first_batch_ is None:
-            first_batch_ = deepcopy(batch)
-            for key in stats_patterns:
-                assert torch.equal(first_batch_[key], first_batch[key])
-        for key, pattern in stats_patterns.items():
-            batch[key] = batch[key].float()
-            # Numerically stable update step for mean computation (where the mean is over squared
-            # residuals).See notes in the mean computation loop above.
-            batch_std = einops.reduce((batch[key] - mean[key]) ** 2, pattern, "mean")
-            std[key] = std[key] + this_batch_size * (batch_std - std[key]) / running_item_count
-
-        if i == ceil(max_num_samples / batch_size) - 1:
-            break
-
-    for key in stats_patterns:
-        std[key] = torch.sqrt(std[key])
-
-    stats = {}
-    for key in stats_patterns:
-        stats[key] = {
-            "mean": mean[key],
-            "std": std[key],
-            "max": max[key],
-            "min": min[key],
-        }
-    return stats
-
-
 def cycle(iterable):
     """The equivalent of itertools.cycle, but safe for Pytorch dataloaders.
 

lerobot/common/datasets/video_utils.py

@@ -0,0 +1,187 @@
+import logging
+import subprocess
+import warnings
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, ClassVar
+
+import pyarrow as pa
+import torch
+import torchvision
+from datasets.features.features import register_feature
+
+
+def load_from_videos(
+    item: dict[str, torch.Tensor], video_frame_keys: list[str], videos_dir: Path, tolerance_s: float
+):
+    """Note: When using data workers (e.g. DataLoader with num_workers>0), do not call this function
+    in the main process (e.g. by using a second Dataloader with num_workers=0). It will result in a 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.
+    """
+    # since video path already contains "videos" (e.g. videos_dir="data/videos", path="videos/episode_0.mp4")
+    data_dir = videos_dir.parent
+
+    for key in video_frame_keys:
+        if isinstance(item[key], list):
+            # load multiple frames at once (expected when delta_timestamps is not None)
+            timestamps = [frame["timestamp"] for frame in item[key]]
+            paths = [frame["path"] for frame in item[key]]
+            if len(set(paths)) > 1:
+                raise NotImplementedError("All video paths are expected to be the same for now.")
+            video_path = data_dir / paths[0]
+
+            frames = decode_video_frames_torchvision(video_path, timestamps, tolerance_s)
+            item[key] = frames
+        else:
+            # load one frame
+            timestamps = [item[key]["timestamp"]]
+            video_path = data_dir / item[key]["path"]
+
+            frames = decode_video_frames_torchvision(video_path, timestamps, tolerance_s)
+            item[key] = frames[0]
+
+    return item
+
+
+def decode_video_frames_torchvision(
+    video_path: str,
+    timestamps: list[float],
+    tolerance_s: float,
+    device: str = "cpu",
+    log_loaded_timestamps: bool = False,
+):
+    """Loads frames associated to the requested timestamps of a video
+
+    Note: Video benefits from inter-frame compression. Instead of storing every frame individually,
+    the encoder stores a reference frame (or a key frame) and subsequent frames as differences relative to
+    that key frame. As a consequence, to access a requested frame, we need to load the preceding key frame,
+    and all subsequent frames until reaching the requested frame. The number of key frames in a video
+    can be adjusted during encoding to take into account decoding time and video size in bytes.
+    """
+    video_path = str(video_path)
+
+    # set backend
+    keyframes_only = False
+    if device == "cpu":
+        # explicitely use pyav
+        torchvision.set_video_backend("pyav")
+        keyframes_only = True  # pyav doesnt support accuracte seek
+    elif device == "cuda":
+        # TODO(rcadene, aliberts): implement video decoding with GPU
+        # torchvision.set_video_backend("cuda")
+        # torchvision.set_video_backend("video_reader")
+        # requires installing torchvision from source, see: https://github.com/pytorch/vision/blob/main/torchvision/csrc/io/decoder/gpu/README.rst
+        # check possible bug: https://github.com/pytorch/vision/issues/7745
+        raise NotImplementedError(
+            "Video decoding on gpu with cuda is currently not supported. Use `device='cpu'`."
+        )
+    else:
+        raise ValueError(device)
+
+    # set a video stream reader
+    # TODO(rcadene): also load audio stream at the same time
+    reader = torchvision.io.VideoReader(video_path, "video")
+
+    # set the first and last requested timestamps
+    # Note: previous timestamps are usually loaded, since we need to access the previous key frame
+    first_ts = timestamps[0]
+    last_ts = timestamps[-1]
+
+    # access closest key frame of the first requested frame
+    # Note: closest key frame timestamp is usally smaller than `first_ts` (e.g. key frame can be the first frame of the video)
+    # for details on what `seek` is doing see: https://pyav.basswood-io.com/docs/stable/api/container.html?highlight=inputcontainer#av.container.InputContainer.seek
+    reader.seek(first_ts, keyframes_only=keyframes_only)
+
+    # load all frames until last requested frame
+    loaded_frames = []
+    loaded_ts = []
+    for frame in reader:
+        current_ts = frame["pts"]
+        if log_loaded_timestamps:
+            logging.info(f"frame loaded at timestamp={current_ts:.4f}")
+        loaded_frames.append(frame["data"])
+        loaded_ts.append(current_ts)
+        if current_ts >= last_ts:
+            break
+
+    reader.container.close()
+    reader = None
+
+    query_ts = torch.tensor(timestamps)
+    loaded_ts = torch.tensor(loaded_ts)
+
+    # compute distances between each query timestamp and timestamps of all loaded frames
+    dist = torch.cdist(query_ts[:, None], loaded_ts[:, None], p=1)
+    min_, argmin_ = dist.min(1)
+
+    is_within_tol = min_ < tolerance_s
+    assert is_within_tol.all(), (
+        f"One or several query timestamps unexpectedly violate the tolerance ({min_[~is_within_tol]} > {tolerance_s=})."
+        "It means that the closest frame that can be loaded from the video is too far away in time."
+        "This might be due to synchronization issues with timestamps during data collection."
+        "To be safe, we advise to ignore this item during training."
+    )
+
+    # get closest frames to the query timestamps
+    closest_frames = torch.stack([loaded_frames[idx] for idx in argmin_])
+    closest_ts = loaded_ts[argmin_]
+
+    if log_loaded_timestamps:
+        logging.info(f"{closest_ts=}")
+
+    # convert to the pytorch format which is float32 in [0,1] range (and channel first)
+    closest_frames = closest_frames.type(torch.float32) / 255
+
+    assert len(timestamps) == len(closest_frames)
+    return closest_frames
+
+
+def encode_video_frames(imgs_dir: Path, video_path: Path, fps: int):
+    """More info on ffmpeg arguments tuning on `lerobot/common/datasets/_video_benchmark/README.md`"""
+    video_path = Path(video_path)
+    video_path.parent.mkdir(parents=True, exist_ok=True)
+
+    ffmpeg_cmd = (
+        f"ffmpeg -r {fps} "
+        "-f image2 "
+        "-loglevel error "
+        f"-i {str(imgs_dir / 'frame_%06d.png')} "
+        "-vcodec libx264 "
+        "-g 2 "
+        "-pix_fmt yuv444p "
+        f"{str(video_path)}"
+    )
+    subprocess.run(ffmpeg_cmd.split(" "), check=True)
+
+
+@dataclass
+class VideoFrame:
+    # TODO(rcadene, lhoestq): move to Hugging Face `datasets` repo
+    """
+    Provides a type for a dataset containing video frames.
+
+    Example:
+
+    ```python
+    data_dict = [{"image": {"path": "videos/episode_0.mp4", "timestamp": 0.3}}]
+    features = {"image": VideoFrame()}
+    Dataset.from_dict(data_dict, features=Features(features))
+    ```
+    """
+
+    pa_type: ClassVar[Any] = pa.struct({"path": pa.string(), "timestamp": pa.float32()})
+    _type: str = field(default="VideoFrame", init=False, repr=False)
+
+    def __call__(self):
+        return self.pa_type
+
+
+with warnings.catch_warnings():
+    warnings.filterwarnings(
+        "ignore",
+        "'register_feature' is experimental and might be subject to breaking changes in the future.",
+        category=UserWarning,
+    )
+    # to make VideoFrame available in HuggingFace `datasets`
+    register_feature(VideoFrame, "VideoFrame")

lerobot/common/envs/factory.py

@@ -1,13 +1,17 @@
 import importlib
 
 import gymnasium as gym
+from omegaconf import DictConfig
 
 
-def make_env(cfg, num_parallel_envs=0) -> gym.Env | gym.vector.SyncVectorEnv:
-    """
-    Note: When `num_parallel_envs > 0`, this function returns a `SyncVectorEnv` which takes batched action as input and
-    returns batched observation, reward, terminated, truncated of `num_parallel_envs` items.
+def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv:
+    """Makes a gym vector environment according to the evaluation config.
+
+    n_envs can be used to override eval.batch_size in the configuration. Must be at least 1.
     """
+    if n_envs is not None and n_envs < 1:
+        raise ValueError("`n_envs must be at least 1")
+
     kwargs = {
         "obs_type": "pixels_agent_pos",
         "render_mode": "rgb_array",
@@ -28,16 +32,13 @@ def make_env(cfg, num_parallel_envs=0) -> gym.Env | gym.vector.SyncVectorEnv:
 
     gym_handle = f"{package_name}/{cfg.env.task}"
 
-    if num_parallel_envs == 0:
-        # non-batched version of the env that returns an observation of shape (c)
-        env = gym.make(gym_handle, disable_env_checker=True, **kwargs)
-    else:
-        # batched version of the env that returns an observation of shape (b, c)
-        env = gym.vector.SyncVectorEnv(
-            [
-                lambda: gym.make(gym_handle, disable_env_checker=True, **kwargs)
-                for _ in range(num_parallel_envs)
-            ]
-        )
+    # batched version of the env that returns an observation of shape (b, c)
+    env_cls = gym.vector.AsyncVectorEnv if cfg.eval.use_async_envs else gym.vector.SyncVectorEnv
+    env = env_cls(
+        [
+            lambda: gym.make(gym_handle, disable_env_checker=True, **kwargs)
+            for _ in range(n_envs if n_envs is not None else cfg.eval.batch_size)
+        ]
+    )
 
     return env

lerobot/common/envs/utils.py

@@ -1,15 +1,23 @@
 import einops
+import numpy as np
 import torch
+from torch import Tensor
 
 
-def preprocess_observation(observation):
+def preprocess_observation(observations: dict[str, np.ndarray]) -> dict[str, Tensor]:
+    """Convert environment observation to LeRobot format observation.
+    Args:
+        observation: Dictionary of observation batches from a Gym vector environment.
+    Returns:
+        Dictionary of observation batches with keys renamed to LeRobot format and values as tensors.
+    """
     # map to expected inputs for the policy
-    obs = {}
+    return_observations = {}
 
-    if isinstance(observation["pixels"], dict):
-        imgs = {f"observation.images.{key}": img for key, img in observation["pixels"].items()}
+    if isinstance(observations["pixels"], dict):
+        imgs = {f"observation.images.{key}": img for key, img in observations["pixels"].items()}
     else:
-        imgs = {"observation.image": observation["pixels"]}
+        imgs = {"observation.image": observations["pixels"]}
 
     for imgkey, img in imgs.items():
         img = torch.from_numpy(img)
@@ -26,17 +34,10 @@ def preprocess_observation(observation):
         img = img.type(torch.float32)
         img /= 255
 
-        obs[imgkey] = img
+        return_observations[imgkey] = img
 
-    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing requirement for "agent_pos"
-    obs["observation.state"] = torch.from_numpy(observation["agent_pos"]).float()
+    # TODO(rcadene): enable pixels only baseline with `obs_type="pixels"` in environment by removing
+    # requirement for "agent_pos"
+    return_observations["observation.state"] = torch.from_numpy(observations["agent_pos"]).float()
 
-    return obs
-
-
-def postprocess_action(action):
-    action = action.to("cpu").numpy()
-    assert (
-        action.ndim == 2
-    ), "we assume dimensions are respectively the number of parallel envs, action dimensions"
-    return action
+    return return_observations

lerobot/common/logger.py

@@ -1,19 +1,26 @@
+# TODO(rcadene, alexander-soare): clean this file
+"""Borrowed from https://github.com/fyhMer/fowm/blob/main/src/logger.py"""
+
 import logging
 import os
 from pathlib import Path
 
+from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
 from omegaconf import OmegaConf
 from termcolor import colored
 
+from lerobot.common.policies.policy_protocol import Policy
+
 
 def log_output_dir(out_dir):
     logging.info(colored("Output dir:", "yellow", attrs=["bold"]) + f" {out_dir}")
 
 
 def cfg_to_group(cfg, return_list=False):
-    """Return a wandb-safe group name for logging. Optionally returns group name as list."""
-    # lst = [cfg.task, cfg.modality, re.sub("[^0-9a-zA-Z]+", "-", cfg.exp_name)]
+    """Return a group name for logging. Optionally returns group name as list."""
     lst = [
+        f"policy:{cfg.policy.name}",
+        f"dataset:{cfg.dataset_repo_id}",
         f"env:{cfg.env.name}",
         f"seed:{cfg.seed}",
     ]
@@ -27,11 +34,11 @@ class Logger:
         self._log_dir = Path(log_dir)
         self._log_dir.mkdir(parents=True, exist_ok=True)
         self._job_name = job_name
-        self._model_dir = self._log_dir / "models"
+        self._model_dir = self._log_dir / "checkpoints"
         self._buffer_dir = self._log_dir / "buffers"
-        self._save_model = cfg.save_model
+        self._save_model = cfg.training.save_model
         self._disable_wandb_artifact = cfg.wandb.disable_artifact
-        self._save_buffer = cfg.save_buffer
+        self._save_buffer = cfg.training.get("save_buffer", False)
         self._group = cfg_to_group(cfg)
         self._seed = cfg.seed
         self._cfg = cfg
@@ -67,18 +74,24 @@ class Logger:
             logging.info(f"Track this run --> {colored(wandb.run.get_url(), 'yellow', attrs=['bold'])}")
             self._wandb = wandb
 
-    def save_model(self, policy, identifier):
+    def save_model(self, policy: Policy, identifier):
         if self._save_model:
             self._model_dir.mkdir(parents=True, exist_ok=True)
-            fp = self._model_dir / f"{str(identifier)}.pt"
-            policy.save(fp)
+            save_dir = self._model_dir / str(identifier)
+            policy.save_pretrained(save_dir)
+            # Also save the full Hydra config for the env configuration.
+            OmegaConf.save(self._cfg, save_dir / "config.yaml")
             if self._wandb and not self._disable_wandb_artifact:
                 # note wandb artifact does not accept ":" in its name
                 artifact = self._wandb.Artifact(
-                    self._group.replace(":", "_") + "-" + str(self._seed) + "-" + str(identifier),
+                    self._group.replace(":", "_").replace("/", "__")
+                    + "-"
+                    + str(self._seed)
+                    + "-"
+                    + str(identifier),
                     type="model",
                 )
-                artifact.add_file(fp)
+                artifact.add_file(save_dir / SAFETENSORS_SINGLE_FILE)
                 self._wandb.log_artifact(artifact)
 
     def save_buffer(self, buffer, identifier):
@@ -107,7 +120,7 @@ class Logger:
             for k, v in d.items():
                 self._wandb.log({f"{mode}/{k}": v}, step=step)
 
-    def log_video(self, video, step, mode="train"):
+    def log_video(self, video_path: str, step: int, mode: str = "train"):
         assert mode in {"train", "eval"}
-        wandb_video = self._wandb.Video(video, fps=self._cfg.fps, format="mp4")
+        wandb_video = self._wandb.Video(video_path, fps=self._cfg.fps, format="mp4")
         self._wandb.log({f"{mode}/video": wandb_video}, step=step)

lerobot/common/policies/act/configuration_act.py

@@ -2,7 +2,7 @@ from dataclasses import dataclass, field
 
 
 @dataclass
-class ActionChunkingTransformerConfig:
+class ACTConfig:
     """Configuration class for the Action Chunking Transformers policy.
 
     Defaults are configured for training on bimanual Aloha tasks like "insertion" or "transfer".
@@ -22,23 +22,24 @@ class ActionChunkingTransformerConfig:
             The key represents the input data name, and the value is a list indicating the dimensions
             of the corresponding data. For example, "observation.images.top" refers to an input from the
             "top" camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
-            Importantly, shapes doesnt include batch dimension or temporal dimension.
+            Importantly, shapes doesn't include batch dimension or temporal dimension.
         output_shapes: A dictionary defining the shapes of the output data for the policy.
             The key represents the output data name, and the value is a list indicating the dimensions
             of the corresponding data. For example, "action" refers to an output shape of [14], indicating
-            14-dimensional actions. Importantly, shapes doesnt include batch dimension or temporal dimension.
-        normalize_input_modes: A dictionary with key represents the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two availables
-            modes are "mean_std" which substracts the mean and divide by the standard
-            deviation and "min_max" which rescale in a [-1, 1] range.
-        unnormalize_output_modes: Similar dictionary as `normalize_input_modes`, but to unormalize in original scale.
+            14-dimensional actions. Importantly, shapes doesn't include batch dimension or temporal dimension.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         pretrained_backbone_weights: Pretrained weights from torchvision to initalize the backbone.
             `None` means no pretrained weights.
         replace_final_stride_with_dilation: Whether to replace the ResNet's final 2x2 stride with a dilated
             convolution.
         pre_norm: Whether to use "pre-norm" in the transformer blocks.
-        d_model: The transformer blocks' main hidden dimension.
+        dim_model: The transformer blocks' main hidden dimension.
         n_heads: The number of heads to use in the transformer blocks' multi-head attention.
         dim_feedforward: The dimension to expand the transformer's hidden dimension to in the feed-forward
             layers.
@@ -62,13 +63,13 @@ class ActionChunkingTransformerConfig:
     chunk_size: int = 100
     n_action_steps: int = 100
 
-    input_shapes: dict[str, list[str]] = field(
+    input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
             "observation.images.top": [3, 480, 640],
             "observation.state": [14],
         }
     )
-    output_shapes: dict[str, list[str]] = field(
+    output_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
             "action": [14],
         }
@@ -94,7 +95,7 @@ class ActionChunkingTransformerConfig:
     replace_final_stride_with_dilation: int = False
     # Transformer layers.
     pre_norm: bool = False
-    d_model: int = 512
+    dim_model: int = 512
     n_heads: int = 8
     dim_feedforward: int = 3200
     feedforward_activation: str = "relu"
@@ -112,15 +113,6 @@ class ActionChunkingTransformerConfig:
     dropout: float = 0.1
     kl_weight: float = 10.0
 
-    # ---
-    # TODO(alexander-soare): Remove these from the policy config.
-    batch_size: int = 8
-    lr: float = 1e-5
-    lr_backbone: float = 1e-5
-    weight_decay: float = 1e-4
-    grad_clip_norm: float = 10
-    utd: int = 1
-
     def __post_init__(self):
         """Input validation (not exhaustive)."""
         if not self.vision_backbone.startswith("resnet"):

lerobot/common/policies/act/modeling_act.py

@@ -5,7 +5,6 @@ The majority of changes here involve removing unused code, unifying naming, and
 """
 
 import math
-import time
 from collections import deque
 from itertools import chain
 from typing import Callable
@@ -15,18 +14,127 @@ import numpy as np
 import torch
 import torch.nn.functional as F  # noqa: N812
 import torchvision
+from huggingface_hub import PyTorchModelHubMixin
 from torch import Tensor, nn
 from torchvision.models._utils import IntermediateLayerGetter
 from torchvision.ops.misc import FrozenBatchNorm2d
 
-from lerobot.common.policies.act.configuration_act import ActionChunkingTransformerConfig
+from lerobot.common.policies.act.configuration_act import ACTConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 
 
-class ActionChunkingTransformerPolicy(nn.Module):
+class ACTPolicy(nn.Module, PyTorchModelHubMixin):
     """
     Action Chunking Transformer Policy as per Learning Fine-Grained Bimanual Manipulation with Low-Cost
     Hardware (paper: https://arxiv.org/abs/2304.13705, code: https://github.com/tonyzhaozh/act)
+    """
+
+    name = "act"
+
+    def __init__(
+        self,
+        config: ACTConfig | None = None,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                    the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+        super().__init__()
+        if config is None:
+            config = ACTConfig()
+        self.config = config
+        self.normalize_inputs = Normalize(
+            config.input_shapes, config.input_normalization_modes, dataset_stats
+        )
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.model = ACT(config)
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        if self.config.n_action_steps is not None:
+            self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    @torch.no_grad
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method wraps `select_actions` in order to return one action at a time for execution in the
+        environment. It works by managing the actions in a queue and only calling `select_actions` when the
+        queue is empty.
+        """
+        assert "observation.images.top" in batch
+        assert "observation.state" in batch
+
+        self.eval()
+
+        batch = self.normalize_inputs(batch)
+        self._stack_images(batch)
+
+        if len(self._action_queue) == 0:
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
+            actions = self.model(batch)[0][: self.config.n_action_steps]
+
+            # TODO(rcadene): make _forward return output dictionary?
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        """Run the batch through the model and compute the loss for training or validation."""
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
+        self._stack_images(batch)
+        actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
+
+        l1_loss = (
+            F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
+        ).mean()
+
+        loss_dict = {"l1_loss": l1_loss}
+        if self.config.use_vae:
+            # Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for
+            # each dimension independently, we sum over the latent dimension to get the total
+            # KL-divergence per batch element, then take the mean over the batch.
+            # (See App. B of https://arxiv.org/abs/1312.6114 for more details).
+            mean_kld = (
+                (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
+            )
+            loss_dict["kld_loss"] = mean_kld
+            loss_dict["loss"] = l1_loss + mean_kld * self.config.kl_weight
+        else:
+            loss_dict["loss"] = l1_loss
+
+        return loss_dict
+
+    def _stack_images(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        """Stacks all the images in a batch and puts them in a new key: "observation.images".
+
+        This function expects `batch` to have (at least):
+        {
+            "observation.state": (B, state_dim) batch of robot states.
+            "observation.images.{name}": (B, C, H, W) tensor of images.
+        }
+        """
+        # Stack images in the order dictated by input_shapes.
+        batch["observation.images"] = torch.stack(
+            [batch[k] for k in self.config.input_shapes if k.startswith("observation.images.")],
+            dim=-4,
+        )
+
+
+class ACT(nn.Module):
+    """Action Chunking Transformer: The underlying neural network for ACTPolicy.
 
     Note: In this code we use the terms `vae_encoder`, 'encoder', `decoder`. The meanings are as follows.
         - The `vae_encoder` is, as per the literature around variational auto-encoders (VAE), the part of the
@@ -60,51 +168,36 @@ class ActionChunkingTransformerPolicy(nn.Module):
                                 └───────────────────────┘
     """
 
-    name = "act"
-
-    def __init__(self, cfg: ActionChunkingTransformerConfig | None = None, dataset_stats=None):
-        """
-        Args:
-            cfg: Policy configuration class instance or None, in which case the default instantiation of the
-                 configuration class is used.
-        """
+    def __init__(self, config: ACTConfig):
         super().__init__()
-        if cfg is None:
-            cfg = ActionChunkingTransformerConfig()
-        self.cfg = cfg
-        self.normalize_inputs = Normalize(cfg.input_shapes, cfg.input_normalization_modes, dataset_stats)
-        self.normalize_targets = Normalize(cfg.output_shapes, cfg.output_normalization_modes, dataset_stats)
-        self.unnormalize_outputs = Unnormalize(
-            cfg.output_shapes, cfg.output_normalization_modes, dataset_stats
-        )
-
+        self.config = config
         # BERT style VAE encoder with input [cls, *joint_space_configuration, *action_sequence].
         # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
-        if self.cfg.use_vae:
-            self.vae_encoder = _TransformerEncoder(cfg)
-            self.vae_encoder_cls_embed = nn.Embedding(1, cfg.d_model)
+        if self.config.use_vae:
+            self.vae_encoder = ACTEncoder(config)
+            self.vae_encoder_cls_embed = nn.Embedding(1, config.dim_model)
             # Projection layer for joint-space configuration to hidden dimension.
             self.vae_encoder_robot_state_input_proj = nn.Linear(
-                cfg.input_shapes["observation.state"][0], cfg.d_model
+                config.input_shapes["observation.state"][0], config.dim_model
             )
             # Projection layer for action (joint-space target) to hidden dimension.
             self.vae_encoder_action_input_proj = nn.Linear(
-                cfg.input_shapes["observation.state"][0], cfg.d_model
+                config.input_shapes["observation.state"][0], config.dim_model
             )
-            self.latent_dim = cfg.latent_dim
+            self.latent_dim = config.latent_dim
             # Projection layer from the VAE encoder's output to the latent distribution's parameter space.
-            self.vae_encoder_latent_output_proj = nn.Linear(cfg.d_model, self.latent_dim * 2)
+            self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, self.latent_dim * 2)
             # Fixed sinusoidal positional embedding the whole input to the VAE encoder. Unsqueeze for batch
             # dimension.
             self.register_buffer(
                 "vae_encoder_pos_enc",
-                _create_sinusoidal_position_embedding(1 + 1 + cfg.chunk_size, cfg.d_model).unsqueeze(0),
+                create_sinusoidal_pos_embedding(1 + 1 + config.chunk_size, config.dim_model).unsqueeze(0),
             )
 
         # Backbone for image feature extraction.
-        backbone_model = getattr(torchvision.models, cfg.vision_backbone)(
-            replace_stride_with_dilation=[False, False, cfg.replace_final_stride_with_dilation],
-            weights=cfg.pretrained_backbone_weights,
+        backbone_model = getattr(torchvision.models, config.vision_backbone)(
+            replace_stride_with_dilation=[False, False, config.replace_final_stride_with_dilation],
+            weights=config.pretrained_backbone_weights,
             norm_layer=FrozenBatchNorm2d,
         )
         # Note: The assumption here is that we are using a ResNet model (and hence layer4 is the final feature
@@ -113,47 +206,30 @@ class ActionChunkingTransformerPolicy(nn.Module):
         self.backbone = IntermediateLayerGetter(backbone_model, return_layers={"layer4": "feature_map"})
 
         # Transformer (acts as VAE decoder when training with the variational objective).
-        self.encoder = _TransformerEncoder(cfg)
-        self.decoder = _TransformerDecoder(cfg)
+        self.encoder = ACTEncoder(config)
+        self.decoder = ACTDecoder(config)
 
         # Transformer encoder input projections. The tokens will be structured like
         # [latent, robot_state, image_feature_map_pixels].
-        self.encoder_robot_state_input_proj = nn.Linear(cfg.input_shapes["observation.state"][0], cfg.d_model)
-        self.encoder_latent_input_proj = nn.Linear(self.latent_dim, cfg.d_model)
+        self.encoder_robot_state_input_proj = nn.Linear(
+            config.input_shapes["observation.state"][0], config.dim_model
+        )
+        self.encoder_latent_input_proj = nn.Linear(self.latent_dim, config.dim_model)
         self.encoder_img_feat_input_proj = nn.Conv2d(
-            backbone_model.fc.in_features, cfg.d_model, kernel_size=1
+            backbone_model.fc.in_features, config.dim_model, kernel_size=1
         )
         # Transformer encoder positional embeddings.
-        self.encoder_robot_and_latent_pos_embed = nn.Embedding(2, cfg.d_model)
-        self.encoder_cam_feat_pos_embed = _SinusoidalPositionEmbedding2D(cfg.d_model // 2)
+        self.encoder_robot_and_latent_pos_embed = nn.Embedding(2, config.dim_model)
+        self.encoder_cam_feat_pos_embed = ACTSinusoidalPositionEmbedding2d(config.dim_model // 2)
 
         # Transformer decoder.
         # Learnable positional embedding for the transformer's decoder (in the style of DETR object queries).
-        self.decoder_pos_embed = nn.Embedding(cfg.chunk_size, cfg.d_model)
+        self.decoder_pos_embed = nn.Embedding(config.chunk_size, config.dim_model)
 
         # Final action regression head on the output of the transformer's decoder.
-        self.action_head = nn.Linear(cfg.d_model, cfg.output_shapes["action"][0])
+        self.action_head = nn.Linear(config.dim_model, config.output_shapes["action"][0])
 
         self._reset_parameters()
-        self._create_optimizer()
-
-    def _create_optimizer(self):
-        optimizer_params_dicts = [
-            {
-                "params": [
-                    p for n, p in self.named_parameters() if not n.startswith("backbone") and p.requires_grad
-                ]
-            },
-            {
-                "params": [
-                    p for n, p in self.named_parameters() if n.startswith("backbone") and p.requires_grad
-                ],
-                "lr": self.cfg.lr_backbone,
-            },
-        ]
-        self.optimizer = torch.optim.AdamW(
-            optimizer_params_dicts, lr=self.cfg.lr, weight_decay=self.cfg.weight_decay
-        )
 
     def _reset_parameters(self):
         """Xavier-uniform initialization of the transformer parameters as in the original code."""
@@ -161,103 +237,7 @@ class ActionChunkingTransformerPolicy(nn.Module):
             if p.dim() > 1:
                 nn.init.xavier_uniform_(p)
 
-    def reset(self):
-        """This should be called whenever the environment is reset."""
-        if self.cfg.n_action_steps is not None:
-            self._action_queue = deque([], maxlen=self.cfg.n_action_steps)
-
-    @torch.no_grad
-    def select_action(self, batch: dict[str, Tensor], **_) -> Tensor:
-        """Select a single action given environment observations.
-
-        This method wraps `select_actions` in order to return one action at a time for execution in the
-        environment. It works by managing the actions in a queue and only calling `select_actions` when the
-        queue is empty.
-        """
-        self.eval()
-
-        batch = self.normalize_inputs(batch)
-
-        if len(self._action_queue) == 0:
-            # `_forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue effectively
-            # has shape (n_action_steps, batch_size, *), hence the transpose.
-            actions = self._forward(batch)[0][: self.cfg.n_action_steps]
-
-            # TODO(rcadene): make _forward return output dictionary?
-            actions = self.unnormalize_outputs({"action": actions})["action"]
-
-            self._action_queue.extend(actions.transpose(0, 1))
-        return self._action_queue.popleft()
-
-    def forward(self, batch, **_) -> dict[str, Tensor]:
-        """Run the batch through the model and compute the loss for training or validation."""
-        actions_hat, (mu_hat, log_sigma_x2_hat) = self._forward(batch)
-
-        l1_loss = (
-            F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
-        ).mean()
-
-        loss_dict = {"l1_loss": l1_loss}
-        if self.cfg.use_vae:
-            # Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for
-            # each dimension independently, we sum over the latent dimension to get the total
-            # KL-divergence per batch element, then take the mean over the batch.
-            # (See App. B of https://arxiv.org/abs/1312.6114 for more details).
-            mean_kld = (
-                (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
-            )
-            loss_dict["kld_loss"] = mean_kld
-            loss_dict["loss"] = l1_loss + mean_kld * self.cfg.kl_weight
-        else:
-            loss_dict["loss"] = l1_loss
-
-        return loss_dict
-
-    def update(self, batch, **_) -> dict:
-        """Run the model in train mode, compute the loss, and do an optimization step."""
-        start_time = time.time()
-        self.train()
-
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
-        loss_dict = self.forward(batch)
-        # TODO(rcadene): self.unnormalize_outputs(out_dict)
-        loss = loss_dict["loss"]
-        loss.backward()
-
-        grad_norm = torch.nn.utils.clip_grad_norm_(
-            self.parameters(), self.cfg.grad_clip_norm, error_if_nonfinite=False
-        )
-
-        self.optimizer.step()
-        self.optimizer.zero_grad()
-
-        info = {
-            "loss": loss.item(),
-            "grad_norm": float(grad_norm),
-            "lr": self.cfg.lr,
-            "update_s": time.time() - start_time,
-        }
-
-        return info
-
-    def _stack_images(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
-        """Stacks all the images in a batch and puts them in a new key: "observation.images".
-
-        This function expects `batch` to have (at least):
-        {
-            "observation.state": (B, state_dim) batch of robot states.
-            "observation.images.{name}": (B, C, H, W) tensor of images.
-        }
-        """
-        # Stack images in the order dictated by input_shapes.
-        batch["observation.images"] = torch.stack(
-            [batch[k] for k in self.cfg.input_shapes if k.startswith("observation.images.")],
-            dim=-4,
-        )
-
-    def _forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, tuple[Tensor, Tensor] | tuple[None, None]]:
+    def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, tuple[Tensor, Tensor] | tuple[None, None]]:
         """A forward pass through the Action Chunking Transformer (with optional VAE encoder).
 
         `batch` should have the following structure:
@@ -273,17 +253,15 @@ class ActionChunkingTransformerPolicy(nn.Module):
             Tuple containing the latent PDF's parameters (mean, log(σ²)) both as (B, L) tensors where L is the
             latent dimension.
         """
-        if self.cfg.use_vae and self.training:
+        if self.config.use_vae and self.training:
             assert (
                 "action" in batch
             ), "actions must be provided when using the variational objective in training mode."
 
-        self._stack_images(batch)
-
         batch_size = batch["observation.state"].shape[0]
 
         # Prepare the latent for input to the transformer encoder.
-        if self.cfg.use_vae and "action" in batch:
+        if self.config.use_vae and "action" in batch:
             # Prepare the input to the VAE encoder: [cls, *joint_space_configuration, *action_sequence].
             cls_embed = einops.repeat(
                 self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
@@ -353,7 +331,7 @@ class ActionChunkingTransformerPolicy(nn.Module):
         # Forward pass through the transformer modules.
         encoder_out = self.encoder(encoder_in, pos_embed=pos_embed)
         decoder_in = torch.zeros(
-            (self.cfg.chunk_size, batch_size, self.cfg.d_model),
+            (self.config.chunk_size, batch_size, self.config.dim_model),
             dtype=pos_embed.dtype,
             device=pos_embed.device,
         )
@@ -371,21 +349,14 @@ class ActionChunkingTransformerPolicy(nn.Module):
 
         return actions, (mu, log_sigma_x2)
 
-    def save(self, fp):
-        torch.save(self.state_dict(), fp)
 
-    def load(self, fp):
-        d = torch.load(fp)
-        self.load_state_dict(d)
-
-
-class _TransformerEncoder(nn.Module):
+class ACTEncoder(nn.Module):
     """Convenience module for running multiple encoder layers, maybe followed by normalization."""
 
-    def __init__(self, cfg: ActionChunkingTransformerConfig):
+    def __init__(self, config: ACTConfig):
         super().__init__()
-        self.layers = nn.ModuleList([_TransformerEncoderLayer(cfg) for _ in range(cfg.n_encoder_layers)])
-        self.norm = nn.LayerNorm(cfg.d_model) if cfg.pre_norm else nn.Identity()
+        self.layers = nn.ModuleList([ACTEncoderLayer(config) for _ in range(config.n_encoder_layers)])
+        self.norm = nn.LayerNorm(config.dim_model) if config.pre_norm else nn.Identity()
 
     def forward(self, x: Tensor, pos_embed: Tensor | None = None) -> Tensor:
         for layer in self.layers:
@@ -394,23 +365,23 @@ class _TransformerEncoder(nn.Module):
         return x
 
 
-class _TransformerEncoderLayer(nn.Module):
-    def __init__(self, cfg: ActionChunkingTransformerConfig):
+class ACTEncoderLayer(nn.Module):
+    def __init__(self, config: ACTConfig):
         super().__init__()
-        self.self_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
+        self.self_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
 
         # Feed forward layers.
-        self.linear1 = nn.Linear(cfg.d_model, cfg.dim_feedforward)
-        self.dropout = nn.Dropout(cfg.dropout)
-        self.linear2 = nn.Linear(cfg.dim_feedforward, cfg.d_model)
+        self.linear1 = nn.Linear(config.dim_model, config.dim_feedforward)
+        self.dropout = nn.Dropout(config.dropout)
+        self.linear2 = nn.Linear(config.dim_feedforward, config.dim_model)
 
-        self.norm1 = nn.LayerNorm(cfg.d_model)
-        self.norm2 = nn.LayerNorm(cfg.d_model)
-        self.dropout1 = nn.Dropout(cfg.dropout)
-        self.dropout2 = nn.Dropout(cfg.dropout)
+        self.norm1 = nn.LayerNorm(config.dim_model)
+        self.norm2 = nn.LayerNorm(config.dim_model)
+        self.dropout1 = nn.Dropout(config.dropout)
+        self.dropout2 = nn.Dropout(config.dropout)
 
-        self.activation = _get_activation_fn(cfg.feedforward_activation)
-        self.pre_norm = cfg.pre_norm
+        self.activation = get_activation_fn(config.feedforward_activation)
+        self.pre_norm = config.pre_norm
 
     def forward(self, x, pos_embed: Tensor | None = None) -> Tensor:
         skip = x
@@ -432,12 +403,12 @@ class _TransformerEncoderLayer(nn.Module):
         return x
 
 
-class _TransformerDecoder(nn.Module):
-    def __init__(self, cfg: ActionChunkingTransformerConfig):
+class ACTDecoder(nn.Module):
+    def __init__(self, config: ACTConfig):
         """Convenience module for running multiple decoder layers followed by normalization."""
         super().__init__()
-        self.layers = nn.ModuleList([_TransformerDecoderLayer(cfg) for _ in range(cfg.n_decoder_layers)])
-        self.norm = nn.LayerNorm(cfg.d_model)
+        self.layers = nn.ModuleList([ACTDecoderLayer(config) for _ in range(config.n_decoder_layers)])
+        self.norm = nn.LayerNorm(config.dim_model)
 
     def forward(
         self,
@@ -455,26 +426,26 @@ class _TransformerDecoder(nn.Module):
         return x
 
 
-class _TransformerDecoderLayer(nn.Module):
-    def __init__(self, cfg: ActionChunkingTransformerConfig):
+class ACTDecoderLayer(nn.Module):
+    def __init__(self, config: ACTConfig):
         super().__init__()
-        self.self_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
-        self.multihead_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
+        self.self_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
+        self.multihead_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
 
         # Feed forward layers.
-        self.linear1 = nn.Linear(cfg.d_model, cfg.dim_feedforward)
-        self.dropout = nn.Dropout(cfg.dropout)
-        self.linear2 = nn.Linear(cfg.dim_feedforward, cfg.d_model)
+        self.linear1 = nn.Linear(config.dim_model, config.dim_feedforward)
+        self.dropout = nn.Dropout(config.dropout)
+        self.linear2 = nn.Linear(config.dim_feedforward, config.dim_model)
 
-        self.norm1 = nn.LayerNorm(cfg.d_model)
-        self.norm2 = nn.LayerNorm(cfg.d_model)
-        self.norm3 = nn.LayerNorm(cfg.d_model)
-        self.dropout1 = nn.Dropout(cfg.dropout)
-        self.dropout2 = nn.Dropout(cfg.dropout)
-        self.dropout3 = nn.Dropout(cfg.dropout)
+        self.norm1 = nn.LayerNorm(config.dim_model)
+        self.norm2 = nn.LayerNorm(config.dim_model)
+        self.norm3 = nn.LayerNorm(config.dim_model)
+        self.dropout1 = nn.Dropout(config.dropout)
+        self.dropout2 = nn.Dropout(config.dropout)
+        self.dropout3 = nn.Dropout(config.dropout)
 
-        self.activation = _get_activation_fn(cfg.feedforward_activation)
-        self.pre_norm = cfg.pre_norm
+        self.activation = get_activation_fn(config.feedforward_activation)
+        self.pre_norm = config.pre_norm
 
     def maybe_add_pos_embed(self, tensor: Tensor, pos_embed: Tensor | None) -> Tensor:
         return tensor if pos_embed is None else tensor + pos_embed
@@ -527,7 +498,7 @@ class _TransformerDecoderLayer(nn.Module):
         return x
 
 
-def _create_sinusoidal_position_embedding(num_positions: int, dimension: int) -> Tensor:
+def create_sinusoidal_pos_embedding(num_positions: int, dimension: int) -> Tensor:
     """1D sinusoidal positional embeddings as in Attention is All You Need.
 
     Args:
@@ -545,7 +516,7 @@ def _create_sinusoidal_position_embedding(num_positions: int, dimension: int) ->
     return torch.from_numpy(sinusoid_table).float()
 
 
-class _SinusoidalPositionEmbedding2D(nn.Module):
+class ACTSinusoidalPositionEmbedding2d(nn.Module):
     """2D sinusoidal positional embeddings similar to what's presented in Attention Is All You Need.
 
     The variation is that the position indices are normalized in [0, 2π] (not quite: the lower bound is 1/H
@@ -599,7 +570,7 @@ class _SinusoidalPositionEmbedding2D(nn.Module):
         return pos_embed
 
 
-def _get_activation_fn(activation: str) -> Callable:
+def get_activation_fn(activation: str) -> Callable:
     """Return an activation function given a string."""
     if activation == "relu":
         return F.relu

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -3,7 +3,7 @@ from dataclasses import dataclass, field
 
 @dataclass
 class DiffusionConfig:
-    """Configuration class for Diffusion Policy.
+    """Configuration class for DiffusionPolicy.
 
     Defaults are configured for training with PushT providing proprioceptive and single camera observations.
 
@@ -25,11 +25,12 @@ class DiffusionConfig:
             The key represents the output data name, and the value is a list indicating the dimensions
             of the corresponding data. For example, "action" refers to an output shape of [14], indicating
             14-dimensional actions. Importantly, shapes doesnt include batch dimension or temporal dimension.
-        normalize_input_modes: A dictionary with key represents the modality (e.g. "observation.state"),
-            and the value specifies the normalization mode to apply. The two availables
-            modes are "mean_std" which substracts the mean and divide by the standard
-            deviation and "min_max" which rescale in a [-1, 1] range.
-        unnormalize_output_modes: Similar dictionary as `normalize_input_modes`, but to unormalize in original scale.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets.
         vision_backbone: Name of the torchvision resnet backbone to use for encoding images.
         crop_shape: (H, W) shape to crop images to as a preprocessing step for the vision backbone. Must fit
             within the image size. If None, no cropping is done.
@@ -63,6 +64,9 @@ class DiffusionConfig:
         clip_sample_range: The magnitude of the clipping range as described above.
         num_inference_steps: Number of reverse diffusion steps to use at inference time (steps are evenly
             spaced). If not provided, this defaults to be the same as `num_train_timesteps`.
+        do_mask_loss_for_padding: Whether to mask the loss when there are copy-padded actions. See
+            `LeRobotDataset` and `load_previous_and_future_frames` for mor information. Note, this defaults
+            to False as the original Diffusion Policy implementation does the same.
     """
 
     # Inputs / output structure.
@@ -70,13 +74,13 @@ class DiffusionConfig:
     horizon: int = 16
     n_action_steps: int = 8
 
-    input_shapes: dict[str, list[str]] = field(
+    input_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
             "observation.image": [3, 96, 96],
             "observation.state": [2],
         }
     )
-    output_shapes: dict[str, list[str]] = field(
+    output_shapes: dict[str, list[int]] = field(
         default_factory=lambda: {
             "action": [2],
         }
@@ -117,23 +121,8 @@ class DiffusionConfig:
     # Inference
     num_inference_steps: int | None = None
 
-    # ---
-    # TODO(alexander-soare): Remove these from the policy config.
-    batch_size: int = 64
-    grad_clip_norm: int = 10
-    lr: float = 1.0e-4
-    lr_scheduler: str = "cosine"
-    lr_warmup_steps: int = 500
-    adam_betas: tuple[float, float] = (0.95, 0.999)
-    adam_eps: float = 1.0e-8
-    adam_weight_decay: float = 1.0e-6
-    utd: int = 1
-    use_ema: bool = True
-    ema_update_after_step: int = 0
-    ema_min_alpha: float = 0.0
-    ema_max_alpha: float = 0.9999
-    ema_inv_gamma: float = 1.0
-    ema_power: float = 0.75
+    # Loss computation
+    do_mask_loss_for_padding: bool = False
 
     def __post_init__(self):
         """Input validation (not exhaustive)."""

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -3,15 +3,9 @@
 TODO(alexander-soare):
   - Remove reliance on Robomimic for SpatialSoftmax.
   - Remove reliance on diffusers for DDPMScheduler and LR scheduler.
-  - Move EMA out of policy.
-  - Consolidate _DiffusionUnetImagePolicy into DiffusionPolicy.
-  - One more pass on comments and documentation.
 """
 
-import copy
-import logging
 import math
-import time
 from collections import deque
 from typing import Callable
 
@@ -19,11 +13,10 @@ import einops
 import torch
 import torch.nn.functional as F  # noqa: N812
 import torchvision
-from diffusers.optimization import get_scheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
+from huggingface_hub import PyTorchModelHubMixin
 from robomimic.models.base_nets import SpatialSoftmax
 from torch import Tensor, nn
-from torch.nn.modules.batchnorm import _BatchNorm
 
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.normalize import Normalize, Unnormalize
@@ -34,7 +27,7 @@ from lerobot.common.policies.utils import (
 )
 
 
-class DiffusionPolicy(nn.Module):
+class DiffusionPolicy(nn.Module, PyTorchModelHubMixin):
     """
     Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
     (paper: https://arxiv.org/abs/2303.04137, code: https://github.com/real-stanford/diffusion_policy).
@@ -43,69 +36,48 @@ class DiffusionPolicy(nn.Module):
     name = "diffusion"
 
     def __init__(
-        self, cfg: DiffusionConfig | None = None, lr_scheduler_num_training_steps: int = 0, dataset_stats=None
+        self,
+        config: DiffusionConfig | None = None,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
     ):
         """
         Args:
-            cfg: Policy configuration class instance or None, in which case the default instantiation of the
-                 configuration class is used.
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
         """
         super().__init__()
-        # TODO(alexander-soare): LR scheduler will be removed.
-        assert lr_scheduler_num_training_steps > 0
-        if cfg is None:
-            cfg = DiffusionConfig()
-        self.cfg = cfg
-        self.normalize_inputs = Normalize(cfg.input_shapes, cfg.input_normalization_modes, dataset_stats)
-        self.normalize_targets = Normalize(cfg.output_shapes, cfg.output_normalization_modes, dataset_stats)
+        if config is None:
+            config = DiffusionConfig()
+        self.config = config
+        self.normalize_inputs = Normalize(
+            config.input_shapes, config.input_normalization_modes, dataset_stats
+        )
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
         self.unnormalize_outputs = Unnormalize(
-            cfg.output_shapes, cfg.output_normalization_modes, dataset_stats
+            config.output_shapes, config.output_normalization_modes, dataset_stats
         )
 
         # queues are populated during rollout of the policy, they contain the n latest observations and actions
         self._queues = None
 
-        self.diffusion = _DiffusionUnetImagePolicy(cfg)
-
-        # TODO(alexander-soare): This should probably be managed outside of the policy class.
-        self.ema_diffusion = None
-        self.ema = None
-        if self.cfg.use_ema:
-            self.ema_diffusion = copy.deepcopy(self.diffusion)
-            self.ema = _EMA(cfg, model=self.ema_diffusion)
-
-        # TODO(alexander-soare): Move optimizer out of policy.
-        self.optimizer = torch.optim.Adam(
-            self.diffusion.parameters(), cfg.lr, cfg.adam_betas, cfg.adam_eps, cfg.adam_weight_decay
-        )
-
-        # TODO(alexander-soare): Move LR scheduler out of policy.
-        # TODO(rcadene): modify lr scheduler so that it doesn't depend on epochs but steps
-        self.global_step = 0
-
-        # configure lr scheduler
-        self.lr_scheduler = get_scheduler(
-            cfg.lr_scheduler,
-            optimizer=self.optimizer,
-            num_warmup_steps=cfg.lr_warmup_steps,
-            num_training_steps=lr_scheduler_num_training_steps,
-            # pytorch assumes stepping LRScheduler every epoch
-            # however huggingface diffusers steps it every batch
-            last_epoch=self.global_step - 1,
-        )
+        self.diffusion = DiffusionModel(config)
 
     def reset(self):
         """
         Clear observation and action queues. Should be called on `env.reset()`
         """
         self._queues = {
-            "observation.image": deque(maxlen=self.cfg.n_obs_steps),
-            "observation.state": deque(maxlen=self.cfg.n_obs_steps),
-            "action": deque(maxlen=self.cfg.n_action_steps),
+            "observation.image": deque(maxlen=self.config.n_obs_steps),
+            "observation.state": deque(maxlen=self.config.n_obs_steps),
+            "action": deque(maxlen=self.config.n_action_steps),
         }
 
     @torch.no_grad
-    def select_action(self, batch: dict[str, Tensor], **_) -> Tensor:
+    def select_action(self, batch: dict[str, Tensor]) -> Tensor:
         """Select a single action given environment observations.
 
         This method handles caching a history of observations and an action trajectory generated by the
@@ -125,13 +97,9 @@ class DiffusionPolicy(nn.Module):
         Note that this means we require: `n_action_steps < horizon - n_obs_steps + 1`. Also, note that
         "horizon" may not the best name to describe what the variable actually means, because this period is
         actually measured from the first observation which (if `n_obs_steps` > 1) happened in the past.
-
-        Note: this method uses the ema model weights if self.training == False, otherwise the non-ema model
-        weights.
         """
         assert "observation.image" in batch
         assert "observation.state" in batch
-        assert len(batch) == 2
 
         batch = self.normalize_inputs(batch)
 
@@ -140,10 +108,7 @@ class DiffusionPolicy(nn.Module):
         if len(self._queues["action"]) == 0:
             # stack n latest observations from the queue
             batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
-            if not self.training and self.ema_diffusion is not None:
-                actions = self.ema_diffusion.generate_actions(batch)
-            else:
-                actions = self.diffusion.generate_actions(batch)
+            actions = self.diffusion.generate_actions(batch)
 
             # TODO(rcadene): make above methods return output dictionary?
             actions = self.unnormalize_outputs({"action": actions})["action"]
@@ -153,87 +118,41 @@ class DiffusionPolicy(nn.Module):
         action = self._queues["action"].popleft()
         return action
 
-    def forward(self, batch: dict[str, Tensor], **_) -> dict[str, Tensor]:
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         """Run the batch through the model and compute the loss for training or validation."""
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
         loss = self.diffusion.compute_loss(batch)
         return {"loss": loss}
 
-    def update(self, batch: dict[str, Tensor], **_) -> dict:
-        """Run the model in train mode, compute the loss, and do an optimization step."""
-        start_time = time.time()
 
-        self.diffusion.train()
-
-        batch = self.normalize_inputs(batch)
-        batch = self.normalize_targets(batch)
-
-        loss = self.forward(batch)["loss"]
-        loss.backward()
-
-        # TODO(rcadene): self.unnormalize_outputs(out_dict)
-
-        grad_norm = torch.nn.utils.clip_grad_norm_(
-            self.diffusion.parameters(),
-            self.cfg.grad_clip_norm,
-            error_if_nonfinite=False,
-        )
-
-        self.optimizer.step()
-        self.optimizer.zero_grad()
-        self.lr_scheduler.step()
-
-        if self.ema is not None:
-            self.ema.step(self.diffusion)
-
-        info = {
-            "loss": loss.item(),
-            "grad_norm": float(grad_norm),
-            "lr": self.lr_scheduler.get_last_lr()[0],
-            "update_s": time.time() - start_time,
-        }
-
-        return info
-
-    def save(self, fp):
-        torch.save(self.state_dict(), fp)
-
-    def load(self, fp):
-        d = torch.load(fp)
-        missing_keys, unexpected_keys = self.load_state_dict(d, strict=False)
-        if len(missing_keys) > 0:
-            assert all(k.startswith("ema_diffusion.") for k in missing_keys)
-            logging.warning(
-                "DiffusionPolicy.load expected ema parameters in loaded state dict but none were found."
-            )
-        assert len(unexpected_keys) == 0
-
-
-class _DiffusionUnetImagePolicy(nn.Module):
-    def __init__(self, cfg: DiffusionConfig):
+class DiffusionModel(nn.Module):
+    def __init__(self, config: DiffusionConfig):
         super().__init__()
-        self.cfg = cfg
+        self.config = config
 
-        self.rgb_encoder = _RgbEncoder(cfg)
-        self.unet = _ConditionalUnet1D(
-            cfg,
-            global_cond_dim=(cfg.output_shapes["action"][0] + self.rgb_encoder.feature_dim) * cfg.n_obs_steps,
+        self.rgb_encoder = DiffusionRgbEncoder(config)
+        self.unet = DiffusionConditionalUnet1d(
+            config,
+            global_cond_dim=(config.output_shapes["action"][0] + self.rgb_encoder.feature_dim)
+            * config.n_obs_steps,
         )
 
         self.noise_scheduler = DDPMScheduler(
-            num_train_timesteps=cfg.num_train_timesteps,
-            beta_start=cfg.beta_start,
-            beta_end=cfg.beta_end,
-            beta_schedule=cfg.beta_schedule,
+            num_train_timesteps=config.num_train_timesteps,
+            beta_start=config.beta_start,
+            beta_end=config.beta_end,
+            beta_schedule=config.beta_schedule,
             variance_type="fixed_small",
-            clip_sample=cfg.clip_sample,
-            clip_sample_range=cfg.clip_sample_range,
-            prediction_type=cfg.prediction_type,
+            clip_sample=config.clip_sample,
+            clip_sample_range=config.clip_sample_range,
+            prediction_type=config.prediction_type,
         )
 
-        if cfg.num_inference_steps is None:
+        if config.num_inference_steps is None:
             self.num_inference_steps = self.noise_scheduler.config.num_train_timesteps
         else:
-            self.num_inference_steps = cfg.num_inference_steps
+            self.num_inference_steps = config.num_inference_steps
 
     # ========= inference  ============
     def conditional_sample(
@@ -244,7 +163,7 @@ class _DiffusionUnetImagePolicy(nn.Module):
 
         # Sample prior.
         sample = torch.randn(
-            size=(batch_size, self.cfg.horizon, self.cfg.output_shapes["action"][0]),
+            size=(batch_size, self.config.horizon, self.config.output_shapes["action"][0]),
             dtype=dtype,
             device=device,
             generator=generator,
@@ -274,7 +193,7 @@ class _DiffusionUnetImagePolicy(nn.Module):
         """
         assert set(batch).issuperset({"observation.state", "observation.image"})
         batch_size, n_obs_steps = batch["observation.state"].shape[:2]
-        assert n_obs_steps == self.cfg.n_obs_steps
+        assert n_obs_steps == self.config.n_obs_steps
 
         # Extract image feature (first combine batch and sequence dims).
         img_features = self.rgb_encoder(einops.rearrange(batch["observation.image"], "b n ... -> (b n) ..."))
@@ -287,10 +206,10 @@ class _DiffusionUnetImagePolicy(nn.Module):
         sample = self.conditional_sample(batch_size, global_cond=global_cond)
 
         # `horizon` steps worth of actions (from the first observation).
-        actions = sample[..., : self.cfg.output_shapes["action"][0]]
+        actions = sample[..., : self.config.output_shapes["action"][0]]
         # Extract `n_action_steps` steps worth of actions (from the current observation).
         start = n_obs_steps - 1
-        end = start + self.cfg.n_action_steps
+        end = start + self.config.n_action_steps
         actions = actions[:, start:end]
 
         return actions
@@ -309,8 +228,8 @@ class _DiffusionUnetImagePolicy(nn.Module):
         assert set(batch).issuperset({"observation.state", "observation.image", "action", "action_is_pad"})
         batch_size, n_obs_steps = batch["observation.state"].shape[:2]
         horizon = batch["action"].shape[1]
-        assert horizon == self.cfg.horizon
-        assert n_obs_steps == self.cfg.n_obs_steps
+        assert horizon == self.config.horizon
+        assert n_obs_steps == self.config.n_obs_steps
 
         # Extract image feature (first combine batch and sequence dims).
         img_features = self.rgb_encoder(einops.rearrange(batch["observation.image"], "b n ... -> (b n) ..."))
@@ -339,52 +258,52 @@ class _DiffusionUnetImagePolicy(nn.Module):
 
         # Compute the loss.
         # The target is either the original trajectory, or the noise.
-        if self.cfg.prediction_type == "epsilon":
+        if self.config.prediction_type == "epsilon":
             target = eps
-        elif self.cfg.prediction_type == "sample":
+        elif self.config.prediction_type == "sample":
             target = batch["action"]
         else:
-            raise ValueError(f"Unsupported prediction type {self.cfg.prediction_type}")
+            raise ValueError(f"Unsupported prediction type {self.config.prediction_type}")
 
         loss = F.mse_loss(pred, target, reduction="none")
 
         # Mask loss wherever the action is padded with copies (edges of the dataset trajectory).
-        if "action_is_pad" in batch:
+        if self.config.do_mask_loss_for_padding and "action_is_pad" in batch:
             in_episode_bound = ~batch["action_is_pad"]
             loss = loss * in_episode_bound.unsqueeze(-1)
 
         return loss.mean()
 
 
-class _RgbEncoder(nn.Module):
+class DiffusionRgbEncoder(nn.Module):
     """Encoder an RGB image into a 1D feature vector.
 
     Includes the ability to normalize and crop the image first.
     """
 
-    def __init__(self, cfg: DiffusionConfig):
+    def __init__(self, config: DiffusionConfig):
         super().__init__()
         # Set up optional preprocessing.
-        if cfg.crop_shape is not None:
+        if config.crop_shape is not None:
             self.do_crop = True
             # Always use center crop for eval
-            self.center_crop = torchvision.transforms.CenterCrop(cfg.crop_shape)
-            if cfg.crop_is_random:
-                self.maybe_random_crop = torchvision.transforms.RandomCrop(cfg.crop_shape)
+            self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
+            if config.crop_is_random:
+                self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
             else:
                 self.maybe_random_crop = self.center_crop
         else:
             self.do_crop = False
 
         # Set up backbone.
-        backbone_model = getattr(torchvision.models, cfg.vision_backbone)(
-            weights=cfg.pretrained_backbone_weights
+        backbone_model = getattr(torchvision.models, config.vision_backbone)(
+            weights=config.pretrained_backbone_weights
         )
         # Note: This assumes that the layer4 feature map is children()[-3]
         # TODO(alexander-soare): Use a safer alternative.
         self.backbone = nn.Sequential(*(list(backbone_model.children())[:-2]))
-        if cfg.use_group_norm:
-            if cfg.pretrained_backbone_weights:
+        if config.use_group_norm:
+            if config.pretrained_backbone_weights:
                 raise ValueError(
                     "You can't replace BatchNorm in a pretrained model without ruining the weights!"
                 )
@@ -398,11 +317,11 @@ class _RgbEncoder(nn.Module):
         # Use a dry run to get the feature map shape.
         with torch.inference_mode():
             feat_map_shape = tuple(
-                self.backbone(torch.zeros(size=(1, *cfg.input_shapes["observation.image"]))).shape[1:]
+                self.backbone(torch.zeros(size=(1, *config.input_shapes["observation.image"]))).shape[1:]
             )
-        self.pool = SpatialSoftmax(feat_map_shape, num_kp=cfg.spatial_softmax_num_keypoints)
-        self.feature_dim = cfg.spatial_softmax_num_keypoints * 2
-        self.out = nn.Linear(cfg.spatial_softmax_num_keypoints * 2, self.feature_dim)
+        self.pool = SpatialSoftmax(feat_map_shape, num_kp=config.spatial_softmax_num_keypoints)
+        self.feature_dim = config.spatial_softmax_num_keypoints * 2
+        self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
         self.relu = nn.ReLU()
 
     def forward(self, x: Tensor) -> Tensor:
@@ -459,7 +378,7 @@ def _replace_submodules(
     return root_module
 
 
-class _SinusoidalPosEmb(nn.Module):
+class DiffusionSinusoidalPosEmb(nn.Module):
     """1D sinusoidal positional embeddings as in Attention is All You Need."""
 
     def __init__(self, dim: int):
@@ -476,7 +395,7 @@ class _SinusoidalPosEmb(nn.Module):
         return emb
 
 
-class _Conv1dBlock(nn.Module):
+class DiffusionConv1dBlock(nn.Module):
     """Conv1d --> GroupNorm --> Mish"""
 
     def __init__(self, inp_channels, out_channels, kernel_size, n_groups=8):
@@ -492,40 +411,40 @@ class _Conv1dBlock(nn.Module):
         return self.block(x)
 
 
-class _ConditionalUnet1D(nn.Module):
+class DiffusionConditionalUnet1d(nn.Module):
     """A 1D convolutional UNet with FiLM modulation for conditioning.
 
     Note: this removes local conditioning as compared to the original diffusion policy code.
     """
 
-    def __init__(self, cfg: DiffusionConfig, global_cond_dim: int):
+    def __init__(self, config: DiffusionConfig, global_cond_dim: int):
         super().__init__()
 
-        self.cfg = cfg
+        self.config = config
 
         # Encoder for the diffusion timestep.
         self.diffusion_step_encoder = nn.Sequential(
-            _SinusoidalPosEmb(cfg.diffusion_step_embed_dim),
-            nn.Linear(cfg.diffusion_step_embed_dim, cfg.diffusion_step_embed_dim * 4),
+            DiffusionSinusoidalPosEmb(config.diffusion_step_embed_dim),
+            nn.Linear(config.diffusion_step_embed_dim, config.diffusion_step_embed_dim * 4),
             nn.Mish(),
-            nn.Linear(cfg.diffusion_step_embed_dim * 4, cfg.diffusion_step_embed_dim),
+            nn.Linear(config.diffusion_step_embed_dim * 4, config.diffusion_step_embed_dim),
         )
 
         # The FiLM conditioning dimension.
-        cond_dim = cfg.diffusion_step_embed_dim + global_cond_dim
+        cond_dim = config.diffusion_step_embed_dim + global_cond_dim
 
         # In channels / out channels for each downsampling block in the Unet's encoder. For the decoder, we
         # just reverse these.
-        in_out = [(cfg.output_shapes["action"][0], cfg.down_dims[0])] + list(
-            zip(cfg.down_dims[:-1], cfg.down_dims[1:], strict=True)
+        in_out = [(config.output_shapes["action"][0], config.down_dims[0])] + list(
+            zip(config.down_dims[:-1], config.down_dims[1:], strict=True)
         )
 
         # Unet encoder.
         common_res_block_kwargs = {
             "cond_dim": cond_dim,
-            "kernel_size": cfg.kernel_size,
-            "n_groups": cfg.n_groups,
-            "use_film_scale_modulation": cfg.use_film_scale_modulation,
+            "kernel_size": config.kernel_size,
+            "n_groups": config.n_groups,
+            "use_film_scale_modulation": config.use_film_scale_modulation,
         }
         self.down_modules = nn.ModuleList([])
         for ind, (dim_in, dim_out) in enumerate(in_out):
@@ -533,8 +452,8 @@ class _ConditionalUnet1D(nn.Module):
             self.down_modules.append(
                 nn.ModuleList(
                     [
-                        _ConditionalResidualBlock1D(dim_in, dim_out, **common_res_block_kwargs),
-                        _ConditionalResidualBlock1D(dim_out, dim_out, **common_res_block_kwargs),
+                        DiffusionConditionalResidualBlock1d(dim_in, dim_out, **common_res_block_kwargs),
+                        DiffusionConditionalResidualBlock1d(dim_out, dim_out, **common_res_block_kwargs),
                         # Downsample as long as it is not the last block.
                         nn.Conv1d(dim_out, dim_out, 3, 2, 1) if not is_last else nn.Identity(),
                     ]
@@ -544,8 +463,12 @@ class _ConditionalUnet1D(nn.Module):
         # Processing in the middle of the auto-encoder.
         self.mid_modules = nn.ModuleList(
             [
-                _ConditionalResidualBlock1D(cfg.down_dims[-1], cfg.down_dims[-1], **common_res_block_kwargs),
-                _ConditionalResidualBlock1D(cfg.down_dims[-1], cfg.down_dims[-1], **common_res_block_kwargs),
+                DiffusionConditionalResidualBlock1d(
+                    config.down_dims[-1], config.down_dims[-1], **common_res_block_kwargs
+                ),
+                DiffusionConditionalResidualBlock1d(
+                    config.down_dims[-1], config.down_dims[-1], **common_res_block_kwargs
+                ),
             ]
         )
 
@@ -557,8 +480,8 @@ class _ConditionalUnet1D(nn.Module):
                 nn.ModuleList(
                     [
                         # dim_in * 2, because it takes the encoder's skip connection as well
-                        _ConditionalResidualBlock1D(dim_in * 2, dim_out, **common_res_block_kwargs),
-                        _ConditionalResidualBlock1D(dim_out, dim_out, **common_res_block_kwargs),
+                        DiffusionConditionalResidualBlock1d(dim_in * 2, dim_out, **common_res_block_kwargs),
+                        DiffusionConditionalResidualBlock1d(dim_out, dim_out, **common_res_block_kwargs),
                         # Upsample as long as it is not the last block.
                         nn.ConvTranspose1d(dim_out, dim_out, 4, 2, 1) if not is_last else nn.Identity(),
                     ]
@@ -566,8 +489,8 @@ class _ConditionalUnet1D(nn.Module):
             )
 
         self.final_conv = nn.Sequential(
-            _Conv1dBlock(cfg.down_dims[0], cfg.down_dims[0], kernel_size=cfg.kernel_size),
-            nn.Conv1d(cfg.down_dims[0], cfg.output_shapes["action"][0], 1),
+            DiffusionConv1dBlock(config.down_dims[0], config.down_dims[0], kernel_size=config.kernel_size),
+            nn.Conv1d(config.down_dims[0], config.output_shapes["action"][0], 1),
         )
 
     def forward(self, x: Tensor, timestep: Tensor | int, global_cond=None) -> Tensor:
@@ -615,7 +538,7 @@ class _ConditionalUnet1D(nn.Module):
         return x
 
 
-class _ConditionalResidualBlock1D(nn.Module):
+class DiffusionConditionalResidualBlock1d(nn.Module):
     """ResNet style 1D convolutional block with FiLM modulation for conditioning."""
 
     def __init__(
@@ -634,13 +557,13 @@ class _ConditionalResidualBlock1D(nn.Module):
         self.use_film_scale_modulation = use_film_scale_modulation
         self.out_channels = out_channels
 
-        self.conv1 = _Conv1dBlock(in_channels, out_channels, kernel_size, n_groups=n_groups)
+        self.conv1 = DiffusionConv1dBlock(in_channels, out_channels, kernel_size, n_groups=n_groups)
 
         # FiLM modulation (https://arxiv.org/abs/1709.07871) outputs per-channel bias and (maybe) scale.
         cond_channels = out_channels * 2 if use_film_scale_modulation else out_channels
         self.cond_encoder = nn.Sequential(nn.Mish(), nn.Linear(cond_dim, cond_channels))
 
-        self.conv2 = _Conv1dBlock(out_channels, out_channels, kernel_size, n_groups=n_groups)
+        self.conv2 = DiffusionConv1dBlock(out_channels, out_channels, kernel_size, n_groups=n_groups)
 
         # A final convolution for dimension matching the residual (if needed).
         self.residual_conv = (
@@ -671,67 +594,3 @@ class _ConditionalResidualBlock1D(nn.Module):
         out = self.conv2(out)
         out = out + self.residual_conv(x)
         return out
-
-
-class _EMA:
-    """
-    Exponential Moving Average of models weights
-    """
-
-    def __init__(self, cfg: DiffusionConfig, model: nn.Module):
-        """
-        @crowsonkb's notes on EMA Warmup:
-            If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are good values for models you plan
-            to train for a million or more steps (reaches decay factor 0.999 at 31.6K steps, 0.9999 at 1M steps),
-            gamma=1, power=3/4 for models you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999
-            at 215.4k steps).
-        Args:
-            inv_gamma (float): Inverse multiplicative factor of EMA warmup. Default: 1.
-            power (float): Exponential factor of EMA warmup. Default: 2/3.
-            min_alpha (float): The minimum EMA decay rate. Default: 0.
-        """
-
-        self.averaged_model = model
-        self.averaged_model.eval()
-        self.averaged_model.requires_grad_(False)
-
-        self.update_after_step = cfg.ema_update_after_step
-        self.inv_gamma = cfg.ema_inv_gamma
-        self.power = cfg.ema_power
-        self.min_alpha = cfg.ema_min_alpha
-        self.max_alpha = cfg.ema_max_alpha
-
-        self.alpha = 0.0
-        self.optimization_step = 0
-
-    def get_decay(self, optimization_step):
-        """
-        Compute the decay factor for the exponential moving average.
-        """
-        step = max(0, optimization_step - self.update_after_step - 1)
-        value = 1 - (1 + step / self.inv_gamma) ** -self.power
-
-        if step <= 0:
-            return 0.0
-
-        return max(self.min_alpha, min(value, self.max_alpha))
-
-    @torch.no_grad()
-    def step(self, new_model):
-        self.alpha = self.get_decay(self.optimization_step)
-
-        for module, ema_module in zip(new_model.modules(), self.averaged_model.modules(), strict=True):
-            # Iterate over immediate parameters only.
-            for param, ema_param in zip(
-                module.parameters(recurse=False), ema_module.parameters(recurse=False), strict=True
-            ):
-                if isinstance(param, dict):
-                    raise RuntimeError("Dict parameter not supported")
-                if isinstance(module, _BatchNorm) or not param.requires_grad:
-                    # Copy BatchNorm parameters, and non-trainable parameters directly.
-                    ema_param.copy_(param.to(dtype=ema_param.dtype).data)
-                else:
-                    ema_param.mul_(self.alpha)
-                    ema_param.add_(param.data.to(dtype=ema_param.dtype), alpha=1 - self.alpha)
-
-        self.optimization_step += 1

lerobot/common/policies/factory.py

@@ -2,6 +2,7 @@ import inspect
 
 from omegaconf import DictConfig, OmegaConf
 
+from lerobot.common.policies.policy_protocol import Policy
 from lerobot.common.utils.utils import get_safe_torch_device
 
 
@@ -20,42 +21,53 @@ def _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg):
     return policy_cfg
 
 
-def make_policy(hydra_cfg: DictConfig, dataset_stats=None):
-    if hydra_cfg.policy.name == "tdmpc":
-        from lerobot.common.policies.tdmpc.policy import TDMPCPolicy
+def get_policy_and_config_classes(name: str) -> tuple[Policy, object]:
+    """Get the policy's class and config class given a name (matching the policy class' `name` attribute)."""
+    if name == "tdmpc":
+        from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
+        from lerobot.common.policies.tdmpc.modeling_tdmpc import TDMPCPolicy
 
-        policy = TDMPCPolicy(
-            hydra_cfg.policy,
-            n_obs_steps=hydra_cfg.n_obs_steps,
-            n_action_steps=hydra_cfg.n_action_steps,
-            device=hydra_cfg.device,
-        )
-    elif hydra_cfg.policy.name == "diffusion":
+        return TDMPCPolicy, TDMPCConfig
+    elif name == "diffusion":
         from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
         from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 
-        policy_cfg = _policy_cfg_from_hydra_cfg(DiffusionConfig, hydra_cfg)
-        policy = DiffusionPolicy(policy_cfg, hydra_cfg.offline_steps, dataset_stats)
-        policy.to(get_safe_torch_device(hydra_cfg.device))
-    elif hydra_cfg.policy.name == "act":
-        from lerobot.common.policies.act.configuration_act import ActionChunkingTransformerConfig
-        from lerobot.common.policies.act.modeling_act import ActionChunkingTransformerPolicy
+        return DiffusionPolicy, DiffusionConfig
+    elif name == "act":
+        from lerobot.common.policies.act.configuration_act import ACTConfig
+        from lerobot.common.policies.act.modeling_act import ACTPolicy
 
-        policy_cfg = _policy_cfg_from_hydra_cfg(ActionChunkingTransformerConfig, hydra_cfg)
-        policy = ActionChunkingTransformerPolicy(policy_cfg, dataset_stats)
-        policy.to(get_safe_torch_device(hydra_cfg.device))
+        return ACTPolicy, ACTConfig
     else:
-        raise ValueError(hydra_cfg.policy.name)
+        raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
-    if hydra_cfg.policy.pretrained_model_path:
-        # TODO(rcadene): hack for old pretrained models from fowm
-        if hydra_cfg.policy.name == "tdmpc" and "fowm" in hydra_cfg.policy.pretrained_model_path:
-            if "offline" in hydra_cfg.policy.pretrained_model_path:
-                policy.step[0] = 25000
-            elif "final" in hydra_cfg.policy.pretrained_model_path:
-                policy.step[0] = 100000
-            else:
-                raise NotImplementedError()
-        policy.load(hydra_cfg.policy.pretrained_model_path)
+
+def make_policy(
+    hydra_cfg: DictConfig, pretrained_policy_name_or_path: str | None = None, dataset_stats=None
+) -> Policy:
+    """Make an instance of a policy class.
+
+    Args:
+        hydra_cfg: A parsed Hydra configuration (see scripts). If `pretrained_policy_name_or_path` is
+            provided, only `hydra_cfg.policy.name` is used while everything else is ignored.
+        pretrained_policy_name_or_path: Either the repo ID of a model hosted on the Hub or a path to a
+            directory containing weights saved using `Policy.save_pretrained`. Note that providing this
+            argument overrides everything in `hydra_cfg.policy` apart from `hydra_cfg.policy.name`.
+        dataset_stats: Dataset statistics to use for (un)normalization of inputs/outputs in the policy. Must
+            be provided when initializing a new policy, and must not be provided when loading a pretrained
+            policy. Therefore, this argument is mutually exclusive with `pretrained_policy_name_or_path`.
+    """
+    if not (pretrained_policy_name_or_path is None) ^ (dataset_stats is None):
+        raise ValueError("Only one of `pretrained_policy_name_or_path` and `dataset_stats` may be provided.")
+
+    policy_cls, policy_cfg_class = get_policy_and_config_classes(hydra_cfg.policy.name)
+
+    if pretrained_policy_name_or_path is None:
+        policy_cfg = _policy_cfg_from_hydra_cfg(policy_cfg_class, hydra_cfg)
+        policy = policy_cls(policy_cfg, dataset_stats)
+    else:
+        policy = policy_cls.from_pretrained(pretrained_policy_name_or_path)
+
+    policy.to(get_safe_torch_device(hydra_cfg.device))
 
     return policy

lerobot/common/policies/normalize.py

@@ -57,17 +57,28 @@ def create_stats_buffers(
             )
 
         if stats is not None:
+            # 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.
             if mode == "mean_std":
-                buffer["mean"].data = stats[key]["mean"]
-                buffer["std"].data = stats[key]["std"]
+                buffer["mean"].data = stats[key]["mean"].clone()
+                buffer["std"].data = stats[key]["std"].clone()
             elif mode == "min_max":
-                buffer["min"].data = stats[key]["min"]
-                buffer["max"].data = stats[key]["max"]
+                buffer["min"].data = stats[key]["min"].clone()
+                buffer["max"].data = stats[key]["max"].clone()
 
         stats_buffers[key] = buffer
     return stats_buffers
 
 
+def _no_stats_error_str(name: str) -> str:
+    return (
+        f"`{name}` is infinity. You should either initialize with `stats` as an argument, or use a "
+        "pretrained model."
+    )
+
+
 class Normalize(nn.Module):
     """Normalizes data (e.g. "observation.image") for more stable and faster convergence during training."""
 
@@ -99,7 +110,6 @@ class Normalize(nn.Module):
         self.shapes = shapes
         self.modes = modes
         self.stats = stats
-        # `self.buffer_observation_state["mean"]` contains `torch.tensor(state_dim)`
         stats_buffers = create_stats_buffers(shapes, modes, stats)
         for key, buffer in stats_buffers.items():
             setattr(self, "buffer_" + key.replace(".", "_"), buffer)
@@ -113,26 +123,14 @@ class Normalize(nn.Module):
             if mode == "mean_std":
                 mean = buffer["mean"]
                 std = buffer["std"]
-                assert not torch.isinf(mean).any(), (
-                    "`mean` is infinity. You forgot to initialize with `stats` as argument, or called "
-                    "`policy.load_state_dict`."
-                )
-                assert not torch.isinf(std).any(), (
-                    "`std` is infinity. You forgot to initialize with `stats` as argument, or called "
-                    "`policy.load_state_dict`."
-                )
+                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)
             elif mode == "min_max":
                 min = buffer["min"]
                 max = buffer["max"]
-                assert not torch.isinf(min).any(), (
-                    "`min` is infinity. You forgot to initialize with `stats` as argument, or called "
-                    "`policy.load_state_dict`."
-                )
-                assert not torch.isinf(max).any(), (
-                    "`max` is infinity. You forgot to initialize with `stats` as argument, or called "
-                    "`policy.load_state_dict`."
-                )
+                assert not torch.isinf(min).any(), _no_stats_error_str("min")
+                assert not torch.isinf(max).any(), _no_stats_error_str("max")
                 # normalize to [0,1]
                 batch[key] = (batch[key] - min) / (max - min)
                 # normalize to [-1, 1]
@@ -190,26 +188,14 @@ class Unnormalize(nn.Module):
             if mode == "mean_std":
                 mean = buffer["mean"]
                 std = buffer["std"]
-                assert not torch.isinf(mean).any(), (
-                    "`mean` is infinity. You forgot to initialize with `stats` as argument, or called "
-                    "`policy.load_state_dict`."
-                )
-                assert not torch.isinf(std).any(), (
-                    "`std` is infinity. You forgot to initialize with `stats` as argument, or called "
-                    "`policy.load_state_dict`."
-                )
+                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
             elif mode == "min_max":
                 min = buffer["min"]
                 max = buffer["max"]
-                assert not torch.isinf(min).any(), (
-                    "`min` is infinity. You forgot to initialize with `stats` as argument, or called "
-                    "`policy.load_state_dict`."
-                )
-                assert not torch.isinf(max).any(), (
-                    "`max` is infinity. You forgot to initialize with `stats` as argument, or called "
-                    "`policy.load_state_dict`."
-                )
+                assert not torch.isinf(min).any(), _no_stats_error_str("min")
+                assert not torch.isinf(max).any(), _no_stats_error_str("max")
                 batch[key] = (batch[key] + 1) / 2
                 batch[key] = batch[key] * (max - min) + min
             else:

lerobot/common/policies/policy_protocol.py

@@ -14,10 +14,21 @@ from torch import Tensor
 
 @runtime_checkable
 class Policy(Protocol):
-    """The required interface for implementing a policy."""
+    """The required interface for implementing a policy.
+
+    We also expect all policies to subclass torch.nn.Module and PyTorchModelHubMixin.
+    """
 
     name: str
 
+    def __init__(self, cfg, dataset_stats: dict[str, dict[str, Tensor]] | None = None):
+        """
+        Args:
+            cfg: Policy configuration class instance or None, in which case the default instantiation of the
+                 configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization.
+        """
+
     def reset(self):
         """To be called whenever the environment is reset.
 
@@ -37,9 +48,12 @@ class Policy(Protocol):
         with caching.
         """
 
-    def update(self, batch):
-        """Does compute_loss then an optimization step.
 
-        TODO(alexander-soare): We will move the optimization step back into the training loop, so this will
-        disappear.
+@runtime_checkable
+class PolicyWithUpdate(Policy, Protocol):
+    def update(self):
+        """An update method that is to be called after a training optimization step.
+
+        Implements an additional updates the model parameters may need (for example, doing an EMA step for a
+        target model, or incrementing an internal buffer).
         """

lerobot/common/policies/tdmpc/configuration_tdmpc.py

@@ -0,0 +1,150 @@
+from dataclasses import dataclass, field
+
+
+@dataclass
+class TDMPCConfig:
+    """Configuration class for TDMPCPolicy.
+
+    Defaults are configured for training with xarm_lift_medium_replay providing proprioceptive and single
+    camera observations.
+
+    The parameters you will most likely need to change are the ones which depend on the environment / sensors.
+    Those are: `input_shapes`, `output_shapes`, and perhaps `max_random_shift`.
+
+    Args:
+        n_action_repeats: The number of times to repeat the action returned by the planning. (hint: Google
+            action repeats in Q-learning or ask your favorite chatbot)
+        horizon: Horizon for model predictive control.
+        input_normalization_modes: A dictionary with key representing the modality (e.g. "observation.state"),
+            and the value specifies the normalization mode to apply. The two available modes are "mean_std"
+            which subtracts the mean and divides by the standard deviation and "min_max" which rescale in a
+            [-1, 1] range. Note that here this defaults to None meaning inputs are not normalized. This is to
+            match the original implementation.
+        output_normalization_modes: Similar dictionary as `normalize_input_modes`, but to unnormalize to the
+            original scale. Note that this is also used for normalizing the training targets. NOTE: Clipping
+            to [-1, +1] is used during MPPI/CEM. Therefore, it is recommended that you stick with "min_max"
+            normalization mode here.
+        image_encoder_hidden_dim: Number of channels for the convolutional layers used for image encoding.
+        state_encoder_hidden_dim: Hidden dimension for MLP used for state vector encoding.
+        latent_dim: Observation's latent embedding dimension.
+        q_ensemble_size: Number of Q function estimators to use in an ensemble for uncertainty estimation.
+        mlp_dim: Hidden dimension of MLPs used for modelling the dynamics encoder, reward function, policy
+            (π), Q ensemble, and V.
+        discount: Discount factor (γ) to use for the reinforcement learning formalism.
+        use_mpc: Whether to use model predictive control. The alternative is to just sample the policy model
+            (π) for each step.
+        cem_iterations: Number of iterations for the MPPI/CEM loop in MPC.
+        max_std: Maximum standard deviation for actions sampled from the gaussian PDF in CEM.
+        min_std: Minimum standard deviation for noise applied to actions sampled from the policy model (π).
+            Doubles up as the minimum standard deviation for actions sampled from the gaussian PDF in CEM.
+        n_gaussian_samples: Number of samples to draw from the gaussian distribution every CEM iteration. Must
+            be non-zero.
+        n_pi_samples: Number of samples to draw from the policy / world model rollout every CEM iteration. Can
+            be zero.
+        uncertainty_regularizer_coeff: Coefficient for the uncertainty regularization used when estimating
+            trajectory values (this is the λ coeffiecient in eqn 4 of FOWM).
+        n_elites: The number of elite samples to use for updating the gaussian parameters every CEM iteration.
+        elite_weighting_temperature: The temperature to use for softmax weighting (by trajectory value) of the
+            elites, when updating the gaussian parameters for CEM.
+        gaussian_mean_momentum: Momentum (α) used for EMA updates of the mean parameter μ of the gaussian
+            paramters optimized in CEM. Updates are calculated as μ⁻ ← αμ⁻ + (1-α)μ.
+        max_random_shift_ratio: Maximum random shift (as a proportion of the image size) to apply to the
+            image(s) (in units of pixels) for training-time augmentation. If set to 0, no such augmentation
+            is applied. Note that the input images are assumed to be square for this augmentation.
+        reward_coeff: Loss weighting coefficient for the reward regression loss.
+        expectile_weight: Weighting (τ) used in expectile regression for the state value function (V).
+            v_pred < v_target is weighted by τ and v_pred >= v_target is weighted by (1-τ). τ is expected to
+            be in [0, 1]. Setting τ closer to 1 results in a more "optimistic" V. This is sensible to do
+            because v_target is obtained by evaluating the learned state-action value functions (Q) with
+            in-sample actions that may not be always optimal.
+        value_coeff: Loss weighting coefficient for both the state-action value (Q) TD loss, and the state
+            value (V) expectile regression loss.
+        consistency_coeff: Loss weighting coefficient for the consistency loss.
+        advantage_scaling: A factor by which the advantages are scaled prior to exponentiation for advantage
+            weighted regression of the policy (π) estimator parameters. Note that the exponentiated advantages
+            are clamped at 100.0.
+        pi_coeff: Loss weighting coefficient for the action regression loss.
+        temporal_decay_coeff: Exponential decay coefficient for decaying the loss coefficient for future time-
+            steps. Hint: each loss computation involves `horizon` steps worth of actions starting from the
+            current time step.
+        target_model_momentum: Momentum (α) used for EMA updates of the target models. Updates are calculated
+            as ϕ ← αϕ + (1-α)θ where ϕ are the parameters of the target model and θ are the parameters of the
+            model being trained.
+    """
+
+    # Input / output structure.
+    n_action_repeats: int = 2
+    horizon: int = 5
+
+    input_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "observation.image": [3, 84, 84],
+            "observation.state": [4],
+        }
+    )
+    output_shapes: dict[str, list[int]] = field(
+        default_factory=lambda: {
+            "action": [4],
+        }
+    )
+
+    # Normalization / Unnormalization
+    input_normalization_modes: dict[str, str] | None = None
+    output_normalization_modes: dict[str, str] = field(
+        default_factory=lambda: {"action": "min_max"},
+    )
+
+    # Architecture / modeling.
+    # Neural networks.
+    image_encoder_hidden_dim: int = 32
+    state_encoder_hidden_dim: int = 256
+    latent_dim: int = 50
+    q_ensemble_size: int = 5
+    mlp_dim: int = 512
+    # Reinforcement learning.
+    discount: float = 0.9
+
+    # Inference.
+    use_mpc: bool = True
+    cem_iterations: int = 6
+    max_std: float = 2.0
+    min_std: float = 0.05
+    n_gaussian_samples: int = 512
+    n_pi_samples: int = 51
+    uncertainty_regularizer_coeff: float = 1.0
+    n_elites: int = 50
+    elite_weighting_temperature: float = 0.5
+    gaussian_mean_momentum: float = 0.1
+
+    # Training and loss computation.
+    max_random_shift_ratio: float = 0.0476
+    # Loss coefficients.
+    reward_coeff: float = 0.5
+    expectile_weight: float = 0.9
+    value_coeff: float = 0.1
+    consistency_coeff: float = 20.0
+    advantage_scaling: float = 3.0
+    pi_coeff: float = 0.5
+    temporal_decay_coeff: float = 0.5
+    # Target model.
+    target_model_momentum: float = 0.995
+
+    def __post_init__(self):
+        """Input validation (not exhaustive)."""
+        if self.input_shapes["observation.image"][-2] != self.input_shapes["observation.image"][-1]:
+            # TODO(alexander-soare): This limitation is solely because of code in the random shift
+            # augmentation. It should be able to be removed.
+            raise ValueError(
+                "Only square images are handled now. Got image shape "
+                f"{self.input_shapes['observation.image']}."
+            )
+        if self.n_gaussian_samples <= 0:
+            raise ValueError(
+                f"The number of guassian samples for CEM should be non-zero. Got `{self.n_gaussian_samples=}`"
+            )
+        if self.output_normalization_modes != {"action": "min_max"}:
+            raise ValueError(
+                "TD-MPC assumes the action space dimensions to all be in [-1, 1]. Therefore it is strongly "
+                f"advised that you stick with the default. See {self.__class__.__name__} docstring for more "
+                "information."
+            )

lerobot/common/policies/tdmpc/helper.py

@@ -1,576 +0,0 @@
-import os
-import pickle
-import re
-
-import numpy as np
-import torch
-import torch.nn as nn
-import torch.nn.functional as F  # noqa: N812
-from torch import distributions as pyd
-from torch.distributions.utils import _standard_normal
-
-DEFAULT_ACT_FN = nn.Mish()
-
-
-def __REDUCE__(b):  # noqa: N802, N807
-    return "mean" if b else "none"
-
-
-def l1(pred, target, reduce=False):
-    """Computes the L1-loss between predictions and targets."""
-    return F.l1_loss(pred, target, reduction=__REDUCE__(reduce))
-
-
-def mse(pred, target, reduce=False):
-    """Computes the MSE loss between predictions and targets."""
-    return F.mse_loss(pred, target, reduction=__REDUCE__(reduce))
-
-
-def l2_expectile(diff, expectile=0.7, reduce=False):
-    weight = torch.where(diff > 0, expectile, (1 - expectile))
-    loss = weight * (diff**2)
-    reduction = __REDUCE__(reduce)
-    if reduction == "mean":
-        return torch.mean(loss)
-    elif reduction == "sum":
-        return torch.sum(loss)
-    return loss
-
-
-def _get_out_shape(in_shape, layers):
-    """Utility function. Returns the output shape of a network for a given input shape."""
-    x = torch.randn(*in_shape).unsqueeze(0)
-    return (nn.Sequential(*layers) if isinstance(layers, list) else layers)(x).squeeze(0).shape
-
-
-def gaussian_logprob(eps, log_std):
-    """Compute Gaussian log probability."""
-    residual = (-0.5 * eps.pow(2) - log_std).sum(-1, keepdim=True)
-    return residual - 0.5 * np.log(2 * np.pi) * eps.size(-1)
-
-
-def squash(mu, pi, log_pi):
-    """Apply squashing function."""
-    mu = torch.tanh(mu)
-    pi = torch.tanh(pi)
-    log_pi -= torch.log(F.relu(1 - pi.pow(2)) + 1e-6).sum(-1, keepdim=True)
-    return mu, pi, log_pi
-
-
-def orthogonal_init(m):
-    """Orthogonal layer initialization."""
-    if isinstance(m, nn.Linear):
-        nn.init.orthogonal_(m.weight.data)
-        if m.bias is not None:
-            nn.init.zeros_(m.bias)
-    elif isinstance(m, nn.Conv2d):
-        gain = nn.init.calculate_gain("relu")
-        nn.init.orthogonal_(m.weight.data, gain)
-        if m.bias is not None:
-            nn.init.zeros_(m.bias)
-
-
-def ema(m, m_target, tau):
-    """Update slow-moving average of online network (target network) at rate tau."""
-    with torch.no_grad():
-        # TODO(rcadene, aliberts): issue with strict=False
-        # for p, p_target in zip(m.parameters(), m_target.parameters(), strict=False):
-        #     p_target.data.lerp_(p.data, tau)
-        m_params_iter = iter(m.parameters())
-        m_target_params_iter = iter(m_target.parameters())
-
-        while True:
-            try:
-                p = next(m_params_iter)
-                p_target = next(m_target_params_iter)
-                p_target.data.lerp_(p.data, tau)
-            except StopIteration:
-                # If any iterator is exhausted, exit the loop
-                break
-
-
-def set_requires_grad(net, value):
-    """Enable/disable gradients for a given (sub)network."""
-    for param in net.parameters():
-        param.requires_grad_(value)
-
-
-class TruncatedNormal(pyd.Normal):
-    """Utility class implementing the truncated normal distribution."""
-
-    default_sample_shape = torch.Size()
-
-    def __init__(self, loc, scale, low=-1.0, high=1.0, eps=1e-6):
-        super().__init__(loc, scale, validate_args=False)
-        self.low = low
-        self.high = high
-        self.eps = eps
-
-    def _clamp(self, x):
-        clamped_x = torch.clamp(x, self.low + self.eps, self.high - self.eps)
-        x = x - x.detach() + clamped_x.detach()
-        return x
-
-    def sample(self, clip=None, sample_shape=default_sample_shape):
-        shape = self._extended_shape(sample_shape)
-        eps = _standard_normal(shape, dtype=self.loc.dtype, device=self.loc.device)
-        eps *= self.scale
-        if clip is not None:
-            eps = torch.clamp(eps, -clip, clip)
-        x = self.loc + eps
-        return self._clamp(x)
-
-
-class NormalizeImg(nn.Module):
-    """Normalizes pixel observations to [0,1) range."""
-
-    def __init__(self):
-        super().__init__()
-
-    def forward(self, x):
-        return x.div(255.0)
-
-
-class Flatten(nn.Module):
-    """Flattens its input to a (batched) vector."""
-
-    def __init__(self):
-        super().__init__()
-
-    def forward(self, x):
-        return x.view(x.size(0), -1)
-
-
-def enc(cfg):
-    obs_shape = {
-        "rgb": (3, cfg.img_size, cfg.img_size),
-        "state": (cfg.state_dim,),
-    }
-
-    """Returns a TOLD encoder."""
-    pixels_enc_layers, state_enc_layers = None, None
-    if cfg.modality in {"pixels", "all"}:
-        C = int(3 * cfg.frame_stack)  # noqa: N806
-        pixels_enc_layers = [
-            NormalizeImg(),
-            nn.Conv2d(C, cfg.num_channels, 7, stride=2),
-            nn.ReLU(),
-            nn.Conv2d(cfg.num_channels, cfg.num_channels, 5, stride=2),
-            nn.ReLU(),
-            nn.Conv2d(cfg.num_channels, cfg.num_channels, 3, stride=2),
-            nn.ReLU(),
-            nn.Conv2d(cfg.num_channels, cfg.num_channels, 3, stride=2),
-            nn.ReLU(),
-        ]
-        out_shape = _get_out_shape((C, cfg.img_size, cfg.img_size), pixels_enc_layers)
-        pixels_enc_layers.extend(
-            [
-                Flatten(),
-                nn.Linear(np.prod(out_shape), cfg.latent_dim),
-                nn.LayerNorm(cfg.latent_dim),
-                nn.Sigmoid(),
-            ]
-        )
-        if cfg.modality == "pixels":
-            return ConvExt(nn.Sequential(*pixels_enc_layers))
-    if cfg.modality in {"state", "all"}:
-        state_dim = obs_shape[0] if cfg.modality == "state" else obs_shape["state"][0]
-        state_enc_layers = [
-            nn.Linear(state_dim, cfg.enc_dim),
-            nn.ELU(),
-            nn.Linear(cfg.enc_dim, cfg.latent_dim),
-            nn.LayerNorm(cfg.latent_dim),
-            nn.Sigmoid(),
-        ]
-        if cfg.modality == "state":
-            return nn.Sequential(*state_enc_layers)
-    else:
-        raise NotImplementedError
-
-    encoders = {}
-    for k in obs_shape:
-        if k == "state":
-            encoders[k] = nn.Sequential(*state_enc_layers)
-        elif k.endswith("rgb"):
-            encoders[k] = ConvExt(nn.Sequential(*pixels_enc_layers))
-        else:
-            raise NotImplementedError
-    return Multiplexer(nn.ModuleDict(encoders))
-
-
-def mlp(in_dim, mlp_dim, out_dim, act_fn=DEFAULT_ACT_FN):
-    """Returns an MLP."""
-    if isinstance(mlp_dim, int):
-        mlp_dim = [mlp_dim, mlp_dim]
-    return nn.Sequential(
-        nn.Linear(in_dim, mlp_dim[0]),
-        nn.LayerNorm(mlp_dim[0]),
-        act_fn,
-        nn.Linear(mlp_dim[0], mlp_dim[1]),
-        nn.LayerNorm(mlp_dim[1]),
-        act_fn,
-        nn.Linear(mlp_dim[1], out_dim),
-    )
-
-
-def dynamics(in_dim, mlp_dim, out_dim, act_fn=DEFAULT_ACT_FN):
-    """Returns a dynamics network."""
-    return nn.Sequential(
-        mlp(in_dim, mlp_dim, out_dim, act_fn),
-        nn.LayerNorm(out_dim),
-        nn.Sigmoid(),
-    )
-
-
-def q(cfg):
-    action_dim = cfg.action_dim
-    """Returns a Q-function that uses Layer Normalization."""
-    return nn.Sequential(
-        nn.Linear(cfg.latent_dim + action_dim, cfg.mlp_dim),
-        nn.LayerNorm(cfg.mlp_dim),
-        nn.Tanh(),
-        nn.Linear(cfg.mlp_dim, cfg.mlp_dim),
-        nn.ELU(),
-        nn.Linear(cfg.mlp_dim, 1),
-    )
-
-
-def v(cfg):
-    """Returns a state value function that uses Layer Normalization."""
-    return nn.Sequential(
-        nn.Linear(cfg.latent_dim, cfg.mlp_dim),
-        nn.LayerNorm(cfg.mlp_dim),
-        nn.Tanh(),
-        nn.Linear(cfg.mlp_dim, cfg.mlp_dim),
-        nn.ELU(),
-        nn.Linear(cfg.mlp_dim, 1),
-    )
-
-
-def aug(cfg):
-    obs_shape = {
-        "rgb": (3, cfg.img_size, cfg.img_size),
-        "state": (4,),
-    }
-
-    """Multiplex augmentation"""
-    if cfg.modality == "state":
-        return nn.Identity()
-    elif cfg.modality == "pixels":
-        return RandomShiftsAug(cfg)
-    else:
-        augs = {}
-        for k in obs_shape:
-            if k == "state":
-                augs[k] = nn.Identity()
-            elif k.endswith("rgb"):
-                augs[k] = RandomShiftsAug(cfg)
-            else:
-                raise NotImplementedError
-        return Multiplexer(nn.ModuleDict(augs))
-
-
-class ConvExt(nn.Module):
-    """Auxiliary conv net accommodating high-dim input"""
-
-    def __init__(self, conv):
-        super().__init__()
-        self.conv = conv
-
-    def forward(self, x):
-        if x.ndim > 4:
-            batch_shape = x.shape[:-3]
-            out = self.conv(x.view(-1, *x.shape[-3:]))
-            out = out.view(*batch_shape, *out.shape[1:])
-        else:
-            out = self.conv(x)
-        return out
-
-
-class Multiplexer(nn.Module):
-    """Model multiplexer"""
-
-    def __init__(self, choices):
-        super().__init__()
-        self.choices = choices
-
-    def forward(self, x, key=None):
-        if isinstance(x, dict):
-            if key is not None:
-                return self.choices[key](x)
-            return {k: self.choices[k](_x) for k, _x in x.items()}
-        return self.choices(x)
-
-
-class RandomShiftsAug(nn.Module):
-    """
-    Random shift image augmentation.
-    Adapted from https://github.com/facebookresearch/drqv2
-    """
-
-    def __init__(self, cfg):
-        super().__init__()
-        assert cfg.modality in {"pixels", "all"}
-        self.pad = int(cfg.img_size / 21)
-
-    def forward(self, x):
-        n, c, h, w = x.size()
-        assert h == w
-        padding = tuple([self.pad] * 4)
-        x = F.pad(x, padding, "replicate")
-        eps = 1.0 / (h + 2 * self.pad)
-        arange = torch.linspace(
-            -1.0 + eps,
-            1.0 - eps,
-            h + 2 * self.pad,
-            device=x.device,
-            dtype=torch.float32,
-        )[:h]
-        arange = arange.unsqueeze(0).repeat(h, 1).unsqueeze(2)
-        base_grid = torch.cat([arange, arange.transpose(1, 0)], dim=2)
-        base_grid = base_grid.unsqueeze(0).repeat(n, 1, 1, 1)
-        shift = torch.randint(
-            0,
-            2 * self.pad + 1,
-            size=(n, 1, 1, 2),
-            device=x.device,
-            dtype=torch.float32,
-        )
-        shift *= 2.0 / (h + 2 * self.pad)
-        grid = base_grid + shift
-        return F.grid_sample(x, grid, padding_mode="zeros", align_corners=False)
-
-
-# TODO(aliberts): remove class
-# class Episode:
-#     """Storage object for a single episode."""
-
-#     def __init__(self, cfg, init_obs):
-#         action_dim = cfg.action_dim
-
-#         self.cfg = cfg
-#         self.device = torch.device(cfg.buffer_device)
-#         if cfg.modality in {"pixels", "state"}:
-#             dtype = torch.float32 if cfg.modality == "state" else torch.uint8
-#             self.obses = torch.empty(
-#                 (cfg.episode_length + 1, *init_obs.shape),
-#                 dtype=dtype,
-#                 device=self.device,
-#             )
-#             self.obses[0] = torch.tensor(init_obs, dtype=dtype, device=self.device)
-#         elif cfg.modality == "all":
-#             self.obses = {}
-#             for k, v in init_obs.items():
-#                 assert k in {"rgb", "state"}
-#                 dtype = torch.float32 if k == "state" else torch.uint8
-#                 self.obses[k] = torch.empty(
-#                     (cfg.episode_length + 1, *v.shape), dtype=dtype, device=self.device
-#                 )
-#                 self.obses[k][0] = torch.tensor(v, dtype=dtype, device=self.device)
-#         else:
-#             raise ValueError
-#         self.actions = torch.empty((cfg.episode_length, action_dim), dtype=torch.float32, device=self.device)
-#         self.rewards = torch.empty((cfg.episode_length,), dtype=torch.float32, device=self.device)
-#         self.dones = torch.empty((cfg.episode_length,), dtype=torch.bool, device=self.device)
-#         self.masks = torch.empty((cfg.episode_length,), dtype=torch.float32, device=self.device)
-#         self.cumulative_reward = 0
-#         self.done = False
-#         self.success = False
-#         self._idx = 0
-
-#     def __len__(self):
-#         return self._idx
-
-#     @classmethod
-#     def from_trajectory(cls, cfg, obses, actions, rewards, dones=None, masks=None):
-#         """Constructs an episode from a trajectory."""
-
-#         if cfg.modality in {"pixels", "state"}:
-#             episode = cls(cfg, obses[0])
-#             episode.obses[1:] = torch.tensor(obses[1:], dtype=episode.obses.dtype, device=episode.device)
-#         elif cfg.modality == "all":
-#             episode = cls(cfg, {k: v[0] for k, v in obses.items()})
-#             for k in obses:
-#                 episode.obses[k][1:] = torch.tensor(
-#                     obses[k][1:], dtype=episode.obses[k].dtype, device=episode.device
-#                 )
-#         else:
-#             raise NotImplementedError
-#         episode.actions = torch.tensor(actions, dtype=episode.actions.dtype, device=episode.device)
-#         episode.rewards = torch.tensor(rewards, dtype=episode.rewards.dtype, device=episode.device)
-#         episode.dones = (
-#             torch.tensor(dones, dtype=episode.dones.dtype, device=episode.device)
-#             if dones is not None
-#             else torch.zeros_like(episode.dones)
-#         )
-#         episode.masks = (
-#             torch.tensor(masks, dtype=episode.masks.dtype, device=episode.device)
-#             if masks is not None
-#             else torch.ones_like(episode.masks)
-#         )
-#         episode.cumulative_reward = torch.sum(episode.rewards)
-#         episode.done = True
-#         episode._idx = cfg.episode_length
-#         return episode
-
-#     @property
-#     def first(self):
-#         return len(self) == 0
-
-#     def __add__(self, transition):
-#         self.add(*transition)
-#         return self
-
-#     def add(self, obs, action, reward, done, mask=1.0, success=False):
-#         """Add a transition into the episode."""
-#         if isinstance(obs, dict):
-#             for k, v in obs.items():
-#                 self.obses[k][self._idx + 1] = torch.tensor(
-#                     v, dtype=self.obses[k].dtype, device=self.obses[k].device
-#                 )
-#         else:
-#             self.obses[self._idx + 1] = torch.tensor(obs, dtype=self.obses.dtype, device=self.obses.device)
-#         self.actions[self._idx] = action
-#         self.rewards[self._idx] = reward
-#         self.dones[self._idx] = done
-#         self.masks[self._idx] = mask
-#         self.cumulative_reward += reward
-#         self.done = done
-#         self.success = self.success or success
-#         self._idx += 1
-
-
-def get_dataset_dict(cfg, env, return_reward_normalizer=False):
-    """Construct a dataset for env"""
-    required_keys = [
-        "observations",
-        "next_observations",
-        "actions",
-        "rewards",
-        "dones",
-        "masks",
-    ]
-
-    if cfg.task.startswith("xarm"):
-        dataset_path = os.path.join(cfg.dataset_dir, "buffer.pkl")
-        print(f"Using offline dataset '{dataset_path}'")
-        with open(dataset_path, "rb") as f:
-            dataset_dict = pickle.load(f)
-        for k in required_keys:
-            if k not in dataset_dict and k[:-1] in dataset_dict:
-                dataset_dict[k] = dataset_dict.pop(k[:-1])
-    elif cfg.task.startswith("legged"):
-        dataset_path = os.path.join(cfg.dataset_dir, "buffer.pkl")
-        print(f"Using offline dataset '{dataset_path}'")
-        with open(dataset_path, "rb") as f:
-            dataset_dict = pickle.load(f)
-        dataset_dict["actions"] /= env.unwrapped.clip_actions
-        print(f"clip_actions={env.unwrapped.clip_actions}")
-    else:
-        import d4rl
-
-        dataset_dict = d4rl.qlearning_dataset(env)
-        dones = np.full_like(dataset_dict["rewards"], False, dtype=bool)
-
-        for i in range(len(dones) - 1):
-            if (
-                np.linalg.norm(dataset_dict["observations"][i + 1] - dataset_dict["next_observations"][i])
-                > 1e-6
-                or dataset_dict["terminals"][i] == 1.0
-            ):
-                dones[i] = True
-
-        dones[-1] = True
-
-        dataset_dict["masks"] = 1.0 - dataset_dict["terminals"]
-        del dataset_dict["terminals"]
-
-        for k, v in dataset_dict.items():
-            dataset_dict[k] = v.astype(np.float32)
-
-        dataset_dict["dones"] = dones
-
-    if cfg.is_data_clip:
-        lim = 1 - cfg.data_clip_eps
-        dataset_dict["actions"] = np.clip(dataset_dict["actions"], -lim, lim)
-    reward_normalizer = get_reward_normalizer(cfg, dataset_dict)
-    dataset_dict["rewards"] = reward_normalizer(dataset_dict["rewards"])
-
-    for key in required_keys:
-        assert key in dataset_dict, f"Missing `{key}` in dataset."
-
-    if return_reward_normalizer:
-        return dataset_dict, reward_normalizer
-    return dataset_dict
-
-
-def get_trajectory_boundaries_and_returns(dataset):
-    """
-    Split dataset into trajectories and compute returns
-    """
-    episode_starts = [0]
-    episode_ends = []
-
-    episode_return = 0
-    episode_returns = []
-
-    n_transitions = len(dataset["rewards"])
-
-    for i in range(n_transitions):
-        episode_return += dataset["rewards"][i]
-
-        if dataset["dones"][i]:
-            episode_returns.append(episode_return)
-            episode_ends.append(i + 1)
-            if i + 1 < n_transitions:
-                episode_starts.append(i + 1)
-            episode_return = 0.0
-
-    return episode_starts, episode_ends, episode_returns
-
-
-def normalize_returns(dataset, scaling=1000):
-    """
-    Normalize returns in the dataset
-    """
-    (_, _, episode_returns) = get_trajectory_boundaries_and_returns(dataset)
-    dataset["rewards"] /= np.max(episode_returns) - np.min(episode_returns)
-    dataset["rewards"] *= scaling
-    return dataset
-
-
-def get_reward_normalizer(cfg, dataset):
-    """
-    Get a reward normalizer for the dataset
-    """
-    if cfg.task.startswith("xarm"):
-        return lambda x: x
-    elif "maze" in cfg.task:
-        return lambda x: x - 1.0
-    elif cfg.task.split("-")[0] in ["hopper", "halfcheetah", "walker2d"]:
-        (_, _, episode_returns) = get_trajectory_boundaries_and_returns(dataset)
-        return lambda x: x / (np.max(episode_returns) - np.min(episode_returns)) * 1000.0
-    elif hasattr(cfg, "reward_scale"):
-        return lambda x: x * cfg.reward_scale
-    return lambda x: x
-
-
-def linear_schedule(schdl, step):
-    """
-    Outputs values following a linear decay schedule.
-    Adapted from https://github.com/facebookresearch/drqv2
-    """
-    try:
-        return float(schdl)
-    except ValueError:
-        match = re.match(r"linear\((.+),(.+),(.+),(.+)\)", schdl)
-        if match:
-            init, final, start, end = (float(g) for g in match.groups())
-            mix = np.clip((step - start) / (end - start), 0.0, 1.0)
-            return (1.0 - mix) * init + mix * final
-        match = re.match(r"linear\((.+),(.+),(.+)\)", schdl)
-        if match:
-            init, final, duration = (float(g) for g in match.groups())
-            mix = np.clip(step / duration, 0.0, 1.0)
-            return (1.0 - mix) * init + mix * final
-    raise NotImplementedError(schdl)

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -0,0 +1,798 @@
+"""Implementation of Finetuning Offline World Models in the Real World.
+
+The comments in this code may sometimes refer to these references:
+    TD-MPC paper: Temporal Difference Learning for Model Predictive Control (https://arxiv.org/abs/2203.04955)
+    FOWM paper: Finetuning Offline World Models in the Real World (https://arxiv.org/abs/2310.16029)
+
+TODO(alexander-soare): Make rollout work for batch sizes larger than 1.
+TODO(alexander-soare): Use batch-first throughout.
+"""
+
+# ruff: noqa: N806
+
+import logging
+from collections import deque
+from copy import deepcopy
+from functools import partial
+from typing import Callable
+
+import einops
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from huggingface_hub import PyTorchModelHubMixin
+from torch import Tensor
+
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
+from lerobot.common.policies.utils import get_device_from_parameters, populate_queues
+
+
+class TDMPCPolicy(nn.Module, PyTorchModelHubMixin):
+    """Implementation of TD-MPC learning + inference.
+
+    Please note several warnings for this policy.
+        - Evaluation of pretrained weights created with the original FOWM code
+            (https://github.com/fyhMer/fowm) works as expected. To be precise: we trained and evaluated a
+            model with the FOWM code for the xarm_lift_medium_replay dataset. We ported the weights across
+            to LeRobot, and were able to evaluate with the same success metric. BUT, we had to use inter-
+            process communication to use the xarm environment from FOWM. This is because our xarm
+            environment uses newer dependencies and does not match the environment in FOWM. See
+            https://github.com/huggingface/lerobot/pull/103 for implementation details.
+        - We have NOT checked that training on LeRobot reproduces SOTA results. This is a TODO.
+        - Our current xarm datasets were generated using the environment from FOWM. Therefore they do not
+            match our xarm environment.
+    """
+
+    name = "tdmpc"
+
+    def __init__(
+        self, config: TDMPCConfig | None = None, dataset_stats: dict[str, dict[str, Tensor]] | None = None
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+        super().__init__()
+        logging.warning(
+            """
+            Please note several warnings for this policy.
+
+            - Evaluation of pretrained weights created with the original FOWM code
+              (https://github.com/fyhMer/fowm) works as expected. To be precise: we trained and evaluated a
+              model with the FOWM code for the xarm_lift_medium_replay dataset. We ported the weights across
+              to LeRobot, and were able to evaluate with the same success metric. BUT, we had to use inter-
+              process communication to use the xarm environment from FOWM. This is because our xarm
+              environment uses newer dependencies and does not match the environment in FOWM. See
+              https://github.com/huggingface/lerobot/pull/103 for implementation details.
+            - We have NOT checked that training on LeRobot reproduces SOTA results. This is a TODO.
+            - Our current xarm datasets were generated using the environment from FOWM. Therefore they do not
+              match our xarm environment.
+            """
+        )
+
+        if config is None:
+            config = TDMPCConfig()
+        self.config = config
+        self.model = TDMPCTOLD(config)
+        self.model_target = deepcopy(self.model)
+        for param in self.model_target.parameters():
+            param.requires_grad = False
+
+        if config.input_normalization_modes is not None:
+            self.normalize_inputs = Normalize(
+                config.input_shapes, config.input_normalization_modes, dataset_stats
+            )
+        else:
+            self.normalize_inputs = nn.Identity()
+        self.normalize_targets = Normalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_shapes, config.output_normalization_modes, dataset_stats
+        )
+
+    def save(self, fp):
+        """Save state dict of TOLD model to filepath."""
+        torch.save(self.state_dict(), fp)
+
+    def load(self, fp):
+        """Load a saved state dict from filepath into current agent."""
+        self.load_state_dict(torch.load(fp))
+
+    def reset(self):
+        """
+        Clear observation and action queues. Clear previous means for warm starting of MPPI/CEM. Should be
+        called on `env.reset()`
+        """
+        self._queues = {
+            "observation.image": deque(maxlen=1),
+            "observation.state": deque(maxlen=1),
+            "action": deque(maxlen=self.config.n_action_repeats),
+        }
+        # Previous mean obtained from the cross-entropy method (CEM) used during MPC. It is used to warm start
+        # CEM for the next step.
+        self._prev_mean: torch.Tensor | None = None
+
+    @torch.no_grad()
+    def select_action(self, batch: dict[str, Tensor]):
+        """Select a single action given environment observations."""
+        assert "observation.image" in batch
+        assert "observation.state" in batch
+
+        batch = self.normalize_inputs(batch)
+
+        self._queues = populate_queues(self._queues, batch)
+
+        # When the action queue is depleted, populate it again by querying the policy.
+        if len(self._queues["action"]) == 0:
+            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
+
+            # Remove the time dimensions as it is not handled yet.
+            for key in batch:
+                assert batch[key].shape[1] == 1
+                batch[key] = batch[key][:, 0]
+
+            # NOTE: Order of observations matters here.
+            z = self.model.encode({k: batch[k] for k in ["observation.image", "observation.state"]})
+            if self.config.use_mpc:
+                batch_size = batch["observation.image"].shape[0]
+                # Batch processing is not handled in MPC mode, so process the batch in a loop.
+                action = []  # will be a batch of actions for one step
+                for i in range(batch_size):
+                    # Note: self.plan does not handle batches, hence the squeeze.
+                    action.append(self.plan(z[i]))
+                action = torch.stack(action)
+            else:
+                # Plan with the policy (π) alone.
+                action = self.model.pi(z)
+
+            self.unnormalize_outputs({"action": action})["action"]
+
+            for _ in range(self.config.n_action_repeats):
+                self._queues["action"].append(action)
+
+        action = self._queues["action"].popleft()
+        return torch.clamp(action, -1, 1)
+
+    @torch.no_grad()
+    def plan(self, z: Tensor) -> Tensor:
+        """Plan next action using TD-MPC inference.
+
+        Args:
+            z: (latent_dim,) tensor for the initial state.
+        Returns:
+            (action_dim,) tensor for the next action.
+
+        TODO(alexander-soare) Extend this to be able to work with batches.
+        """
+        device = get_device_from_parameters(self)
+
+        # Sample Nπ trajectories from the policy.
+        pi_actions = torch.empty(
+            self.config.horizon,
+            self.config.n_pi_samples,
+            self.config.output_shapes["action"][0],
+            device=device,
+        )
+        if self.config.n_pi_samples > 0:
+            _z = einops.repeat(z, "d -> n d", n=self.config.n_pi_samples)
+            for t in range(self.config.horizon):
+                # Note: Adding a small amount of noise here doesn't hurt during inference and may even be
+                # helpful for CEM.
+                pi_actions[t] = self.model.pi(_z, self.config.min_std)
+                _z = self.model.latent_dynamics(_z, pi_actions[t])
+
+        # In the CEM loop we will need this for a call to estimate_value with the gaussian sampled
+        # trajectories.
+        z = einops.repeat(z, "d -> n d", n=self.config.n_gaussian_samples + self.config.n_pi_samples)
+
+        # Model Predictive Path Integral (MPPI) with the cross-entropy method (CEM) as the optimization
+        # algorithm.
+        # The initial mean and standard deviation for the cross-entropy method (CEM).
+        mean = torch.zeros(self.config.horizon, self.config.output_shapes["action"][0], device=device)
+        # Maybe warm start CEM with the mean from the previous step.
+        if self._prev_mean is not None:
+            mean[:-1] = self._prev_mean[1:]
+        std = self.config.max_std * torch.ones_like(mean)
+
+        for _ in range(self.config.cem_iterations):
+            # Randomly sample action trajectories for the gaussian distribution.
+            std_normal_noise = torch.randn(
+                self.config.horizon,
+                self.config.n_gaussian_samples,
+                self.config.output_shapes["action"][0],
+                device=std.device,
+            )
+            gaussian_actions = torch.clamp(mean.unsqueeze(1) + std.unsqueeze(1) * std_normal_noise, -1, 1)
+
+            # Compute elite actions.
+            actions = torch.cat([gaussian_actions, pi_actions], dim=1)
+            value = self.estimate_value(z, actions).nan_to_num_(0)
+            elite_idxs = torch.topk(value, self.config.n_elites, dim=0).indices
+            elite_value, elite_actions = value[elite_idxs], actions[:, elite_idxs]
+
+            # Update guassian PDF parameters to be the (weighted) mean and standard deviation of the elites.
+            max_value = elite_value.max(0)[0]
+            # The weighting is a softmax over trajectory values. Note that this is not the same as the usage
+            # of Ω in eqn 4 of the TD-MPC paper. Instead it is the normalized version of it: s = Ω/ΣΩ. This
+            # makes the equations: μ = Σ(s⋅Γ), σ = Σ(s⋅(Γ-μ)²).
+            score = torch.exp(self.config.elite_weighting_temperature * (elite_value - max_value))
+            score /= score.sum()
+            _mean = torch.sum(einops.rearrange(score, "n -> n 1") * elite_actions, dim=1)
+            _std = torch.sqrt(
+                torch.sum(
+                    einops.rearrange(score, "n -> n 1")
+                    * (elite_actions - einops.rearrange(_mean, "h d -> h 1 d")) ** 2,
+                    dim=1,
+                )
+            )
+            # Update mean with an exponential moving average, and std with a direct replacement.
+            mean = (
+                self.config.gaussian_mean_momentum * mean + (1 - self.config.gaussian_mean_momentum) * _mean
+            )
+            std = _std.clamp_(self.config.min_std, self.config.max_std)
+
+        # Keep track of the mean for warm-starting subsequent steps.
+        self._prev_mean = mean
+
+        # Randomly select one of the elite actions from the last iteration of MPPI/CEM using the softmax
+        # scores from the last iteration.
+        actions = elite_actions[:, torch.multinomial(score, 1).item()]
+
+        # Select only the first action
+        action = actions[0]
+        return action
+
+    @torch.no_grad()
+    def estimate_value(self, z: Tensor, actions: Tensor):
+        """Estimates the value of a trajectory as per eqn 4 of the FOWM paper.
+
+        Args:
+            z: (batch, latent_dim) tensor of initial latent states.
+            actions: (horizon, batch, action_dim) tensor of action trajectories.
+        Returns:
+            (batch,) tensor of values.
+        """
+        # Initialize return and running discount factor.
+        G, running_discount = 0, 1
+        # Iterate over the actions in the trajectory to simulate the trajectory using the latent dynamics
+        # model. Keep track of return.
+        for t in range(actions.shape[0]):
+            # We will compute the reward in a moment. First compute the uncertainty regularizer from eqn 4
+            # of the FOWM paper.
+            if self.config.uncertainty_regularizer_coeff > 0:
+                regularization = -(
+                    self.config.uncertainty_regularizer_coeff * self.model.Qs(z, actions[t]).std(0)
+                )
+            else:
+                regularization = 0
+            # Estimate the next state (latent) and reward.
+            z, reward = self.model.latent_dynamics_and_reward(z, actions[t])
+            # Update the return and running discount.
+            G += running_discount * (reward + regularization)
+            running_discount *= self.config.discount
+        # Add the estimated value of the final state (using the minimum for a conservative estimate).
+        # Do so by predicting the next action, then taking a minimum over the ensemble of state-action value
+        # estimators.
+        # Note: This small amount of added noise seems to help a bit at inference time as observed by success
+        # metrics over 50 episodes of xarm_lift_medium_replay.
+        next_action = self.model.pi(z, self.config.min_std)  # (batch, action_dim)
+        terminal_values = self.model.Qs(z, next_action)  # (ensemble, batch)
+        # Randomly choose 2 of the Qs for terminal value estimation (as in App C. of the FOWM paper).
+        if self.config.q_ensemble_size > 2:
+            G += (
+                running_discount
+                * torch.min(terminal_values[torch.randint(0, self.config.q_ensemble_size, size=(2,))], dim=0)[
+                    0
+                ]
+            )
+        else:
+            G += running_discount * torch.min(terminal_values, dim=0)[0]
+        # Finally, also regularize the terminal value.
+        if self.config.uncertainty_regularizer_coeff > 0:
+            G -= running_discount * self.config.uncertainty_regularizer_coeff * terminal_values.std(0)
+        return G
+
+    def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+        """Run the batch through the model and compute the loss."""
+        device = get_device_from_parameters(self)
+
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
+
+        info = {}
+
+        # TODO(alexander-soare): Refactor TDMPC and make it comply with the policy interface documentation.
+        batch_size = batch["index"].shape[0]
+
+        # (b, t) -> (t, b)
+        for key in batch:
+            if batch[key].ndim > 1:
+                batch[key] = batch[key].transpose(1, 0)
+
+        action = batch["action"]  # (t, b)
+        reward = batch["next.reward"]  # (t,)
+        observations = {k: v for k, v in batch.items() if k.startswith("observation.")}
+
+        # Apply random image augmentations.
+        if self.config.max_random_shift_ratio > 0:
+            observations["observation.image"] = flatten_forward_unflatten(
+                partial(random_shifts_aug, max_random_shift_ratio=self.config.max_random_shift_ratio),
+                observations["observation.image"],
+            )
+
+        # Get the current observation for predicting trajectories, and all future observations for use in
+        # the latent consistency loss and TD loss.
+        current_observation, next_observations = {}, {}
+        for k in observations:
+            current_observation[k] = observations[k][0]
+            next_observations[k] = observations[k][1:]
+        horizon = next_observations["observation.image"].shape[0]
+
+        # Run latent rollout using the latent dynamics model and policy model.
+        # Note this has shape `horizon+1` because there are `horizon` actions and a current `z`. Each action
+        # gives us a next `z`.
+        z_preds = torch.empty(horizon + 1, batch_size, self.config.latent_dim, device=device)
+        z_preds[0] = self.model.encode(current_observation)
+        reward_preds = torch.empty_like(reward, device=device)
+        for t in range(horizon):
+            z_preds[t + 1], reward_preds[t] = self.model.latent_dynamics_and_reward(z_preds[t], action[t])
+
+        # Compute Q and V value predictions based on the latent rollout.
+        q_preds_ensemble = self.model.Qs(z_preds[:-1], action)  # (ensemble, horizon, batch)
+        v_preds = self.model.V(z_preds[:-1])
+        info.update({"Q": q_preds_ensemble.mean().item(), "V": v_preds.mean().item()})
+
+        # Compute various targets with stopgrad.
+        with torch.no_grad():
+            # Latent state consistency targets.
+            z_targets = self.model_target.encode(next_observations)
+            # State-action value targets (or TD targets) as in eqn 3 of the FOWM. Unlike TD-MPC which uses the
+            # learned state-action value function in conjunction with the learned policy: Q(z, π(z)), FOWM
+            # uses a learned state value function: V(z). This means the TD targets only depend on in-sample
+            # actions (not actions estimated by π).
+            # Note: Here we do not use self.model_target, but self.model. This is to follow the original code
+            # and the FOWM paper.
+            q_targets = reward + self.config.discount * self.model.V(self.model.encode(next_observations))
+            # From eqn 3 of FOWM. These appear as Q(z, a). Here we call them v_targets to emphasize that we
+            # are using them to compute loss for V.
+            v_targets = self.model_target.Qs(z_preds[:-1].detach(), action, return_min=True)
+
+        # Compute losses.
+        # Exponentially decay the loss weight with respect to the timestep. Steps that are more distant in the
+        # future have less impact on the loss. Note: unsqueeze will let us broadcast to (seq, batch).
+        temporal_loss_coeffs = torch.pow(
+            self.config.temporal_decay_coeff, torch.arange(horizon, device=device)
+        ).unsqueeze(-1)
+        # Compute consistency loss as MSE loss between latents predicted from the rollout and latents
+        # predicted from the (target model's) observation encoder.
+        consistency_loss = (
+            (
+                temporal_loss_coeffs
+                * F.mse_loss(z_preds[1:], z_targets, reduction="none").mean(dim=-1)
+                # `z_preds` depends on the current observation and the actions.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+                # `z_targets` depends on the next observation.
+                * ~batch["observation.state_is_pad"][1:]
+            )
+            .sum(0)
+            .mean()
+        )
+        # Compute the reward loss as MSE loss between rewards predicted from the rollout and the dataset
+        # rewards.
+        reward_loss = (
+            (
+                temporal_loss_coeffs
+                * F.mse_loss(reward_preds, reward, reduction="none")
+                * ~batch["next.reward_is_pad"]
+                # `reward_preds` depends on the current observation and the actions.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+            )
+            .sum(0)
+            .mean()
+        )
+        # Compute state-action value loss (TD loss) for all of the Q functions in the ensemble.
+        q_value_loss = (
+            (
+                F.mse_loss(
+                    q_preds_ensemble,
+                    einops.repeat(q_targets, "t b -> e t b", e=q_preds_ensemble.shape[0]),
+                    reduction="none",
+                ).sum(0)  # sum over ensemble
+                # `q_preds_ensemble` depends on the first observation and the actions.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+                # q_targets depends on the reward and the next observations.
+                * ~batch["next.reward_is_pad"]
+                * ~batch["observation.state_is_pad"][1:]
+            )
+            .sum(0)
+            .mean()
+        )
+        # Compute state value loss as in eqn 3 of FOWM.
+        diff = v_targets - v_preds
+        # Expectile loss penalizes:
+        #   - `v_preds <  v_targets` with weighting `expectile_weight`
+        #   - `v_preds >= v_targets` with weighting `1 - expectile_weight`
+        raw_v_value_loss = torch.where(
+            diff > 0, self.config.expectile_weight, (1 - self.config.expectile_weight)
+        ) * (diff**2)
+        v_value_loss = (
+            (
+                temporal_loss_coeffs
+                * raw_v_value_loss
+                # `v_targets` depends on the first observation and the actions, as does `v_preds`.
+                * ~batch["observation.state_is_pad"][0]
+                * ~batch["action_is_pad"]
+            )
+            .sum(0)
+            .mean()
+        )
+
+        # Calculate the advantage weighted regression loss for π as detailed in FOWM 3.1.
+        # We won't need these gradients again so detach.
+        z_preds = z_preds.detach()
+        # Use stopgrad for the advantage calculation.
+        with torch.no_grad():
+            advantage = self.model_target.Qs(z_preds[:-1], action, return_min=True) - self.model.V(
+                z_preds[:-1]
+            )
+            info["advantage"] = advantage[0]
+            # (t, b)
+            exp_advantage = torch.clamp(torch.exp(advantage * self.config.advantage_scaling), max=100.0)
+        action_preds = self.model.pi(z_preds[:-1])  # (t, b, a)
+        # Calculate the MSE between the actions and the action predictions.
+        # Note: FOWM's original code calculates the log probability (wrt to a unit standard deviation
+        # gaussian) and sums over the action dimension. Computing the log probability amounts to multiplying
+        # the MSE by 0.5 and adding a constant offset (the log(2*pi) term) . Here we drop the constant offset
+        # as it doesn't change the optimization step, and we drop the 0.5 as we instead make a configuration
+        # parameter for it (see below where we compute the total loss).
+        mse = F.mse_loss(action_preds, action, reduction="none").sum(-1)  # (t, b)
+        # NOTE: The original implementation does not take the sum over the temporal dimension like with the
+        # other losses.
+        # TODO(alexander-soare): Take the sum over the temporal dimension and check that training still works
+        # as well as expected.
+        pi_loss = (
+            exp_advantage
+            * mse
+            * temporal_loss_coeffs
+            # `action_preds` depends on the first observation and the actions.
+            * ~batch["observation.state_is_pad"][0]
+            * ~batch["action_is_pad"]
+        ).mean()
+
+        loss = (
+            self.config.consistency_coeff * consistency_loss
+            + self.config.reward_coeff * reward_loss
+            + self.config.value_coeff * q_value_loss
+            + self.config.value_coeff * v_value_loss
+            + self.config.pi_coeff * pi_loss
+        )
+
+        info.update(
+            {
+                "consistency_loss": consistency_loss.item(),
+                "reward_loss": reward_loss.item(),
+                "Q_value_loss": q_value_loss.item(),
+                "V_value_loss": v_value_loss.item(),
+                "pi_loss": pi_loss.item(),
+                "loss": loss,
+                "sum_loss": loss.item() * self.config.horizon,
+            }
+        )
+
+        # Undo (b, t) -> (t, b).
+        for key in batch:
+            if batch[key].ndim > 1:
+                batch[key] = batch[key].transpose(1, 0)
+
+        return info
+
+    def update(self):
+        """Update the target model's parameters with an EMA step."""
+        # Note a minor variation with respect to the original FOWM code. Here they do this based on an EMA
+        # update frequency parameter which is set to 2 (every 2 steps an update is done). To simplify the code
+        # we update every step and adjust the decay parameter `alpha` accordingly (0.99 -> 0.995)
+        update_ema_parameters(self.model_target, self.model, self.config.target_model_momentum)
+
+
+class TDMPCTOLD(nn.Module):
+    """Task-Oriented Latent Dynamics (TOLD) model used in TD-MPC."""
+
+    def __init__(self, config: TDMPCConfig):
+        super().__init__()
+        self.config = config
+        self._encoder = TDMPCObservationEncoder(config)
+        self._dynamics = nn.Sequential(
+            nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.latent_dim),
+            nn.LayerNorm(config.latent_dim),
+            nn.Sigmoid(),
+        )
+        self._reward = nn.Sequential(
+            nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, 1),
+        )
+        self._pi = nn.Sequential(
+            nn.Linear(config.latent_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Mish(),
+            nn.Linear(config.mlp_dim, config.output_shapes["action"][0]),
+        )
+        self._Qs = nn.ModuleList(
+            [
+                nn.Sequential(
+                    nn.Linear(config.latent_dim + config.output_shapes["action"][0], config.mlp_dim),
+                    nn.LayerNorm(config.mlp_dim),
+                    nn.Tanh(),
+                    nn.Linear(config.mlp_dim, config.mlp_dim),
+                    nn.ELU(),
+                    nn.Linear(config.mlp_dim, 1),
+                )
+                for _ in range(config.q_ensemble_size)
+            ]
+        )
+        self._V = nn.Sequential(
+            nn.Linear(config.latent_dim, config.mlp_dim),
+            nn.LayerNorm(config.mlp_dim),
+            nn.Tanh(),
+            nn.Linear(config.mlp_dim, config.mlp_dim),
+            nn.ELU(),
+            nn.Linear(config.mlp_dim, 1),
+        )
+        self._init_weights()
+
+    def _init_weights(self):
+        """Initialize model weights.
+
+        Orthogonal initialization for all linear and convolutional layers' weights (apart from final layers
+        of reward network and Q networks which get zero initialization).
+        Zero initialization for all linear and convolutional layers' biases.
+        """
+
+        def _apply_fn(m):
+            if isinstance(m, nn.Linear):
+                nn.init.orthogonal_(m.weight.data)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Conv2d):
+                gain = nn.init.calculate_gain("relu")
+                nn.init.orthogonal_(m.weight.data, gain)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+        self.apply(_apply_fn)
+        for m in [self._reward, *self._Qs]:
+            assert isinstance(
+                m[-1], nn.Linear
+            ), "Sanity check. The last linear layer needs 0 initialization on weights."
+            nn.init.zeros_(m[-1].weight)
+            nn.init.zeros_(m[-1].bias)  # this has already been done, but keep this line here for good measure
+
+    def encode(self, obs: dict[str, Tensor]) -> Tensor:
+        """Encodes an observation into its latent representation."""
+        return self._encoder(obs)
+
+    def latent_dynamics_and_reward(self, z: Tensor, a: Tensor) -> tuple[Tensor, Tensor]:
+        """Predict the next state's latent representation and the reward given a current latent and action.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+        Returns:
+            A tuple containing:
+                - (*, latent_dim) tensor for the next state's latent representation.
+                - (*,) tensor for the estimated reward.
+        """
+        x = torch.cat([z, a], dim=-1)
+        return self._dynamics(x), self._reward(x).squeeze(-1)
+
+    def latent_dynamics(self, z: Tensor, a: Tensor) -> Tensor:
+        """Predict the next state's latent representation given a current latent and action.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+        Returns:
+            (*, latent_dim) tensor for the next state's latent representation.
+        """
+        x = torch.cat([z, a], dim=-1)
+        return self._dynamics(x)
+
+    def pi(self, z: Tensor, std: float = 0.0) -> Tensor:
+        """Samples an action from the learned policy.
+
+        The policy can also have added (truncated) Gaussian noise injected for encouraging exploration when
+        generating rollouts for online training.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            std: The standard deviation of the injected noise.
+        Returns:
+            (*, action_dim) tensor for the sampled action.
+        """
+        action = torch.tanh(self._pi(z))
+        if std > 0:
+            std = torch.ones_like(action) * std
+            action += torch.randn_like(action) * std
+        return action
+
+    def V(self, z: Tensor) -> Tensor:  # noqa: N802
+        """Predict state value (V).
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+        Returns:
+            (*,) tensor of estimated state values.
+        """
+        return self._V(z).squeeze(-1)
+
+    def Qs(self, z: Tensor, a: Tensor, return_min: bool = False) -> Tensor:  # noqa: N802
+        """Predict state-action value for all of the learned Q functions.
+
+        Args:
+            z: (*, latent_dim) tensor for the current state's latent representation.
+            a: (*, action_dim) tensor for the action to be applied.
+            return_min: Set to true for implementing the detail in App. C of the FOWM paper: randomly select
+                2 of the Qs and return the minimum
+        Returns:
+            (q_ensemble, *) tensor for the value predictions of each learned Q function in the ensemble OR
+            (*,) tensor if return_min=True.
+        """
+        x = torch.cat([z, a], dim=-1)
+        if not return_min:
+            return torch.stack([q(x).squeeze(-1) for q in self._Qs], dim=0)
+        else:
+            if len(self._Qs) > 2:  # noqa: SIM108
+                Qs = [self._Qs[i] for i in np.random.choice(len(self._Qs), size=2)]
+            else:
+                Qs = self._Qs
+            return torch.stack([q(x).squeeze(-1) for q in Qs], dim=0).min(dim=0)[0]
+
+
+class TDMPCObservationEncoder(nn.Module):
+    """Encode image and/or state vector observations."""
+
+    def __init__(self, config: TDMPCConfig):
+        """
+        Creates encoders for pixel and/or state modalities.
+        TODO(alexander-soare): The original work allows for multiple images by concatenating them along the
+            channel dimension. Re-implement this capability.
+        """
+        super().__init__()
+        self.config = config
+
+        if "observation.image" in config.input_shapes:
+            self.image_enc_layers = nn.Sequential(
+                nn.Conv2d(
+                    config.input_shapes["observation.image"][0], config.image_encoder_hidden_dim, 7, stride=2
+                ),
+                nn.ReLU(),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 5, stride=2),
+                nn.ReLU(),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
+                nn.ReLU(),
+                nn.Conv2d(config.image_encoder_hidden_dim, config.image_encoder_hidden_dim, 3, stride=2),
+                nn.ReLU(),
+            )
+            dummy_batch = torch.zeros(1, *config.input_shapes["observation.image"])
+            with torch.inference_mode():
+                out_shape = self.image_enc_layers(dummy_batch).shape[1:]
+            self.image_enc_layers.extend(
+                nn.Sequential(
+                    nn.Flatten(),
+                    nn.Linear(np.prod(out_shape), config.latent_dim),
+                    nn.LayerNorm(config.latent_dim),
+                    nn.Sigmoid(),
+                )
+            )
+        if "observation.state" in config.input_shapes:
+            self.state_enc_layers = nn.Sequential(
+                nn.Linear(config.input_shapes["observation.state"][0], config.state_encoder_hidden_dim),
+                nn.ELU(),
+                nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
+                nn.LayerNorm(config.latent_dim),
+                nn.Sigmoid(),
+            )
+
+    def forward(self, obs_dict: dict[str, Tensor]) -> Tensor:
+        """Encode the image and/or state vector.
+
+        Each modality is encoded into a feature vector of size (latent_dim,) and then a uniform mean is taken
+        over all features.
+        """
+        feat = []
+        if "observation.image" in self.config.input_shapes:
+            feat.append(flatten_forward_unflatten(self.image_enc_layers, obs_dict["observation.image"]))
+        if "observation.state" in self.config.input_shapes:
+            feat.append(self.state_enc_layers(obs_dict["observation.state"]))
+        return torch.stack(feat, dim=0).mean(0)
+
+
+def random_shifts_aug(x: Tensor, max_random_shift_ratio: float) -> Tensor:
+    """Randomly shifts images horizontally and vertically.
+
+    Adapted from https://github.com/facebookresearch/drqv2
+    """
+    b, _, h, w = x.size()
+    assert h == w, "non-square images not handled yet"
+    pad = int(round(max_random_shift_ratio * h))
+    x = F.pad(x, tuple([pad] * 4), "replicate")
+    eps = 1.0 / (h + 2 * pad)
+    arange = torch.linspace(
+        -1.0 + eps,
+        1.0 - eps,
+        h + 2 * pad,
+        device=x.device,
+        dtype=torch.float32,
+    )[:h]
+    arange = einops.repeat(arange, "w -> h w 1", h=h)
+    base_grid = torch.cat([arange, arange.transpose(1, 0)], dim=2)
+    base_grid = einops.repeat(base_grid, "h w c -> b h w c", b=b)
+    # A random shift in units of pixels and within the boundaries of the padding.
+    shift = torch.randint(
+        0,
+        2 * pad + 1,
+        size=(b, 1, 1, 2),
+        device=x.device,
+        dtype=torch.float32,
+    )
+    shift *= 2.0 / (h + 2 * pad)
+    grid = base_grid + shift
+    return F.grid_sample(x, grid, padding_mode="zeros", align_corners=False)
+
+
+def update_ema_parameters(ema_net: nn.Module, net: nn.Module, alpha: float):
+    """Update EMA parameters in place with ema_param <- alpha * ema_param + (1 - alpha) * param."""
+    for ema_module, module in zip(ema_net.modules(), net.modules(), strict=True):
+        for (n_p_ema, p_ema), (n_p, p) in zip(
+            ema_module.named_parameters(recurse=False), module.named_parameters(recurse=False), strict=True
+        ):
+            assert n_p_ema == n_p, "Parameter names don't match for EMA model update"
+            if isinstance(p, dict):
+                raise RuntimeError("Dict parameter not supported")
+            if isinstance(module, nn.modules.batchnorm._BatchNorm) or not p.requires_grad:
+                # Copy BatchNorm parameters, and non-trainable parameters directly.
+                p_ema.copy_(p.to(dtype=p_ema.dtype).data)
+            with torch.no_grad():
+                p_ema.mul_(alpha)
+                p_ema.add_(p.to(dtype=p_ema.dtype).data, alpha=1 - alpha)
+
+
+def flatten_forward_unflatten(fn: Callable[[Tensor], Tensor], image_tensor: Tensor) -> Tensor:
+    """Helper to temporarily flatten extra dims at the start of the image tensor.
+
+    Args:
+        fn: Callable that the image tensor will be passed to. It should accept (B, C, H, W) and return
+            (B, *), where * is any number of dimensions.
+        image_tensor: An image tensor of shape (**, C, H, W), where ** is any number of dimensions, generally
+            different from *.
+    Returns:
+        A return value from the callable reshaped to (**, *).
+    """
+    if image_tensor.ndim == 4:
+        return fn(image_tensor)
+    start_dims = image_tensor.shape[:-3]
+    inp = torch.flatten(image_tensor, end_dim=-4)
+    flat_out = fn(inp)
+    return torch.reshape(flat_out, (*start_dims, *flat_out.shape[1:]))

lerobot/common/policies/tdmpc/policy.py

@@ -1,495 +0,0 @@
-# ruff: noqa: N806
-
-import time
-from collections import deque
-from copy import deepcopy
-
-import einops
-import numpy as np
-import torch
-import torch.nn as nn
-
-import lerobot.common.policies.tdmpc.helper as h
-from lerobot.common.policies.utils import populate_queues
-from lerobot.common.utils.utils import get_safe_torch_device
-
-FIRST_FRAME = 0
-
-
-class TOLD(nn.Module):
-    """Task-Oriented Latent Dynamics (TOLD) model used in TD-MPC."""
-
-    def __init__(self, cfg):
-        super().__init__()
-        action_dim = cfg.action_dim
-
-        self.cfg = cfg
-        self._encoder = h.enc(cfg)
-        self._dynamics = h.dynamics(cfg.latent_dim + action_dim, cfg.mlp_dim, cfg.latent_dim)
-        self._reward = h.mlp(cfg.latent_dim + action_dim, cfg.mlp_dim, 1)
-        self._pi = h.mlp(cfg.latent_dim, cfg.mlp_dim, action_dim)
-        self._Qs = nn.ModuleList([h.q(cfg) for _ in range(cfg.num_q)])
-        self._V = h.v(cfg)
-        self.apply(h.orthogonal_init)
-        for m in [self._reward, *self._Qs]:
-            m[-1].weight.data.fill_(0)
-            m[-1].bias.data.fill_(0)
-
-    def track_q_grad(self, enable=True):
-        """Utility function. Enables/disables gradient tracking of Q-networks."""
-        for m in self._Qs:
-            h.set_requires_grad(m, enable)
-
-    def track_v_grad(self, enable=True):
-        """Utility function. Enables/disables gradient tracking of Q-networks."""
-        if hasattr(self, "_V"):
-            h.set_requires_grad(self._V, enable)
-
-    def encode(self, obs):
-        """Encodes an observation into its latent representation."""
-        out = self._encoder(obs)
-        if isinstance(obs, dict):
-            # fusion
-            out = torch.stack([v for k, v in out.items()]).mean(dim=0)
-        return out
-
-    def next(self, z, a):
-        """Predicts next latent state (d) and single-step reward (R)."""
-        x = torch.cat([z, a], dim=-1)
-        return self._dynamics(x), self._reward(x)
-
-    def next_dynamics(self, z, a):
-        """Predicts next latent state (d)."""
-        x = torch.cat([z, a], dim=-1)
-        return self._dynamics(x)
-
-    def pi(self, z, std=0):
-        """Samples an action from the learned policy (pi)."""
-        mu = torch.tanh(self._pi(z))
-        if std > 0:
-            std = torch.ones_like(mu) * std
-            return h.TruncatedNormal(mu, std).sample(clip=0.3)
-        return mu
-
-    def V(self, z):  # noqa: N802
-        """Predict state value (V)."""
-        return self._V(z)
-
-    def Q(self, z, a, return_type):  # noqa: N802
-        """Predict state-action value (Q)."""
-        assert return_type in {"min", "avg", "all"}
-        x = torch.cat([z, a], dim=-1)
-
-        if return_type == "all":
-            return torch.stack([q(x) for q in self._Qs], dim=0)
-
-        idxs = np.random.choice(self.cfg.num_q, 2, replace=False)
-        Q1, Q2 = self._Qs[idxs[0]](x), self._Qs[idxs[1]](x)
-        return torch.min(Q1, Q2) if return_type == "min" else (Q1 + Q2) / 2
-
-
-class TDMPCPolicy(nn.Module):
-    """Implementation of TD-MPC learning + inference."""
-
-    name = "tdmpc"
-
-    def __init__(self, cfg, n_obs_steps, n_action_steps, device):
-        super().__init__()
-        self.action_dim = cfg.action_dim
-
-        self.cfg = cfg
-        self.n_obs_steps = n_obs_steps
-        self.n_action_steps = n_action_steps
-        self.device = get_safe_torch_device(device)
-        self.std = h.linear_schedule(cfg.std_schedule, 0)
-        self.model = TOLD(cfg)
-        self.model.to(self.device)
-        self.model_target = deepcopy(self.model)
-        self.optim = torch.optim.Adam(self.model.parameters(), lr=self.cfg.lr)
-        self.pi_optim = torch.optim.Adam(self.model._pi.parameters(), lr=self.cfg.lr)
-        # self.bc_optim = torch.optim.Adam(self.model.parameters(), lr=self.cfg.lr)
-        self.model.eval()
-        self.model_target.eval()
-
-        self.register_buffer("step", torch.zeros(1))
-
-    def state_dict(self):
-        """Retrieve state dict of TOLD model, including slow-moving target network."""
-        return {
-            "model": self.model.state_dict(),
-            "model_target": self.model_target.state_dict(),
-        }
-
-    def save(self, fp):
-        """Save state dict of TOLD model to filepath."""
-        torch.save(self.state_dict(), fp)
-
-    def load(self, fp):
-        """Load a saved state dict from filepath into current agent."""
-        d = torch.load(fp)
-        self.model.load_state_dict(d["model"])
-        self.model_target.load_state_dict(d["model_target"])
-
-    def reset(self):
-        """
-        Clear observation and action queues. Should be called on `env.reset()`
-        """
-        self._queues = {
-            "observation.image": deque(maxlen=self.n_obs_steps),
-            "observation.state": deque(maxlen=self.n_obs_steps),
-            "action": deque(maxlen=self.n_action_steps),
-        }
-
-    @torch.no_grad()
-    def select_action(self, batch, step):
-        assert "observation.image" in batch
-        assert "observation.state" in batch
-        assert len(batch) == 2
-
-        self._queues = populate_queues(self._queues, batch)
-
-        t0 = step == 0
-
-        self.eval()
-
-        if len(self._queues["action"]) == 0:
-            batch = {key: torch.stack(list(self._queues[key]), dim=1) for key in batch}
-
-            if self.n_obs_steps == 1:
-                # hack to remove the time dimension
-                for key in batch:
-                    assert batch[key].shape[1] == 1
-                    batch[key] = batch[key][:, 0]
-
-            actions = []
-            batch_size = batch["observation.image"].shape[0]
-            for i in range(batch_size):
-                obs = {
-                    "rgb": batch["observation.image"][[i]],
-                    "state": batch["observation.state"][[i]],
-                }
-                # Note: unsqueeze needed because `act` still uses non-batch logic.
-                action = self.act(obs, t0=t0, step=self.step)
-                actions.append(action)
-            action = torch.stack(actions)
-
-            # tdmpc returns an action for 1 timestep only, so we copy it over `n_action_steps` time
-            if i in range(self.n_action_steps):
-                self._queues["action"].append(action)
-
-        action = self._queues["action"].popleft()
-        return action
-
-    @torch.no_grad()
-    def act(self, obs, t0=False, step=None):
-        """Take an action. Uses either MPC or the learned policy, depending on the self.cfg.mpc flag."""
-        obs = {k: o.detach() for k, o in obs.items()} if isinstance(obs, dict) else obs.detach()
-        z = self.model.encode(obs)
-        if self.cfg.mpc:
-            a = self.plan(z, t0=t0, step=step)
-        else:
-            a = self.model.pi(z, self.cfg.min_std * self.model.training).squeeze(0)
-        return a
-
-    @torch.no_grad()
-    def estimate_value(self, z, actions, horizon):
-        """Estimate value of a trajectory starting at latent state z and executing given actions."""
-        G, discount = 0, 1
-        for t in range(horizon):
-            if self.cfg.uncertainty_cost > 0:
-                G -= (
-                    discount
-                    * self.cfg.uncertainty_cost
-                    * self.model.Q(z, actions[t], return_type="all").std(dim=0)
-                )
-            z, reward = self.model.next(z, actions[t])
-            G += discount * reward
-            discount *= self.cfg.discount
-        pi = self.model.pi(z, self.cfg.min_std)
-        G += discount * self.model.Q(z, pi, return_type="min")
-        if self.cfg.uncertainty_cost > 0:
-            G -= discount * self.cfg.uncertainty_cost * self.model.Q(z, pi, return_type="all").std(dim=0)
-        return G
-
-    @torch.no_grad()
-    def plan(self, z, step=None, t0=True):
-        """
-        Plan next action using TD-MPC inference.
-        z: latent state.
-        step: current time step. determines e.g. planning horizon.
-        t0: whether current step is the first step of an episode.
-        """
-        # during eval: eval_mode: uniform sampling and action noise is disabled during evaluation.
-
-        assert step is not None
-        # Seed steps
-        if step < self.cfg.seed_steps and self.model.training:
-            return torch.empty(self.action_dim, dtype=torch.float32, device=self.device).uniform_(-1, 1)
-
-        # Sample policy trajectories
-        horizon = int(min(self.cfg.horizon, h.linear_schedule(self.cfg.horizon_schedule, step)))
-        num_pi_trajs = int(self.cfg.mixture_coef * self.cfg.num_samples)
-        if num_pi_trajs > 0:
-            pi_actions = torch.empty(horizon, num_pi_trajs, self.action_dim, device=self.device)
-            _z = z.repeat(num_pi_trajs, 1)
-            for t in range(horizon):
-                pi_actions[t] = self.model.pi(_z, self.cfg.min_std)
-                _z = self.model.next_dynamics(_z, pi_actions[t])
-
-        # Initialize state and parameters
-        z = z.repeat(self.cfg.num_samples + num_pi_trajs, 1)
-        mean = torch.zeros(horizon, self.action_dim, device=self.device)
-        std = self.cfg.max_std * torch.ones(horizon, self.action_dim, device=self.device)
-        if not t0 and hasattr(self, "_prev_mean"):
-            mean[:-1] = self._prev_mean[1:]
-
-        # Iterate CEM
-        for _ in range(self.cfg.iterations):
-            actions = torch.clamp(
-                mean.unsqueeze(1)
-                + std.unsqueeze(1)
-                * torch.randn(horizon, self.cfg.num_samples, self.action_dim, device=std.device),
-                -1,
-                1,
-            )
-            if num_pi_trajs > 0:
-                actions = torch.cat([actions, pi_actions], dim=1)
-
-            # Compute elite actions
-            value = self.estimate_value(z, actions, horizon).nan_to_num_(0)
-            elite_idxs = torch.topk(value.squeeze(1), self.cfg.num_elites, dim=0).indices
-            elite_value, elite_actions = value[elite_idxs], actions[:, elite_idxs]
-
-            # Update parameters
-            max_value = elite_value.max(0)[0]
-            score = torch.exp(self.cfg.temperature * (elite_value - max_value))
-            score /= score.sum(0)
-            _mean = torch.sum(score.unsqueeze(0) * elite_actions, dim=1) / (score.sum(0) + 1e-9)
-            _std = torch.sqrt(
-                torch.sum(
-                    score.unsqueeze(0) * (elite_actions - _mean.unsqueeze(1)) ** 2,
-                    dim=1,
-                )
-                / (score.sum(0) + 1e-9)
-            )
-            _std = _std.clamp_(self.std, self.cfg.max_std)
-            mean, std = self.cfg.momentum * mean + (1 - self.cfg.momentum) * _mean, _std
-
-        # Outputs
-        # TODO(rcadene): remove numpy with
-        # # Convert score tensor to probabilities using softmax
-        # probabilities = torch.softmax(score, dim=0)
-        # # Generate a random sample index based on the probabilities
-        # sample_index = torch.multinomial(probabilities, 1).item()
-        score = score.squeeze(1).cpu().numpy()
-        actions = elite_actions[:, np.random.choice(np.arange(score.shape[0]), p=score)]
-        self._prev_mean = mean
-        mean, std = actions[0], _std[0]
-        a = mean
-        if self.model.training:
-            a += std * torch.randn(self.action_dim, device=std.device)
-        return torch.clamp(a, -1, 1)
-
-    def update_pi(self, zs, acts=None):
-        """Update policy using a sequence of latent states."""
-        self.pi_optim.zero_grad(set_to_none=True)
-        self.model.track_q_grad(False)
-        self.model.track_v_grad(False)
-
-        info = {}
-        # Advantage Weighted Regression
-        assert acts is not None
-        vs = self.model.V(zs)
-        qs = self.model_target.Q(zs, acts, return_type="min")
-        adv = qs - vs
-        exp_a = torch.exp(adv * self.cfg.A_scaling)
-        exp_a = torch.clamp(exp_a, max=100.0)
-        log_probs = h.gaussian_logprob(self.model.pi(zs) - acts, 0)
-        rho = torch.pow(self.cfg.rho, torch.arange(len(qs), device=self.device))
-        pi_loss = -((exp_a * log_probs).mean(dim=(1, 2)) * rho).mean()
-        info["adv"] = adv[0]
-
-        pi_loss.backward()
-        torch.nn.utils.clip_grad_norm_(
-            self.model._pi.parameters(),
-            self.cfg.grad_clip_norm,
-            error_if_nonfinite=False,
-        )
-        self.pi_optim.step()
-        self.model.track_q_grad(True)
-        self.model.track_v_grad(True)
-
-        info["pi_loss"] = pi_loss.item()
-        return pi_loss.item(), info
-
-    @torch.no_grad()
-    def _td_target(self, next_z, reward, mask):
-        """Compute the TD-target from a reward and the observation at the following time step."""
-        next_v = self.model.V(next_z)
-        td_target = reward + self.cfg.discount * mask * next_v.squeeze(2)
-        return td_target
-
-    def forward(self, batch, step):
-        # TODO(alexander-soare): Refactor TDMPC and make it comply with the policy interface documentation.
-        raise NotImplementedError()
-
-    def update(self, batch, step):
-        """Main update function. Corresponds to one iteration of the model learning."""
-        start_time = time.time()
-
-        batch_size = batch["index"].shape[0]
-
-        # TODO(rcadene): convert tdmpc with (batch size, time/horizon, channels)
-        # instead of currently (time/horizon, batch size, channels) which is not the pytorch convention
-        # batch size b = 256, time/horizon t = 5
-        # b t ... -> t b ...
-        for key in batch:
-            if batch[key].ndim > 1:
-                batch[key] = batch[key].transpose(1, 0)
-
-        action = batch["action"]
-        reward = batch["next.reward"]
-        #  idxs = batch["index"]  # TODO(rcadene): use idxs to update sampling weights
-        done = torch.zeros_like(reward, dtype=torch.bool, device=reward.device)
-        mask = torch.ones_like(reward, dtype=torch.bool, device=reward.device)
-        weights = torch.ones(batch_size, dtype=torch.bool, device=reward.device)
-
-        obses = {
-            "rgb": batch["observation.image"],
-            "state": batch["observation.state"],
-        }
-
-        shapes = {}
-        for k in obses:
-            shapes[k] = obses[k].shape
-            obses[k] = einops.rearrange(obses[k], "t b ... -> (t b) ... ")
-
-        # Apply augmentations
-        aug_tf = h.aug(self.cfg)
-        obses = aug_tf(obses)
-
-        for k in obses:
-            t, b = shapes[k][:2]
-            obses[k] = einops.rearrange(obses[k], "(t b) ... -> t b ... ", b=b, t=t)
-
-        obs, next_obses = {}, {}
-        for k in obses:
-            obs[k] = obses[k][0]
-            next_obses[k] = obses[k][1:].clone()
-
-        horizon = next_obses["rgb"].shape[0]
-        loss_mask = torch.ones_like(mask, device=self.device)
-        for t in range(1, horizon):
-            loss_mask[t] = loss_mask[t - 1] * (~done[t - 1])
-
-        self.optim.zero_grad(set_to_none=True)
-        self.std = h.linear_schedule(self.cfg.std_schedule, step)
-        self.model.train()
-
-        data_s = time.time() - start_time
-
-        # Compute targets
-        with torch.no_grad():
-            next_z = self.model.encode(next_obses)
-            z_targets = self.model_target.encode(next_obses)
-            td_targets = self._td_target(next_z, reward, mask)
-
-        # Latent rollout
-        zs = torch.empty(horizon + 1, batch_size, self.cfg.latent_dim, device=self.device)
-        reward_preds = torch.empty_like(reward, device=self.device)
-        assert reward.shape[0] == horizon
-        z = self.model.encode(obs)
-        zs[0] = z
-        value_info = {"Q": 0.0, "V": 0.0}
-        for t in range(horizon):
-            z, reward_pred = self.model.next(z, action[t])
-            zs[t + 1] = z
-            reward_preds[t] = reward_pred.squeeze(1)
-
-        with torch.no_grad():
-            v_target = self.model_target.Q(zs[:-1].detach(), action, return_type="min")
-
-        # Predictions
-        qs = self.model.Q(zs[:-1], action, return_type="all")
-        qs = qs.squeeze(3)
-        value_info["Q"] = qs.mean().item()
-        v = self.model.V(zs[:-1])
-        value_info["V"] = v.mean().item()
-
-        # Losses
-        rho = torch.pow(self.cfg.rho, torch.arange(horizon, device=self.device)).view(-1, 1)
-        consistency_loss = (rho * torch.mean(h.mse(zs[1:], z_targets), dim=2) * loss_mask).sum(dim=0)
-        reward_loss = (rho * h.mse(reward_preds, reward) * loss_mask).sum(dim=0)
-        q_value_loss, priority_loss = 0, 0
-        for q in range(self.cfg.num_q):
-            q_value_loss += (rho * h.mse(qs[q], td_targets) * loss_mask).sum(dim=0)
-            priority_loss += (rho * h.l1(qs[q], td_targets) * loss_mask).sum(dim=0)
-
-        expectile = h.linear_schedule(self.cfg.expectile, step)
-        v_value_loss = (rho * h.l2_expectile(v_target - v, expectile=expectile).squeeze(2) * loss_mask).sum(
-            dim=0
-        )
-
-        total_loss = (
-            self.cfg.consistency_coef * consistency_loss
-            + self.cfg.reward_coef * reward_loss
-            + self.cfg.value_coef * q_value_loss
-            + self.cfg.value_coef * v_value_loss
-        )
-
-        weighted_loss = (total_loss * weights).mean()
-        weighted_loss.register_hook(lambda grad: grad * (1 / self.cfg.horizon))
-        has_nan = torch.isnan(weighted_loss).item()
-        if has_nan:
-            print(f"weighted_loss has nan: {total_loss=} {weights=}")
-        else:
-            weighted_loss.backward()
-
-        grad_norm = torch.nn.utils.clip_grad_norm_(
-            self.model.parameters(), self.cfg.grad_clip_norm, error_if_nonfinite=False
-        )
-        self.optim.step()
-
-        # TODO(rcadene): implement PrioritizedSampling by modifying sampler.weights with priorities computed by a criterion
-        # if self.cfg.per:
-        #     # Update priorities
-        #     priorities = priority_loss.clamp(max=1e4).detach()
-        #     has_nan = torch.isnan(priorities).any().item()
-        #     if has_nan:
-        #         print(f"priorities has nan: {priorities=}")
-        #     else:
-        #         replay_buffer.update_priority(
-        #             idxs[:num_slices],
-        #             priorities[:num_slices],
-        #         )
-        #         if demo_batch_size > 0:
-        #             demo_buffer.update_priority(demo_idxs, priorities[num_slices:])
-
-        # Update policy + target network
-        _, pi_update_info = self.update_pi(zs[:-1].detach(), acts=action)
-
-        if step % self.cfg.update_freq == 0:
-            h.ema(self.model._encoder, self.model_target._encoder, self.cfg.tau)
-            h.ema(self.model._Qs, self.model_target._Qs, self.cfg.tau)
-
-        self.model.eval()
-
-        info = {
-            "consistency_loss": float(consistency_loss.mean().item()),
-            "reward_loss": float(reward_loss.mean().item()),
-            "Q_value_loss": float(q_value_loss.mean().item()),
-            "V_value_loss": float(v_value_loss.mean().item()),
-            "sum_loss": float(total_loss.mean().item()),
-            "loss": float(weighted_loss.mean().item()),
-            "grad_norm": float(grad_norm),
-            "lr": self.cfg.lr,
-            "data_s": data_s,
-            "update_s": time.time() - start_time,
-        }
-        # info["demo_batch_size"] = demo_batch_size
-        info["expectile"] = expectile
-        info.update(value_info)
-        info.update(pi_update_info)
-
-        self.step[0] = step
-        return info

lerobot/configs/default.yaml

@@ -9,31 +9,30 @@ hydra:
   job:
     name: default
 
-seed: 1337
-# batch size for TorchRL SerialEnv. Each underlying env will get the seed = seed + env_index
-# NOTE: only diffusion policy supports rollout_batch_size > 1
-rollout_batch_size: 1
 device: cuda  # cpu
-prefetch: 4
-eval_freq: ???
-save_freq: ???
-eval_episodes: ???
-save_video: false
-save_model: false
-save_buffer: false
-train_steps: ???
-fps: ???
+# `seed` is used for training (eg: model initialization, dataset shuffling)
+# AND for the evaluation environments.
+seed: ???
+dataset_repo_id: lerobot/pusht
 
-offline_prioritized_sampler: true
+training:
+  offline_steps: ???
+  online_steps: ???
+  online_steps_between_rollouts: ???
+  online_sampling_ratio: 0.5
+  # `online_env_seed` is used for environments for online training data rollouts.
+  online_env_seed: ???
+  eval_freq: ???
+  save_freq: ???
+  log_freq: 250
+  save_model: true
 
-dataset:
-  repo_id: ???
-
-n_action_steps: ???
-n_obs_steps: ???
-env: ???
-
-policy: ???
+eval:
+  n_episodes: 1
+  # `batch_size` specifies the number of environments to use in a gym.vector.VectorEnv.
+  batch_size: 1
+  # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
+  use_async_envs: false
 
 wandb:
   enable: true

lerobot/configs/env/aloha.yaml

@@ -1,18 +1,7 @@
 # @package _global_
 
-eval_episodes: 50
-eval_freq: 7500
-save_freq: 75000
-log_freq: 250
-# TODO: same as xarm, need to adjust
-offline_steps: 25000
-online_steps: 25000
-
 fps: 50
 
-dataset:
-  repo_id: lerobot/aloha_sim_insertion_human
-
 env:
   name: aloha
   task: AlohaInsertion-v0

lerobot/configs/env/pusht.yaml

@@ -1,18 +1,7 @@
 # @package _global_
 
-eval_episodes: 50
-eval_freq: 7500
-save_freq: 75000
-log_freq: 250
-# TODO: same as xarm, need to adjust
-offline_steps: 25000
-online_steps: 25000
-
 fps: 10
 
-dataset:
-  repo_id: lerobot/pusht
-
 env:
   name: pusht
   task: PushT-v0

lerobot/configs/env/xarm.yaml

@@ -1,17 +1,7 @@
 # @package _global_
 
-eval_episodes: 20
-eval_freq: 1000
-save_freq: 10000
-log_freq: 50
-offline_steps: 25000
-online_steps: 25000
-
 fps: 15
 
-dataset:
-  repo_id: lerobot/xarm_lift_medium
-
 env:
   name: xarm
   task: XarmLift-v0

lerobot/configs/policy/act.yaml

@@ -1,30 +1,42 @@
 # @package _global_
 
-offline_steps: 80000
-online_steps: 0
+seed: 1000
+dataset_repo_id: lerobot/aloha_sim_insertion_human
 
-eval_episodes: 1
-eval_freq: 10000
-save_freq: 100000
-log_freq: 250
+training:
+  offline_steps: 80000
+  online_steps: 0
+  eval_freq: 10000
+  save_freq: 100000
+  log_freq: 250
+  save_model: true
 
-n_obs_steps: 1
-# when temporal_agg=False, n_action_steps=horizon
+  batch_size: 8
+  lr: 1e-5
+  lr_backbone: 1e-5
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
 
-override_dataset_stats:
-  observation.images.top:
-    # stats from imagenet, since we use a pretrained vision model
-    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
-    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+  override_dataset_stats:
+    observation.images.top:
+      # stats from imagenet, since we use a pretrained vision model
+      mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+      std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+  delta_timestamps:
+    action: "[i / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
 
 # See `configuration_act.py` for more details.
 policy:
   name: act
 
-  pretrained_model_path:
-
   # Input / output structure.
-  n_obs_steps: ${n_obs_steps}
+  n_obs_steps: 1
   chunk_size: 100 # chunk_size
   n_action_steps: 100
 
@@ -49,7 +61,7 @@ policy:
   replace_final_stride_with_dilation: false
   # Transformer layers.
   pre_norm: false
-  d_model: 512
+  dim_model: 512
   n_heads: 8
   dim_feedforward: 3200
   feedforward_activation: relu
@@ -66,15 +78,3 @@ policy:
   # Training and loss computation.
   dropout: 0.1
   kl_weight: 10.0
-
-  # ---
-  # TODO(alexander-soare): Remove these from the policy config.
-  batch_size: 8
-  lr: 1e-5
-  lr_backbone: 1e-5
-  weight_decay: 1e-4
-  grad_clip_norm: 10
-  utd: 1
-
-  delta_timestamps:
-    action: "[i / ${fps} for i in range(${policy.chunk_size})]"

lerobot/configs/policy/diffusion.yaml

@@ -1,22 +1,38 @@
 # @package _global_
 
+# Defaults for training for the PushT dataset as per https://github.com/real-stanford/diffusion_policy.
+# Note: We do not track EMA model weights as we discovered it does not improve the results. See
+#       https://github.com/huggingface/lerobot/pull/134 for more details.
+
 seed: 100000
-horizon: 16
-n_obs_steps: 2
-n_action_steps: 8
-dataset_obs_steps: ${n_obs_steps}
-past_action_visible: False
-keypoint_visible_rate: 1.0
+dataset_repo_id: lerobot/pusht
 
-eval_episodes: 50
-eval_freq: 5000
-save_freq: 5000
-log_freq: 250
+training:
+  offline_steps: 200000
+  online_steps: 0
+  eval_freq: 5000
+  save_freq: 5000
+  log_freq: 250
+  save_model: true
 
-offline_steps: 200000
-online_steps: 0
+  batch_size: 64
+  grad_clip_norm: 10
+  lr: 1.0e-4
+  lr_scheduler: cosine
+  lr_warmup_steps: 500
+  adam_betas: [0.95, 0.999]
+  adam_eps: 1.0e-8
+  adam_weight_decay: 1.0e-6
+  online_steps_between_rollouts: 1
 
-offline_prioritized_sampler: true
+  delta_timestamps:
+    observation.image: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    observation.state: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
+    action: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1 - ${policy.n_obs_steps} + ${policy.horizon})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
 
 override_dataset_stats:
   # TODO(rcadene, alexander-soare): should we remove image stats as well? do we use a pretrained vision model?
@@ -35,12 +51,10 @@ override_dataset_stats:
 policy:
   name: diffusion
 
-  pretrained_model_path:
-
   # Input / output structure.
-  n_obs_steps: ${n_obs_steps}
-  horizon: ${horizon}
-  n_action_steps: ${n_action_steps}
+  n_obs_steps: 2
+  horizon: 16
+  n_action_steps: 8
 
   input_shapes:
     # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
@@ -82,25 +96,5 @@ policy:
   # Inference
   num_inference_steps: 100
 
-  # ---
-  # TODO(alexander-soare): Remove these from the policy config.
-  batch_size: 64
-  grad_clip_norm: 10
-  lr: 1.0e-4
-  lr_scheduler: cosine
-  lr_warmup_steps: 500
-  adam_betas: [0.95, 0.999]
-  adam_eps: 1.0e-8
-  adam_weight_decay: 1.0e-6
-  utd: 1
-  use_ema: true
-  ema_update_after_step: 0
-  ema_min_alpha: 0.0
-  ema_max_alpha: 0.9999
-  ema_inv_gamma: 1.0
-  ema_power: 0.75
-
-  delta_timestamps:
-    observation.image: "[i / ${fps} for i in range(1 - ${n_obs_steps}, 1)]"
-    observation.state: "[i / ${fps} for i in range(1 - ${n_obs_steps}, 1)]"
-    action: "[i / ${fps} for i in range(1 - ${n_obs_steps}, 1 - ${n_obs_steps} + ${policy.horizon})]"
+  # Loss computation
+  do_mask_loss_for_padding: false

lerobot/configs/policy/tdmpc.yaml

@@ -1,85 +1,78 @@
 # @package _global_
 
-n_action_steps: 2
-n_obs_steps: 1
+seed: 1
+dataset_repo_id: lerobot/xarm_lift_medium_replay
+
+training:
+  offline_steps: 25000
+  online_steps: 25000
+  eval_freq: 5000
+  online_steps_between_rollouts: 1
+  online_sampling_ratio: 0.5
+  online_env_seed: 10000
+
+  batch_size: 256
+  grad_clip_norm: 10.0
+  lr: 3e-4
+
+  delta_timestamps:
+    observation.image: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
+    observation.state: "[i / ${fps} for i in range(${policy.horizon} + 1)]"
+    action: "[i / ${fps} for i in range(${policy.horizon})]"
+    next.reward: "[i / ${fps} for i in range(${policy.horizon})]"
 
 policy:
   name: tdmpc
 
-  reward_scale: 1.0
+  pretrained_model_path:
 
-  episode_length: ${env.episode_length}
-  discount: 0.9
-  modality: 'all'
-
-  # pixels
-  frame_stack: 1
-  num_channels: 32
-  img_size: ${env.image_size}
-  state_dim: ${env.action_dim}
-  action_dim: ${env.action_dim}
-
-  # planning
-  mpc: true
-  iterations: 6
-  num_samples: 512
-  num_elites: 50
-  mixture_coef: 0.1
-  min_std: 0.05
-  max_std: 2.0
-  temperature: 0.5
-  momentum: 0.1
-  uncertainty_cost: 1
-
-  # actor
-  log_std_min: -10
-  log_std_max: 2
-
-  # learning
-  batch_size: 256
-  max_buffer_size: 10000
+  # Input / output structure.
+  n_action_repeats: 2
   horizon: 5
-  reward_coef: 0.5
-  value_coef: 0.1
-  consistency_coef: 20
-  rho: 0.5
-  kappa: 0.1
-  lr: 3e-4
-  std_schedule: ${policy.min_std}
-  horizon_schedule: ${policy.horizon}
-  per: true
-  per_alpha: 0.6
-  per_beta: 0.4
-  grad_clip_norm: 10
-  seed_steps: 0
-  update_freq: 2
-  tau: 0.01
-  utd: 1
 
-  # offline rl
-  # dataset_dir: ???
-  data_first_percent: 1.0
-  is_data_clip: true
-  data_clip_eps: 1e-5
-  expectile: 0.9
-  A_scaling: 3.0
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.image: [3, 84, 84]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
 
-  # offline->online
-  offline_steps: ${offline_steps}
-  pretrained_model_path: ""
-  # pretrained_model_path: "/home/rcadene/code/fowm/logs/xarm_lift/all/default/2/models/offline.pt"
-  # pretrained_model_path: "/home/rcadene/code/fowm/logs/xarm_lift/all/default/2/models/final.pt"
-  balanced_sampling: true
-  demo_schedule: 0.5
+  # Normalization / Unnormalization
+  input_normalization_modes: null
+  output_normalization_modes:
+    action: min_max
 
-  # architecture
-  enc_dim: 256
-  num_q: 5
-  mlp_dim: 512
+  # Architecture / modeling.
+  # Neural networks.
+  image_encoder_hidden_dim: 32
+  state_encoder_hidden_dim: 256
   latent_dim: 50
+  q_ensemble_size: 5
+  mlp_dim: 512
+  # Reinforcement learning.
+  discount: 0.9
 
-  delta_timestamps:
-    observation.image: "[i / ${fps} for i in range(6)]"
-    observation.state: "[i / ${fps} for i in range(6)]"
-    action: "[i / ${fps} for i in range(5)]"
-    next.reward: "[i / ${fps} for i in range(5)]"
+  # Inference.
+  use_mpc: false
+  cem_iterations: 6
+  max_std: 2.0
+  min_std: 0.05
+  n_gaussian_samples: 512
+  n_pi_samples: 51
+  uncertainty_regularizer_coeff: 1.0
+  n_elites: 50
+  elite_weighting_temperature: 0.5
+  gaussian_mean_momentum: 0.1
+
+  # Training and loss computation.
+  max_random_shift_ratio: 0.0476
+  # Loss coefficients.
+  reward_coeff: 0.5
+  expectile_weight: 0.9
+  value_coeff: 0.1
+  consistency_coeff: 20.0
+  advantage_scaling: 3.0
+  pi_coeff: 0.5
+  temporal_decay_coeff: 0.5
+  # Target model.
+  target_model_momentum: 0.995

lerobot/scripts/eval.py

@@ -1,30 +1,29 @@
 """Evaluate a policy on an environment by running rollouts and computing metrics.
 
-The script may be run in one of two ways:
+Usage examples:
 
-1. By providing the path to a config file with the --config argument.
-2. By providing a HuggingFace Hub ID with the --hub-id argument. You may also provide a revision number with the
-    --revision argument.
+You want to evaluate a model from the hub (eg: https://huggingface.co/lerobot/diffusion_pusht)
+for 10 episodes.
 
-In either case, it is possible to override config arguments by adding a list of config.key=value arguments.
+```
+python lerobot/scripts/eval.py -p lerobot/diffusion_pusht eval.n_episodes=10
+```
 
-Examples:
-
-You have a specific config file to go with trained model weights, and want to run 10 episodes.
+OR, you want to evaluate a model checkpoint from the LeRobot training script for 10 episodes.
 
 ```
 python lerobot/scripts/eval.py \
---config PATH/TO/FOLDER/config.yaml \
-policy.pretrained_model_path=PATH/TO/FOLDER/weights.pth \
-eval_episodes=10
+    -p outputs/train/diffusion_pusht/checkpoints/005000 \
+    eval.n_episodes=10
 ```
 
-You have a HuggingFace Hub ID, you know which revision you want, and want to run 10 episodes (note that in this case,
-you don't need to specify which weights to use):
+Note that in both examples, the repo/folder should contain at least `config.json`, `config.yaml` and
+`model.safetensors`.
 
-```
-python lerobot/scripts/eval.py --hub-id HUB/ID --revision v1.0 eval_episodes=10
-```
+Note the formatting for providing the number of episodes. Generally, you may provide any number of arguments
+with `qualified.parameter.name=value`. In this case, the parameter eval.n_episodes appears as `n_episodes`
+nested under `eval` in the `config.yaml` found at
+https://huggingface.co/lerobot/diffusion_pusht/tree/main.
 """
 
 import argparse
@@ -35,283 +34,327 @@ import time
 from copy import deepcopy
 from datetime import datetime as dt
 from pathlib import Path
+from typing import Callable
 
 import einops
 import gymnasium as gym
 import numpy as np
 import torch
-from datasets import Dataset, Features, Image, Sequence, Value
+from datasets import Dataset, Features, Image, Sequence, Value, concatenate_datasets
 from huggingface_hub import snapshot_download
+from huggingface_hub.utils._errors import RepositoryNotFoundError
+from huggingface_hub.utils._validators import HFValidationError
 from PIL import Image as PILImage
+from torch import Tensor
 from tqdm import trange
 
+from lerobot.common.datasets.factory import make_dataset
 from lerobot.common.datasets.utils import hf_transform_to_torch
 from lerobot.common.envs.factory import make_env
-from lerobot.common.envs.utils import postprocess_action, preprocess_observation
+from lerobot.common.envs.utils import preprocess_observation
 from lerobot.common.logger import log_output_dir
 from lerobot.common.policies.factory import make_policy
+from lerobot.common.policies.policy_protocol import Policy
+from lerobot.common.policies.utils import get_device_from_parameters
 from lerobot.common.utils.io_utils import write_video
 from lerobot.common.utils.utils import get_safe_torch_device, init_hydra_config, init_logging, set_global_seed
 
 
+def rollout(
+    env: gym.vector.VectorEnv,
+    policy: Policy,
+    seeds: list[int] | None = None,
+    return_observations: bool = False,
+    render_callback: Callable[[gym.vector.VectorEnv], None] | None = None,
+    enable_progbar: bool = False,
+) -> dict:
+    """Run a batched policy rollout once through a batch of environments.
+
+    Note that all environments in the batch are run until the last environment is done. This means some
+    data will probably need to be discarded (for environments that aren't the first one to be done).
+
+    The return dictionary contains:
+        (optional) "observation": A a dictionary of (batch, sequence + 1, *) tensors mapped to observation
+            keys. NOTE the that this has an extra sequence element relative to the other keys in the
+            dictionary. This is because an extra observation is included for after the environment is
+            terminated or truncated.
+        "action": A (batch, sequence, action_dim) tensor of actions applied based on the observations (not
+            including the last observations).
+        "reward": A (batch, sequence) tensor of rewards received for applying the actions.
+        "success": A (batch, sequence) tensor of success conditions (the only time this can be True is upon
+            environment termination/truncation).
+        "don": A (batch, sequence) tensor of **cumulative** done conditions. For any given batch element,
+            the first True is followed by True's all the way till the end. This can be used for masking
+            extraneous elements from the sequences above.
+
+    Args:
+        env: The batch of environments.
+        policy: The policy.
+        seeds: The environments are seeded once at the start of the rollout. If provided, this argument
+            specifies the seeds for each of the environments.
+        return_observations: Whether to include all observations in the returned rollout data. Observations
+            are returned optionally because they typically take more memory to cache. Defaults to False.
+        render_callback: Optional rendering callback to be used after the environments are reset, and after
+            every step.
+        enable_progbar: Enable a progress bar over rollout steps.
+    Returns:
+        The dictionary described above.
+    """
+    device = get_device_from_parameters(policy)
+
+    # Reset the policy and environments.
+    policy.reset()
+
+    observation, info = env.reset(seed=seeds)
+    if render_callback is not None:
+        render_callback(env)
+
+    all_observations = []
+    all_actions = []
+    all_rewards = []
+    all_successes = []
+    all_dones = []
+
+    step = 0
+    # Keep track of which environments are done.
+    done = np.array([False] * env.num_envs)
+    max_steps = env.call("_max_episode_steps")[0]
+    progbar = trange(
+        max_steps,
+        desc=f"Running rollout with at most {max_steps} steps",
+        disable=not enable_progbar,
+        leave=False,
+    )
+    while not np.all(done):
+        # Numpy array to tensor and changing dictionary keys to LeRobot policy format.
+        observation = preprocess_observation(observation)
+        if return_observations:
+            all_observations.append(deepcopy(observation))
+
+        observation = {key: observation[key].to(device, non_blocking=True) for key in observation}
+
+        with torch.inference_mode():
+            action = policy.select_action(observation)
+
+        # Convert to CPU / numpy.
+        action = action.to("cpu").numpy()
+        assert action.ndim == 2, "Action dimensions should be (batch, action_dim)"
+
+        # Apply the next action.
+        observation, reward, terminated, truncated, info = env.step(action)
+        if render_callback is not None:
+            render_callback(env)
+
+        # VectorEnv stores is_success in `info["final_info"][env_index]["is_success"]`. "final_info" isn't
+        # available of none of the envs finished.
+        if "final_info" in info:
+            successes = [info["is_success"] if info is not None else False for info in info["final_info"]]
+        else:
+            successes = [False] * env.num_envs
+
+        # Keep track of which environments are done so far.
+        done = terminated | truncated | done
+
+        all_actions.append(torch.from_numpy(action))
+        all_rewards.append(torch.from_numpy(reward))
+        all_dones.append(torch.from_numpy(done))
+        all_successes.append(torch.tensor(successes))
+
+        step += 1
+        running_success_rate = (
+            einops.reduce(torch.stack(all_successes, dim=1), "b n -> b", "any").numpy().mean()
+        )
+        progbar.set_postfix({"running_success_rate": f"{running_success_rate.item() * 100:.1f}%"})
+        progbar.update()
+
+    # Track the final observation.
+    if return_observations:
+        observation = preprocess_observation(observation)
+        all_observations.append(deepcopy(observation))
+
+    # Stack the sequence along the first dimension so that we have (batch, sequence, *) tensors.
+    ret = {
+        "action": torch.stack(all_actions, dim=1),
+        "reward": torch.stack(all_rewards, dim=1),
+        "success": torch.stack(all_successes, dim=1),
+        "done": torch.stack(all_dones, dim=1),
+    }
+    if return_observations:
+        stacked_observations = {}
+        for key in all_observations[0]:
+            stacked_observations[key] = torch.stack([obs[key] for obs in all_observations], dim=1)
+        ret["observation"] = stacked_observations
+
+    return ret
+
+
 def eval_policy(
     env: gym.vector.VectorEnv,
     policy: torch.nn.Module,
+    n_episodes: int,
     max_episodes_rendered: int = 0,
-    video_dir: Path = None,
+    video_dir: Path | None = None,
     return_episode_data: bool = False,
-    seed=None,
-):
+    start_seed: int | None = None,
+    enable_progbar: bool = False,
+    enable_inner_progbar: bool = False,
+) -> dict:
     """
-    set `return_episode_data` to return a Hugging Face dataset object in an "episodes" key of the return dict.
+    Args:
+        env: The batch of environments.
+        policy: The policy.
+        n_episodes: The number of episodes to evaluate.
+        max_episodes_rendered: Maximum number of episodes to render into videos.
+        video_dir: Where to save rendered videos.
+        return_episode_data: Whether to return episode data for online training. Incorporates the data into
+            the "episodes" key of the returned dictionary.
+        start_seed: The first seed to use for the first individual rollout. For all subsequent rollouts the
+            seed is incremented by 1. If not provided, the environments are not manually seeded.
+        enable_progbar: Enable progress bar over batches.
+        enable_inner_progbar: Enable progress bar over steps in each batch.
+    Returns:
+        Dictionary with metrics and data regarding the rollouts.
     """
-    fps = env.unwrapped.metadata["render_fps"]
-
-    if policy is not None:
-        policy.eval()
-    device = "cpu" if policy is None else next(policy.parameters()).device
-
     start = time.time()
+    policy.eval()
+
+    # Determine how many batched rollouts we need to get n_episodes. Note that if n_episodes is not evenly
+    # divisible by env.num_envs we end up discarding some data in the last batch.
+    n_batches = n_episodes // env.num_envs + int((n_episodes % env.num_envs) != 0)
+
+    # Keep track of some metrics.
     sum_rewards = []
     max_rewards = []
     all_successes = []
-    seeds = []
+    all_seeds = []
     threads = []  # for video saving threads
-    episode_counter = 0  # for saving the correct number of videos
+    n_episodes_rendered = 0  # for saving the correct number of videos
 
-    num_episodes = len(env.envs)
-
-    # TODO(alexander-soare): if num_episodes is not evenly divisible by the batch size, this will do more work than
-    # needed as I'm currently taking a ceil.
-    ep_frames = []
-
-    def render_frame(env):
+    # Callback for visualization.
+    def render_frame(env: gym.vector.VectorEnv):
         # noqa: B023
-        eps_rendered = min(max_episodes_rendered, len(env.envs))
-        visu = np.stack([env.envs[i].render() for i in range(eps_rendered)])
-        ep_frames.append(visu)  # noqa: B023
+        if n_episodes_rendered >= max_episodes_rendered:
+            return
+        n_to_render_now = min(max_episodes_rendered - n_episodes_rendered, env.num_envs)
+        if isinstance(env, gym.vector.SyncVectorEnv):
+            ep_frames.append(np.stack([env.envs[i].render() for i in range(n_to_render_now)]))  # noqa: B023
+        elif isinstance(env, gym.vector.AsyncVectorEnv):
+            # Here we must render all frames and discard any we don't need.
+            ep_frames.append(np.stack(env.call("render")[:n_to_render_now]))
 
-    for _ in range(num_episodes):
-        seeds.append("TODO")
-
-    if hasattr(policy, "reset"):
-        policy.reset()
-    else:
-        logging.warning(
-            f"Policy {policy} doesnt have a `reset` method. It is required if the policy relies on an internal state during rollout."
-        )
-
-    # reset the environment
-    observation, info = env.reset(seed=seed)
     if max_episodes_rendered > 0:
-        render_frame(env)
+        video_paths: list[str] = []
 
-    observations = []
-    actions = []
-    # episode
-    # frame_id
-    # timestamp
-    rewards = []
-    successes = []
-    dones = []
+    if return_episode_data:
+        episode_data: dict | None = None
 
-    done = torch.tensor([False for _ in env.envs])
-    step = 0
-    max_steps = env.envs[0]._max_episode_steps
-    progbar = trange(max_steps, desc=f"Running eval with {max_steps} steps (maximum) per rollout.")
-    while not done.all():
-        # format from env keys to lerobot keys
-        observation = preprocess_observation(observation)
-        if return_episode_data:
-            observations.append(deepcopy(observation))
-
-        # send observation to device/gpu
-        observation = {key: observation[key].to(device, non_blocking=True) for key in observation}
-
-        # get the next action for the environment
-        with torch.inference_mode():
-            action = policy.select_action(observation, step=step)
-
-        # convert to cpu numpy
-        action = postprocess_action(action)
-
-        # apply the next action
-        observation, reward, terminated, truncated, info = env.step(action)
+    progbar = trange(n_batches, desc="Stepping through eval batches", disable=not enable_progbar)
+    for batch_ix in progbar:
+        # Cache frames for rendering videos. Each item will be (b, h, w, c), and the list indexes the rollout
+        # step.
         if max_episodes_rendered > 0:
-            render_frame(env)
+            ep_frames: list[np.ndarray] = []
 
-        # TODO(rcadene): implement a wrapper over env to return torch tensors in float32 (and cuda?)
-        action = torch.from_numpy(action)
-        reward = torch.from_numpy(reward)
-        terminated = torch.from_numpy(terminated)
-        truncated = torch.from_numpy(truncated)
-        # environment is considered done (no more steps), when success state is reached (terminated is True),
-        # or time limit is reached (truncated is True), or it was previsouly done.
-        done = terminated | truncated | done
-
-        if "final_info" in info:
-            # VectorEnv stores is_success into `info["final_info"][env_id]["is_success"]` instead of `info["is_success"]`
-            success = [
-                env_info["is_success"] if env_info is not None else False for env_info in info["final_info"]
-            ]
-        else:
-            success = [False for _ in env.envs]
-        success = torch.tensor(success)
-
-        actions.append(action)
-        rewards.append(reward)
-        dones.append(done)
-        successes.append(success)
-
-        step += 1
-        progbar.update()
-
-    env.close()
-
-    # add the last observation when the env is done
-    if return_episode_data:
-        observation = preprocess_observation(observation)
-        observations.append(deepcopy(observation))
-
-    if return_episode_data:
-        new_obses = {}
-        for key in observations[0].keys():  # noqa: SIM118
-            new_obses[key] = torch.stack([obs[key] for obs in observations], dim=1)
-        observations = new_obses
-    actions = torch.stack(actions, dim=1)
-    rewards = torch.stack(rewards, dim=1)
-    successes = torch.stack(successes, dim=1)
-    dones = torch.stack(dones, dim=1)
-
-    # Figure out where in each rollout sequence the first done condition was encountered (results after
-    # this won't be included).
-    # Note: this assumes that the shape of the done key is (batch_size, max_steps).
-    # Note: this relies on a property of argmax: that it returns the first occurrence as a tiebreaker.
-    done_indices = torch.argmax(dones.to(int), axis=1)  # (batch_size, rollout_steps)
-    expand_done_indices = done_indices[:, None].expand(-1, step)
-    expand_step_indices = torch.arange(step)[None, :].expand(num_episodes, -1)
-    mask = (expand_step_indices <= expand_done_indices).int()  # (batch_size, rollout_steps)
-    batch_sum_reward = einops.reduce((rewards * mask), "b n -> b", "sum")
-    batch_max_reward = einops.reduce((rewards * mask), "b n -> b", "max")
-    batch_success = einops.reduce((successes * mask), "b n -> b", "any")
-    sum_rewards.extend(batch_sum_reward.tolist())
-    max_rewards.extend(batch_max_reward.tolist())
-    all_successes.extend(batch_success.tolist())
-
-    # similar logic is implemented when datasets are pushed to hub (see: `push_to_hub`)
-    ep_dicts = []
-    episode_data_index = {"from": [], "to": []}
-    num_episodes = dones.shape[0]
-    total_frames = 0
-    id_from = 0
-    for ep_id in range(num_episodes):
-        num_frames = done_indices[ep_id].item() + 1
-        total_frames += num_frames
-
-        # TODO(rcadene): We need to add a missing last frame which is the observation
-        # of a done state. it is critical to have this frame for tdmpc to predict a "done observation/state"
-        if return_episode_data:
-            ep_dict = {
-                "action": actions[ep_id, :num_frames],
-                "episode_index": torch.tensor([ep_id] * num_frames),
-                "frame_index": torch.arange(0, num_frames, 1),
-                "timestamp": torch.arange(0, num_frames, 1) / fps,
-                "next.done": dones[ep_id, :num_frames],
-                "next.reward": rewards[ep_id, :num_frames].type(torch.float32),
-            }
-            for key in observations:
-                ep_dict[key] = observations[key][ep_id][:num_frames]
-            ep_dicts.append(ep_dict)
-
-            episode_data_index["from"].append(id_from)
-            episode_data_index["to"].append(id_from + num_frames)
-
-        id_from += num_frames
-
-    # similar logic is implemented in dataset preprocessing
-    if return_episode_data:
-        data_dict = {}
-        keys = ep_dicts[0].keys()
-        for key in keys:
-            if "image" not in key:
-                data_dict[key] = torch.cat([x[key] for x in ep_dicts])
-            else:
-                if key not in data_dict:
-                    data_dict[key] = []
-                for ep_dict in ep_dicts:
-                    for img in ep_dict[key]:
-                        # sanity check that images are channel first
-                        c, h, w = img.shape
-                        assert c < h and c < w, f"expect channel first images, but instead {img.shape}"
-
-                        # sanity check that images are float32 in range [0,1]
-                        assert img.dtype == torch.float32, f"expect torch.float32, but instead {img.dtype=}"
-                        assert img.max() <= 1, f"expect pixels lower than 1, but instead {img.max()=}"
-                        assert img.min() >= 0, f"expect pixels greater than 1, but instead {img.min()=}"
-
-                        # from float32 in range [0,1] to uint8 in range [0,255]
-                        img *= 255
-                        img = img.type(torch.uint8)
-
-                        # convert to channel last and numpy as expected by PIL
-                        img = PILImage.fromarray(img.permute(1, 2, 0).numpy())
-
-                        data_dict[key].append(img)
-
-        data_dict["index"] = torch.arange(0, total_frames, 1)
-        episode_data_index["from"] = torch.tensor(episode_data_index["from"])
-        episode_data_index["to"] = torch.tensor(episode_data_index["to"])
-
-        # TODO(rcadene): clean this
-        features = {}
-        for key in observations:
-            if "image" in key:
-                features[key] = Image()
-            else:
-                features[key] = Sequence(
-                    length=data_dict[key].shape[1], feature=Value(dtype="float32", id=None)
-                )
-        features.update(
-            {
-                "action": Sequence(
-                    length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
-                ),
-                "episode_index": Value(dtype="int64", id=None),
-                "frame_index": Value(dtype="int64", id=None),
-                "timestamp": Value(dtype="float32", id=None),
-                "next.reward": Value(dtype="float32", id=None),
-                "next.done": Value(dtype="bool", id=None),
-                #'next.success': Value(dtype='bool', id=None),
-                "index": Value(dtype="int64", id=None),
-            }
+        seeds = range(start_seed + (batch_ix * env.num_envs), start_seed + ((batch_ix + 1) * env.num_envs))
+        rollout_data = rollout(
+            env,
+            policy,
+            seeds=seeds,
+            return_observations=return_episode_data,
+            render_callback=render_frame if max_episodes_rendered > 0 else None,
+            enable_progbar=enable_inner_progbar,
         )
-        features = Features(features)
-        hf_dataset = Dataset.from_dict(data_dict, features=features)
-        hf_dataset.set_transform(hf_transform_to_torch)
 
-    if max_episodes_rendered > 0:
-        batch_stacked_frames = np.stack(ep_frames, 1)  # (b, t, *)
+        # Figure out where in each rollout sequence the first done condition was encountered (results after
+        # this won't be included).
+        n_steps = rollout_data["done"].shape[1]
+        # Note: this relies on a property of argmax: that it returns the first occurrence as a tiebreaker.
+        done_indices = torch.argmax(rollout_data["done"].to(int), axis=1)  # (batch_size, rollout_steps)
+        # Make a mask with shape (batch, n_steps) to mask out rollout data after the first done
+        # (batch-element-wise). Note the `done_indices + 1` to make sure to keep the data from the done step.
+        mask = (torch.arange(n_steps) <= einops.repeat(done_indices + 1, "b -> b s", s=n_steps)).int()
+        # Extend metrics.
+        batch_sum_rewards = einops.reduce((rollout_data["reward"] * mask), "b n -> b", "sum")
+        sum_rewards.extend(batch_sum_rewards.tolist())
+        batch_max_rewards = einops.reduce((rollout_data["reward"] * mask), "b n -> b", "max")
+        max_rewards.extend(batch_max_rewards.tolist())
+        batch_successes = einops.reduce((rollout_data["success"] * mask), "b n -> b", "any")
+        all_successes.extend(batch_successes.tolist())
+        all_seeds.extend(seeds)
 
-        for stacked_frames, done_index in zip(
-            batch_stacked_frames, done_indices.flatten().tolist(), strict=False
-        ):
-            if episode_counter >= max_episodes_rendered:
-                continue
-            video_dir.mkdir(parents=True, exist_ok=True)
-            video_path = video_dir / f"eval_episode_{episode_counter}.mp4"
-            thread = threading.Thread(
-                target=write_video,
-                args=(str(video_path), stacked_frames[:done_index], fps),
+        if return_episode_data:
+            this_episode_data = _compile_episode_data(
+                rollout_data,
+                done_indices,
+                start_episode_index=batch_ix * env.num_envs,
+                start_data_index=(
+                    0 if episode_data is None else (episode_data["episode_data_index"]["to"][-1].item())
+                ),
+                fps=env.unwrapped.metadata["render_fps"],
             )
-            thread.start()
-            threads.append(thread)
-            episode_counter += 1
+            if episode_data is None:
+                episode_data = this_episode_data
+            else:
+                # Some sanity checks to make sure we are not correctly compiling the data.
+                assert (
+                    episode_data["hf_dataset"]["episode_index"][-1] + 1
+                    == this_episode_data["hf_dataset"]["episode_index"][0]
+                )
+                assert (
+                    episode_data["hf_dataset"]["index"][-1] + 1 == this_episode_data["hf_dataset"]["index"][0]
+                )
+                assert torch.equal(
+                    episode_data["episode_data_index"]["to"][-1],
+                    this_episode_data["episode_data_index"]["from"][0],
+                )
+                # Concatenate the episode data.
+                episode_data = {
+                    "hf_dataset": concatenate_datasets(
+                        [episode_data["hf_dataset"], this_episode_data["hf_dataset"]]
+                    ),
+                    "episode_data_index": {
+                        k: torch.cat(
+                            [
+                                episode_data["episode_data_index"][k],
+                                this_episode_data["episode_data_index"][k],
+                            ]
+                        )
+                        for k in ["from", "to"]
+                    },
+                }
 
-        videos = einops.rearrange(batch_stacked_frames, "b t h w c -> b t c h w")
+        # Maybe render video for visualization.
+        if max_episodes_rendered > 0 and len(ep_frames) > 0:
+            batch_stacked_frames = np.stack(ep_frames, axis=1)  # (b, t, *)
+            for stacked_frames, done_index in zip(
+                batch_stacked_frames, done_indices.flatten().tolist(), strict=False
+            ):
+                if n_episodes_rendered >= max_episodes_rendered:
+                    break
+                video_dir.mkdir(parents=True, exist_ok=True)
+                video_path = video_dir / f"eval_episode_{n_episodes_rendered}.mp4"
+                video_paths.append(str(video_path))
+                thread = threading.Thread(
+                    target=write_video,
+                    args=(
+                        str(video_path),
+                        stacked_frames[: done_index + 1],  # + 1 to capture the last observation
+                        env.unwrapped.metadata["render_fps"],
+                    ),
+                )
+                thread.start()
+                threads.append(thread)
+                n_episodes_rendered += 1
 
+        progbar.set_postfix(
+            {"running_success_rate": f"{np.mean(all_successes[:n_episodes]).item() * 100:.1f}%"}
+        )
+
+    # Wait till all video rendering threads are done.
     for thread in threads:
         thread.join()
 
+    # Compile eval info.
     info = {
         "per_episode": [
             {
@@ -323,80 +366,209 @@ def eval_policy(
             }
             for i, (sum_reward, max_reward, success, seed) in enumerate(
                 zip(
-                    sum_rewards[:num_episodes],
-                    max_rewards[:num_episodes],
-                    all_successes[:num_episodes],
-                    seeds[:num_episodes],
+                    sum_rewards[:n_episodes],
+                    max_rewards[:n_episodes],
+                    all_successes[:n_episodes],
+                    all_seeds[:n_episodes],
                     strict=True,
                 )
             )
         ],
         "aggregated": {
-            "avg_sum_reward": float(np.nanmean(sum_rewards[:num_episodes])),
-            "avg_max_reward": float(np.nanmean(max_rewards[:num_episodes])),
-            "pc_success": float(np.nanmean(all_successes[:num_episodes]) * 100),
+            "avg_sum_reward": float(np.nanmean(sum_rewards[:n_episodes])),
+            "avg_max_reward": float(np.nanmean(max_rewards[:n_episodes])),
+            "pc_success": float(np.nanmean(all_successes[:n_episodes]) * 100),
             "eval_s": time.time() - start,
-            "eval_ep_s": (time.time() - start) / num_episodes,
+            "eval_ep_s": (time.time() - start) / n_episodes,
         },
     }
+
     if return_episode_data:
-        info["episodes"] = {
-            "hf_dataset": hf_dataset,
-            "episode_data_index": episode_data_index,
-        }
+        info["episodes"] = episode_data
+
     if max_episodes_rendered > 0:
-        info["videos"] = videos
+        info["video_paths"] = video_paths
+
     return info
 
 
-def eval(cfg: dict, out_dir=None):
+def _compile_episode_data(
+    rollout_data: dict, done_indices: Tensor, start_episode_index: int, start_data_index: int, fps: float
+) -> dict:
+    """Convenience function for `eval_policy(return_episode_data=True)`
+
+    Compiles all the rollout data into a Hugging Face dataset.
+
+    Similar logic is implemented when datasets are pushed to hub (see: `push_to_hub`).
+    """
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+    total_frames = 0
+    data_index_from = start_data_index
+    for ep_ix in range(rollout_data["action"].shape[0]):
+        num_frames = done_indices[ep_ix].item() + 1  # + 1 to include the first done frame
+        total_frames += num_frames
+
+        # TODO(rcadene): We need to add a missing last frame which is the observation
+        # of a done state. it is critical to have this frame for tdmpc to predict a "done observation/state"
+        ep_dict = {
+            "action": rollout_data["action"][ep_ix, :num_frames],
+            "episode_index": torch.tensor([start_episode_index + ep_ix] * num_frames),
+            "frame_index": torch.arange(0, num_frames, 1),
+            "timestamp": torch.arange(0, num_frames, 1) / fps,
+            "next.done": rollout_data["done"][ep_ix, :num_frames],
+            "next.reward": rollout_data["reward"][ep_ix, :num_frames].type(torch.float32),
+        }
+        for key in rollout_data["observation"]:
+            ep_dict[key] = rollout_data["observation"][key][ep_ix][:num_frames]
+        ep_dicts.append(ep_dict)
+
+        episode_data_index["from"].append(data_index_from)
+        episode_data_index["to"].append(data_index_from + num_frames)
+
+        data_index_from += num_frames
+
+    data_dict = {}
+    for key in ep_dicts[0]:
+        if "image" not in key:
+            data_dict[key] = torch.cat([x[key] for x in ep_dicts])
+        else:
+            if key not in data_dict:
+                data_dict[key] = []
+            for ep_dict in ep_dicts:
+                for img in ep_dict[key]:
+                    # sanity check that images are channel first
+                    c, h, w = img.shape
+                    assert c < h and c < w, f"expect channel first images, but instead {img.shape}"
+
+                    # sanity check that images are float32 in range [0,1]
+                    assert img.dtype == torch.float32, f"expect torch.float32, but instead {img.dtype=}"
+                    assert img.max() <= 1, f"expect pixels lower than 1, but instead {img.max()=}"
+                    assert img.min() >= 0, f"expect pixels greater than 1, but instead {img.min()=}"
+
+                    # from float32 in range [0,1] to uint8 in range [0,255]
+                    img *= 255
+                    img = img.type(torch.uint8)
+
+                    # convert to channel last and numpy as expected by PIL
+                    img = PILImage.fromarray(img.permute(1, 2, 0).numpy())
+
+                    data_dict[key].append(img)
+
+    data_dict["index"] = torch.arange(start_data_index, start_data_index + total_frames, 1)
+    episode_data_index["from"] = torch.tensor(episode_data_index["from"])
+    episode_data_index["to"] = torch.tensor(episode_data_index["to"])
+
+    # TODO(rcadene): clean this
+    features = {}
+    for key in rollout_data["observation"]:
+        if "image" in key:
+            features[key] = Image()
+        else:
+            features[key] = Sequence(length=data_dict[key].shape[1], feature=Value(dtype="float32", id=None))
+    features.update(
+        {
+            "action": Sequence(length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)),
+            "episode_index": Value(dtype="int64", id=None),
+            "frame_index": Value(dtype="int64", id=None),
+            "timestamp": Value(dtype="float32", id=None),
+            "next.reward": Value(dtype="float32", id=None),
+            "next.done": Value(dtype="bool", id=None),
+            #'next.success': Value(dtype='bool', id=None),
+            "index": Value(dtype="int64", id=None),
+        }
+    )
+    features = Features(features)
+    hf_dataset = Dataset.from_dict(data_dict, features=features)
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return {
+        "hf_dataset": hf_dataset,
+        "episode_data_index": episode_data_index,
+    }
+
+
+def eval(
+    pretrained_policy_path: str | None = None,
+    hydra_cfg_path: str | None = None,
+    config_overrides: list[str] | None = None,
+):
+    assert (pretrained_policy_path is None) ^ (hydra_cfg_path is None)
+    if hydra_cfg_path is None:
+        hydra_cfg = init_hydra_config(pretrained_policy_path / "config.yaml", config_overrides)
+    else:
+        hydra_cfg = init_hydra_config(hydra_cfg_path, config_overrides)
+    out_dir = (
+        f"outputs/eval/{dt.now().strftime('%Y-%m-%d/%H-%M-%S')}_{hydra_cfg.env.name}_{hydra_cfg.policy.name}"
+    )
+
     if out_dir is None:
         raise NotImplementedError()
 
-    init_logging()
-
     # Check device is available
-    get_safe_torch_device(cfg.device, log=True)
+    get_safe_torch_device(hydra_cfg.device, log=True)
 
     torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
-    set_global_seed(cfg.seed)
+    set_global_seed(hydra_cfg.seed)
 
     log_output_dir(out_dir)
 
     logging.info("Making environment.")
-    env = make_env(cfg, num_parallel_envs=cfg.eval_episodes)
+    env = make_env(hydra_cfg)
 
     logging.info("Making policy.")
-    policy = make_policy(cfg)
+    if hydra_cfg_path is None:
+        policy = make_policy(hydra_cfg=hydra_cfg, pretrained_policy_name_or_path=pretrained_policy_path)
+    else:
+        # Note: We need the dataset stats to pass to the policy's normalization modules.
+        policy = make_policy(hydra_cfg=hydra_cfg, dataset_stats=make_dataset(hydra_cfg).stats)
+    policy.eval()
 
     info = eval_policy(
         env,
         policy,
+        hydra_cfg.eval.n_episodes,
         max_episodes_rendered=10,
         video_dir=Path(out_dir) / "eval",
-        return_episode_data=False,
-        seed=cfg.seed,
+        start_seed=hydra_cfg.seed,
+        enable_progbar=True,
+        enable_inner_progbar=True,
     )
     print(info["aggregated"])
 
     # Save info
     with open(Path(out_dir) / "eval_info.json", "w") as f:
-        # remove pytorch tensors which are not serializable to save the evaluation results only
-        del info["videos"]
         json.dump(info, f, indent=2)
 
+    env.close()
+
     logging.info("End of eval")
 
 
 if __name__ == "__main__":
+    init_logging()
+
     parser = argparse.ArgumentParser(
         description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
     )
     group = parser.add_mutually_exclusive_group(required=True)
-    group.add_argument("--config", help="Path to a specific yaml config you want to use.")
-    group.add_argument("--hub-id", help="HuggingFace Hub ID for a pretrained model.")
-    parser.add_argument("--revision", help="Optionally provide the HuggingFace Hub revision ID.")
+    group.add_argument(
+        "-p",
+        "--pretrained-policy-name-or-path",
+        help=(
+            "Either the repo ID of a model hosted on the Hub or a path to a directory containing weights "
+            "saved using `Policy.save_pretrained`. If not provided, the policy is initialized from scratch "
+            "(useful for debugging). This argument is mutually exclusive with `--config`."
+        ),
+    )
+    group.add_argument(
+        "--config",
+        help=(
+            "Path to a yaml config you want to use for initializing a policy from scratch (useful for "
+            "debugging). This argument is mutually exclusive with `--pretrained-policy-name-or-path` (`-p`)."
+        ),
+    )
+    parser.add_argument("--revision", help="Optionally provide the Hugging Face Hub revision ID.")
     parser.add_argument(
         "overrides",
         nargs="*",
@@ -404,16 +576,29 @@ if __name__ == "__main__":
     )
     args = parser.parse_args()
 
-    if args.config is not None:
-        # Note: For the config_path, Hydra wants a path relative to this script file.
-        cfg = init_hydra_config(args.config, args.overrides)
-    elif args.hub_id is not None:
-        folder = Path(snapshot_download(args.hub_id, revision=args.revision))
-        cfg = init_hydra_config(
-            folder / "config.yaml", [f"policy.pretrained_model_path={folder / 'model.pt'}", *args.overrides]
-        )
+    if args.pretrained_policy_name_or_path is None:
+        eval(hydra_cfg_path=args.config, config_overrides=args.overrides)
+    else:
+        try:
+            pretrained_policy_path = Path(
+                snapshot_download(args.pretrained_policy_name_or_path, revision=args.revision)
+            )
+        except (HFValidationError, RepositoryNotFoundError) as e:
+            if isinstance(e, HFValidationError):
+                error_message = (
+                    "The provided pretrained_policy_name_or_path is not a valid Hugging Face Hub repo ID."
+                )
+            else:
+                error_message = (
+                    "The provided pretrained_policy_name_or_path was not found on the Hugging Face Hub."
+                )
 
-    eval(
-        cfg,
-        out_dir=f"outputs/eval/{dt.now().strftime('%Y-%m-%d/%H-%M-%S')}_{cfg.env.name}_{cfg.policy.name}",
-    )
+            logging.warning(f"{error_message} Treating it as a local directory.")
+            pretrained_policy_path = Path(args.pretrained_policy_name_or_path)
+        if not pretrained_policy_path.is_dir() or not pretrained_policy_path.exists():
+            raise ValueError(
+                "The provided pretrained_policy_name_or_path is not a valid/existing Hugging Face Hub "
+                "repo ID, nor is it an existing local directory."
+            )
+
+        eval(pretrained_policy_path=pretrained_policy_path, config_overrides=args.overrides)

lerobot/scripts/push_dataset_to_hub.py

@@ -1,295 +1,247 @@
+"""
+Use this script to convert your dataset into LeRobot dataset format and upload it to the Hugging Face hub,
+or store it locally. LeRobot dataset format is lightweight, fast to load from, and does not require any
+installation of neural net specific packages like pytorch, tensorflow, jax.
+
+Example:
+```
+python lerobot/scripts/push_dataset_to_hub.py \
+--data-dir data \
+--dataset-id pusht \
+--raw-format pusht_zarr \
+--community-id lerobot \
+--revision v1.2 \
+--dry-run 1 \
+--save-to-disk 1 \
+--save-tests-to-disk 0 \
+--debug 1
+
+python lerobot/scripts/push_dataset_to_hub.py \
+--data-dir data \
+--dataset-id xarm_lift_medium \
+--raw-format xarm_pkl \
+--community-id lerobot \
+--revision v1.2 \
+--dry-run 1 \
+--save-to-disk 1 \
+--save-tests-to-disk 0 \
+--debug 1
+
+python lerobot/scripts/push_dataset_to_hub.py \
+--data-dir data \
+--dataset-id aloha_sim_insertion_scripted \
+--raw-format aloha_hdf5 \
+--community-id lerobot \
+--revision v1.2 \
+--dry-run 1 \
+--save-to-disk 1 \
+--save-tests-to-disk 0 \
+--debug 1
+
+python lerobot/scripts/push_dataset_to_hub.py \
+--data-dir data \
+--dataset-id umi_cup_in_the_wild \
+--raw-format umi_zarr \
+--community-id lerobot \
+--revision v1.2 \
+--dry-run 1 \
+--save-to-disk 1 \
+--save-tests-to-disk 0 \
+--debug 1
+```
+"""
+
 import argparse
 import json
 import shutil
 from pathlib import Path
-from typing import Any, Protocol
 
 import torch
-from datasets import Dataset
 from huggingface_hub import HfApi
 from safetensors.torch import save_file
 
+from lerobot.common.datasets.lerobot_dataset import CODEBASE_VERSION, LeRobotDataset
 from lerobot.common.datasets.push_dataset_to_hub._download_raw import download_raw
-from lerobot.common.datasets.push_dataset_to_hub.aloha_processor import (
-    AlohaProcessor,
-)
-from lerobot.common.datasets.push_dataset_to_hub.pusht_processor import PushTProcessor
-from lerobot.common.datasets.push_dataset_to_hub.umi_processor import UmiProcessor
-from lerobot.common.datasets.push_dataset_to_hub.xarm_processor import XarmProcessor
-from lerobot.common.datasets.utils import compute_stats, flatten_dict
+from lerobot.common.datasets.push_dataset_to_hub.compute_stats import compute_stats
+from lerobot.common.datasets.utils import flatten_dict
 
 
-def push_lerobot_dataset_to_hub(
-    hf_dataset: Dataset,
-    episode_data_index: dict[str, list[int]],
-    info: dict[str, Any],
-    stats: dict[str, dict[str, torch.Tensor]],
-    root: Path,
-    revision: str,
-    dataset_id: str,
-    community_id: str = "lerobot",
-    dry_run: bool = False,
-) -> None:
-    """
-    Pushes a dataset to the Hugging Face Hub.
+def get_from_raw_to_lerobot_format_fn(raw_format):
+    if raw_format == "pusht_zarr":
+        from lerobot.common.datasets.push_dataset_to_hub.pusht_zarr_format import from_raw_to_lerobot_format
+    elif raw_format == "umi_zarr":
+        from lerobot.common.datasets.push_dataset_to_hub.umi_zarr_format import from_raw_to_lerobot_format
+    elif raw_format == "aloha_hdf5":
+        from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import from_raw_to_lerobot_format
+    elif raw_format == "xarm_pkl":
+        from lerobot.common.datasets.push_dataset_to_hub.xarm_pkl_format import from_raw_to_lerobot_format
+    else:
+        raise ValueError(raw_format)
 
-    Args:
-        hf_dataset (Dataset): The dataset to be pushed.
-        episode_data_index (dict[str, list[int]]): The index of episode data.
-        info (dict[str, Any]): Information about the dataset, eg. fps.
-        stats (dict[str, dict[str, torch.Tensor]]): Statistics of the dataset.
-        root (Path): The root directory of the dataset.
-        revision (str): The revision of the dataset.
-        dataset_id (str): The ID of the dataset.
-        community_id (str, optional): The ID of the community or the user where the
-            dataset will be stored. Defaults to "lerobot".
-        dry_run (bool, optional): If True, performs a dry run without actually pushing the dataset. Defaults to False.
-    """
-    if not dry_run:
-        # push to main to indicate latest version
-        hf_dataset.push_to_hub(f"{community_id}/{dataset_id}", token=True)
+    return from_raw_to_lerobot_format
 
-        # push to version branch
-        hf_dataset.push_to_hub(f"{community_id}/{dataset_id}", token=True, revision=revision)
 
-    # create and store meta_data
-    meta_data_dir = root / community_id / dataset_id / "meta_data"
+def save_meta_data(info, stats, episode_data_index, meta_data_dir):
     meta_data_dir.mkdir(parents=True, exist_ok=True)
 
-    # info
+    # save info
     info_path = meta_data_dir / "info.json"
-
     with open(str(info_path), "w") as f:
         json.dump(info, f, indent=4)
-    # stats
+
+    # save stats
     stats_path = meta_data_dir / "stats.safetensors"
     save_file(flatten_dict(stats), stats_path)
 
-    # episode_data_index
+    # save episode_data_index
     episode_data_index = {key: torch.tensor(episode_data_index[key]) for key in episode_data_index}
     ep_data_idx_path = meta_data_dir / "episode_data_index.safetensors"
     save_file(episode_data_index, ep_data_idx_path)
 
-    if not dry_run:
-        api = HfApi()
 
-        api.upload_file(
-            path_or_fileobj=info_path,
-            path_in_repo=str(info_path).replace(f"{root}/{community_id}/{dataset_id}", ""),
-            repo_id=f"{community_id}/{dataset_id}",
-            repo_type="dataset",
-        )
-        api.upload_file(
-            path_or_fileobj=info_path,
-            path_in_repo=str(info_path).replace(f"{root}/{community_id}/{dataset_id}", ""),
-            repo_id=f"{community_id}/{dataset_id}",
-            repo_type="dataset",
-            revision=revision,
-        )
-
-        # stats
-        api.upload_file(
-            path_or_fileobj=stats_path,
-            path_in_repo=str(stats_path).replace(f"{root}/{community_id}/{dataset_id}", ""),
-            repo_id=f"{community_id}/{dataset_id}",
-            repo_type="dataset",
-        )
-        api.upload_file(
-            path_or_fileobj=stats_path,
-            path_in_repo=str(stats_path).replace(f"{root}/{community_id}/{dataset_id}", ""),
-            repo_id=f"{community_id}/{dataset_id}",
-            repo_type="dataset",
-            revision=revision,
-        )
-
-        api.upload_file(
-            path_or_fileobj=ep_data_idx_path,
-            path_in_repo=str(ep_data_idx_path).replace(f"{root}/{community_id}/{dataset_id}", ""),
-            repo_id=f"{community_id}/{dataset_id}",
-            repo_type="dataset",
-        )
-        api.upload_file(
-            path_or_fileobj=ep_data_idx_path,
-            path_in_repo=str(ep_data_idx_path).replace(f"{root}/{community_id}/{dataset_id}", ""),
-            repo_id=f"{community_id}/{dataset_id}",
-            repo_type="dataset",
-            revision=revision,
-        )
-
-    # copy in tests folder, the first episode and the meta_data directory
-    num_items_first_ep = episode_data_index["to"][0] - episode_data_index["from"][0]
-    hf_dataset.select(range(num_items_first_ep)).with_format("torch").save_to_disk(
-        f"tests/data/{community_id}/{dataset_id}/train"
+def push_meta_data_to_hub(repo_id, meta_data_dir, revision):
+    """Expect all meta data files to be all stored in a single "meta_data" directory.
+    On the hugging face repositery, they will be uploaded in a "meta_data" directory at the root.
+    """
+    api = HfApi()
+    api.upload_folder(
+        folder_path=meta_data_dir,
+        path_in_repo="meta_data",
+        repo_id=repo_id,
+        revision=revision,
+        repo_type="dataset",
+    )
+
+
+def push_videos_to_hub(repo_id, videos_dir, revision):
+    """Expect mp4 files to be all stored in a single "videos" directory.
+    On the hugging face repositery, they will be uploaded in a "videos" directory at the root.
+    """
+    api = HfApi()
+    api.upload_folder(
+        folder_path=videos_dir,
+        path_in_repo="videos",
+        repo_id=repo_id,
+        revision=revision,
+        repo_type="dataset",
+        allow_patterns="*.mp4",
     )
-    if Path(f"tests/data/{community_id}/{dataset_id}/meta_data").exists():
-        shutil.rmtree(f"tests/data/{community_id}/{dataset_id}/meta_data")
-    shutil.copytree(meta_data_dir, f"tests/data/{community_id}/{dataset_id}/meta_data")
 
 
 def push_dataset_to_hub(
+    data_dir: Path,
     dataset_id: str,
-    root: Path,
+    raw_format: str | None,
+    community_id: str,
+    revision: str,
+    dry_run: bool,
+    save_to_disk: bool,
+    tests_data_dir: Path,
+    save_tests_to_disk: bool,
     fps: int | None,
-    dataset_folder: Path | None = None,
-    dry_run: bool = False,
-    revision: str = "v1.1",
-    community_id: str = "lerobot",
-    no_preprocess: bool = False,
-    path_save_to_disk: str | None = None,
-    **kwargs,
-) -> None:
-    """
-    Download a raw dataset if needed or access a local raw dataset, detect the raw format (e.g. aloha, pusht, umi) and process it accordingly in a common data format which is then pushed to the Hugging Face Hub.
+    video: bool,
+    batch_size: int,
+    num_workers: int,
+    debug: bool,
+):
+    repo_id = f"{community_id}/{dataset_id}"
 
-    Args:
-        dataset_id (str): The ID of the dataset.
-        root (Path): The root directory where the dataset will be downloaded.
-        fps (int | None): The desired frames per second for the dataset.
-        dataset_folder (Path | None, optional): The path to the dataset folder. If not provided, the dataset will be downloaded using the dataset ID. Defaults to None.
-        dry_run (bool, optional): If True, performs a dry run without actually pushing the dataset. Defaults to False.
-        revision (str, optional): Version of the `push_dataset_to_hub.py` codebase used to preprocess the dataset. Defaults to "v1.1".
-        community_id (str, optional): The ID of the community. Defaults to "lerobot".
-        no_preprocess (bool, optional): If True, does not preprocesses the dataset. Defaults to False.
-        path_save_to_disk (str | None, optional): The path to save the dataset to disk. Works when `dry_run` is True, which allows to only save on disk without uploading. By default, the dataset is not saved on disk.
-        **kwargs: Additional keyword arguments for the preprocessor init method.
+    raw_dir = data_dir / f"{dataset_id}_raw"
 
+    out_dir = data_dir / repo_id
+    meta_data_dir = out_dir / "meta_data"
+    videos_dir = out_dir / "videos"
 
-    """
-    if dataset_folder is None:
-        dataset_folder = download_raw(root=root, dataset_id=dataset_id)
+    tests_out_dir = tests_data_dir / repo_id
+    tests_meta_data_dir = tests_out_dir / "meta_data"
+    tests_videos_dir = tests_out_dir / "videos"
 
-    if not no_preprocess:
-        processor = guess_dataset_type(dataset_folder=dataset_folder, fps=fps, **kwargs)
-        data_dict, episode_data_index = processor.preprocess()
-        hf_dataset = processor.to_hf_dataset(data_dict)
+    if out_dir.exists():
+        shutil.rmtree(out_dir)
 
-        info = {
-            "fps": processor.fps,
-        }
-        stats: dict[str, dict[str, torch.Tensor]] = compute_stats(hf_dataset)
+    if tests_out_dir.exists() and save_tests_to_disk:
+        shutil.rmtree(tests_out_dir)
 
-        push_lerobot_dataset_to_hub(
-            hf_dataset=hf_dataset,
-            episode_data_index=episode_data_index,
-            info=info,
-            stats=stats,
-            root=root,
-            revision=revision,
-            dataset_id=dataset_id,
-            community_id=community_id,
-            dry_run=dry_run,
-        )
-        if path_save_to_disk:
-            hf_dataset.with_format("torch").save_to_disk(dataset_path=str(path_save_to_disk))
+    if not raw_dir.exists():
+        download_raw(raw_dir, dataset_id)
 
-        processor.cleanup()
+    if raw_format is None:
+        # TODO(rcadene, adilzouitine): implement auto_find_raw_format
+        raise NotImplementedError()
+        # raw_format = auto_find_raw_format(raw_dir)
 
+    from_raw_to_lerobot_format = get_from_raw_to_lerobot_format_fn(raw_format)
 
-class DatasetProcessor(Protocol):
-    """A class for processing datasets.
+    # convert dataset from original raw format to LeRobot format
+    hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(raw_dir, out_dir, fps, video, debug)
 
-    This class provides methods for validating, preprocessing, and converting datasets.
+    lerobot_dataset = LeRobotDataset.from_preloaded(
+        repo_id=repo_id,
+        version=revision,
+        hf_dataset=hf_dataset,
+        episode_data_index=episode_data_index,
+        info=info,
+        videos_dir=videos_dir,
+    )
+    stats = compute_stats(lerobot_dataset, batch_size, num_workers)
 
-    Args:
-        folder_path (str): The path to the folder containing the dataset.
-        fps (int | None): The frames per second of the dataset. If None, the default value is used.
-        *args: Additional positional arguments.
-        **kwargs: Additional keyword arguments.
-    """
+    if save_to_disk:
+        hf_dataset = hf_dataset.with_format(None)  # to remove transforms that cant be saved
+        hf_dataset.save_to_disk(str(out_dir / "train"))
 
-    def __init__(self, folder_path: str, fps: int | None, *args, **kwargs) -> None: ...
+    if not dry_run or save_to_disk:
+        # mandatory for upload
+        save_meta_data(info, stats, episode_data_index, meta_data_dir)
 
-    def is_valid(self) -> bool:
-        """Check if the dataset is valid.
+    if not dry_run:
+        hf_dataset.push_to_hub(repo_id, token=True, revision="main")
+        hf_dataset.push_to_hub(repo_id, token=True, revision=revision)
 
-        Returns:
-            bool: True if the dataset is valid, False otherwise.
-        """
-        ...
+        push_meta_data_to_hub(repo_id, meta_data_dir, revision="main")
+        push_meta_data_to_hub(repo_id, meta_data_dir, revision=revision)
 
-    def preprocess(self) -> tuple[dict, dict]:
-        """Preprocess the dataset.
+        if video:
+            push_videos_to_hub(repo_id, videos_dir, revision="main")
+            push_videos_to_hub(repo_id, videos_dir, revision=revision)
 
-        Returns:
-            tuple[dict, dict]: A tuple containing two dictionaries representing the preprocessed data.
-        """
-        ...
+    if save_tests_to_disk:
+        # get the first episode
+        num_items_first_ep = episode_data_index["to"][0] - episode_data_index["from"][0]
+        test_hf_dataset = hf_dataset.select(range(num_items_first_ep))
 
-    def to_hf_dataset(self, data_dict: dict) -> Dataset:
-        """Convert the preprocessed data to a Hugging Face dataset.
+        test_hf_dataset = test_hf_dataset.with_format(None)
+        test_hf_dataset.save_to_disk(str(tests_out_dir / "train"))
 
-        Args:
-            data_dict (dict): The preprocessed data.
+        # copy meta data to tests directory
+        shutil.copytree(meta_data_dir, tests_meta_data_dir)
 
-        Returns:
-            Dataset: The converted Hugging Face dataset.
-        """
-        ...
-
-    @property
-    def fps(self) -> int:
-        """Get the frames per second of the dataset.
-
-        Returns:
-            int: The frames per second.
-        """
-        ...
-
-    def cleanup(self):
-        """Clean up any resources used by the dataset processor."""
-        ...
-
-
-def guess_dataset_type(dataset_folder: Path, **processor_kwargs) -> DatasetProcessor:
-    if (processor := AlohaProcessor(folder_path=dataset_folder, **processor_kwargs)).is_valid():
-        return processor
-    if (processor := XarmProcessor(folder_path=dataset_folder, **processor_kwargs)).is_valid():
-        return processor
-    if (processor := PushTProcessor(folder_path=dataset_folder, **processor_kwargs)).is_valid():
-        return processor
-    if (processor := UmiProcessor(folder_path=dataset_folder, **processor_kwargs)).is_valid():
-        return processor
-    # TODO: Propose a registration mechanism for new dataset types
-    raise ValueError(f"Could not guess dataset type for folder {dataset_folder}")
+        # copy videos of first episode to tests directory
+        episode_index = 0
+        tests_videos_dir.mkdir(parents=True, exist_ok=True)
+        for key in lerobot_dataset.video_frame_keys:
+            fname = f"{key}_episode_{episode_index:06d}.mp4"
+            shutil.copy(videos_dir / fname, tests_videos_dir / fname)
 
 
 def main():
-    """
-    Main function to process command line arguments and push dataset to Hugging Face Hub.
-
-    Parses command line arguments to get dataset details and conditions under which the dataset
-    is processed and pushed. It manages dataset preparation and uploading based on the user-defined parameters.
-    """
-    parser = argparse.ArgumentParser(
-        description="Push a dataset to the Hugging Face Hub with optional parameters for customization.",
-        epilog="""
-        Example usage:
-            python -m lerobot.scripts.push_dataset_to_hub --dataset-folder /path/to/dataset --dataset-id example_dataset --root /path/to/root --dry-run --revision v2.0 --community-id example_community --fps 30 --path-save-to-disk /path/to/save --no-preprocess
-
-        This processes and optionally pushes 'example_dataset' located in '/path/to/dataset' to Hugging Face Hub,
-        with various parameters to control the processing and uploading behavior.
-        """,
-    )
+    parser = argparse.ArgumentParser()
 
     parser.add_argument(
-        "--dataset-folder",
+        "--data-dir",
         type=Path,
-        default=None,
-        help="The filesystem path to the dataset folder. If not provided, the dataset must be identified and managed by other means.",
+        required=True,
+        help="Root directory containing datasets (e.g. `data` or `tmp/data` or `/tmp/lerobot/data`).",
     )
     parser.add_argument(
         "--dataset-id",
         type=str,
         required=True,
-        help="Unique identifier for the dataset to be processed and uploaded.",
+        help="Name of the dataset (e.g. `pusht`, `aloha_sim_insertion_human`), which matches the folder where the data is stored (e.g. `data/pusht`).",
     )
     parser.add_argument(
-        "--root", type=Path, required=True, help="Root directory where the dataset operations are managed."
-    )
-    parser.add_argument(
-        "--dry-run",
-        action="store_true",
-        help="Simulate the push process without uploading any data, for testing purposes.",
+        "--raw-format",
+        type=str,
+        help="Dataset type (e.g. `pusht_zarr`, `umi_zarr`, `aloha_hdf5`, `xarm_pkl`). If not provided, will be detected automatically.",
     )
     parser.add_argument(
         "--community-id",
@@ -297,41 +249,68 @@ def main():
         default="lerobot",
         help="Community or user ID under which the dataset will be hosted on the Hub.",
     )
-    parser.add_argument(
-        "--fps",
-        type=int,
-        help="Target frame rate for video or image sequence datasets. Optional and applicable only if the dataset includes temporal media.",
-    )
     parser.add_argument(
         "--revision",
         type=str,
-        default="v1.0",
-        help="Dataset version identifier to manage different iterations of the dataset.",
+        default=CODEBASE_VERSION,
+        help="Codebase version used to generate the dataset.",
     )
     parser.add_argument(
-        "--no-preprocess",
-        action="store_true",
-        help="Does not preprocess the dataset, set this flag if you only want dowload the dataset raw.",
+        "--dry-run",
+        type=int,
+        default=0,
+        help="Run everything without uploading to hub, for testing purposes or storing a dataset locally.",
     )
     parser.add_argument(
-        "--path-save-to-disk",
+        "--save-to-disk",
+        type=int,
+        default=1,
+        help="Save the dataset in the directory specified by `--data-dir`.",
+    )
+    parser.add_argument(
+        "--tests-data-dir",
         type=Path,
-        help="Optional path where the processed dataset can be saved locally.",
+        default="tests/data",
+        help="Directory containing tests artifacts datasets.",
+    )
+    parser.add_argument(
+        "--save-tests-to-disk",
+        type=int,
+        default=1,
+        help="Save the dataset with 1 episode used for unit tests in the directory specified by `--tests-data-dir`.",
+    )
+    parser.add_argument(
+        "--fps",
+        type=int,
+        help="Frame rate used to collect videos. If not provided, use the default one specified in the code.",
+    )
+    parser.add_argument(
+        "--video",
+        type=int,
+        default=1,
+        help="Convert each episode of the raw dataset to an mp4 video. This option allows 60 times lower disk space consumption and 25 faster loading time during training.",
+    )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=32,
+        help="Batch size loaded by DataLoader for computing the dataset statistics.",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=16,
+        help="Number of processes of Dataloader for computing the dataset statistics.",
+    )
+    parser.add_argument(
+        "--debug",
+        type=int,
+        default=0,
+        help="Debug mode process the first episode only.",
     )
 
     args = parser.parse_args()
-
-    push_dataset_to_hub(
-        dataset_folder=args.dataset_folder,
-        dataset_id=args.dataset_id,
-        root=args.root,
-        fps=args.fps,
-        dry_run=args.dry_run,
-        community_id=args.community_id,
-        revision=args.revision,
-        no_preprocess=args.no_preprocess,
-        path_save_to_disk=args.path_save_to_disk,
-    )
+    push_dataset_to_hub(**vars(args))
 
 
 if __name__ == "__main__":

lerobot/scripts/train.py

@@ -1,4 +1,5 @@
 import logging
+import time
 from copy import deepcopy
 from pathlib import Path
 
@@ -13,6 +14,7 @@ from lerobot.common.datasets.utils import cycle
 from lerobot.common.envs.factory import make_env
 from lerobot.common.logger import Logger, log_output_dir
 from lerobot.common.policies.factory import make_policy
+from lerobot.common.policies.policy_protocol import PolicyWithUpdate
 from lerobot.common.utils.utils import (
     format_big_number,
     get_safe_torch_device,
@@ -22,6 +24,86 @@ from lerobot.common.utils.utils import (
 from lerobot.scripts.eval import eval_policy
 
 
+def make_optimizer_and_scheduler(cfg, policy):
+    if cfg.policy.name == "act":
+        optimizer_params_dicts = [
+            {
+                "params": [
+                    p
+                    for n, p in policy.named_parameters()
+                    if not n.startswith("backbone") and p.requires_grad
+                ]
+            },
+            {
+                "params": [
+                    p for n, p in policy.named_parameters() if n.startswith("backbone") and p.requires_grad
+                ],
+                "lr": cfg.training.lr_backbone,
+            },
+        ]
+        optimizer = torch.optim.AdamW(
+            optimizer_params_dicts, lr=cfg.training.lr, weight_decay=cfg.training.weight_decay
+        )
+        lr_scheduler = None
+    elif cfg.policy.name == "diffusion":
+        optimizer = torch.optim.Adam(
+            policy.diffusion.parameters(),
+            cfg.training.lr,
+            cfg.training.adam_betas,
+            cfg.training.adam_eps,
+            cfg.training.adam_weight_decay,
+        )
+        assert cfg.training.online_steps == 0, "Diffusion Policy does not handle online training."
+        from diffusers.optimization import get_scheduler
+
+        lr_scheduler = get_scheduler(
+            cfg.training.lr_scheduler,
+            optimizer=optimizer,
+            num_warmup_steps=cfg.training.lr_warmup_steps,
+            num_training_steps=cfg.training.offline_steps,
+        )
+    elif policy.name == "tdmpc":
+        optimizer = torch.optim.Adam(policy.parameters(), cfg.training.lr)
+        lr_scheduler = None
+    else:
+        raise NotImplementedError()
+
+    return optimizer, lr_scheduler
+
+
+def update_policy(policy, batch, optimizer, grad_clip_norm, lr_scheduler=None):
+    start_time = time.time()
+    policy.train()
+    output_dict = policy.forward(batch)
+    # TODO(rcadene): policy.unnormalize_outputs(out_dict)
+    loss = output_dict["loss"]
+    loss.backward()
+    grad_norm = torch.nn.utils.clip_grad_norm_(
+        policy.parameters(),
+        grad_clip_norm,
+        error_if_nonfinite=False,
+    )
+
+    optimizer.step()
+    optimizer.zero_grad()
+
+    if lr_scheduler is not None:
+        lr_scheduler.step()
+
+    if isinstance(policy, PolicyWithUpdate):
+        # To possibly update an internal buffer (for instance an Exponential Moving Average like in TDMPC).
+        policy.update()
+
+    info = {
+        "loss": loss.item(),
+        "grad_norm": float(grad_norm),
+        "lr": optimizer.param_groups[0]["lr"],
+        "update_s": time.time() - start_time,
+    }
+
+    return info
+
+
 @hydra.main(version_base="1.2", config_name="default", config_path="../configs")
 def train_cli(cfg: dict):
     train(
@@ -48,7 +130,7 @@ def log_train_info(logger, info, step, cfg, dataset, is_offline):
 
     # A sample is an (observation,action) pair, where observation and action
     # can be on multiple timestamps. In a batch, we have `batch_size`` number of samples.
-    num_samples = (step + 1) * cfg.policy.batch_size
+    num_samples = (step + 1) * cfg.training.batch_size
     avg_samples_per_ep = dataset.num_samples / dataset.num_episodes
     num_episodes = num_samples / avg_samples_per_ep
     num_epochs = num_samples / dataset.num_samples
@@ -84,7 +166,7 @@ def log_eval_info(logger, info, step, cfg, dataset, is_offline):
 
     # A sample is an (observation,action) pair, where observation and action
     # can be on multiple timestamps. In a batch, we have `batch_size`` number of samples.
-    num_samples = (step + 1) * cfg.policy.batch_size
+    num_samples = (step + 1) * cfg.training.batch_size
     avg_samples_per_ep = dataset.num_samples / dataset.num_episodes
     num_episodes = num_samples / avg_samples_per_ep
     num_epochs = num_samples / dataset.num_samples
@@ -213,11 +295,12 @@ def train(cfg: dict, out_dir=None, job_name=None):
         raise NotImplementedError()
     if job_name is None:
         raise NotImplementedError()
-    if cfg.online_steps > 0:
-        assert cfg.rollout_batch_size == 1, "rollout_batch_size > 1 not supported for online training steps"
 
     init_logging()
 
+    if cfg.training.online_steps > 0 and cfg.eval.batch_size > 1:
+        logging.warning("eval.batch_size > 1 not supported for online training steps")
+
     # Check device is available
     get_safe_torch_device(cfg.device, log=True)
 
@@ -229,10 +312,14 @@ def train(cfg: dict, out_dir=None, job_name=None):
     offline_dataset = make_dataset(cfg)
 
     logging.info("make_env")
-    env = make_env(cfg, num_parallel_envs=cfg.eval_episodes)
+    eval_env = make_env(cfg)
 
     logging.info("make_policy")
-    policy = make_policy(cfg, dataset_stats=offline_dataset.stats)
+    policy = make_policy(hydra_cfg=cfg, dataset_stats=offline_dataset.stats)
+
+    # Create optimizer and scheduler
+    # Temporary hack to move optimizer out of policy
+    optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
 
     num_learnable_params = sum(p.numel() for p in policy.parameters() if p.requires_grad)
     num_total_params = sum(p.numel() for p in policy.parameters())
@@ -242,70 +329,78 @@ def train(cfg: dict, out_dir=None, job_name=None):
 
     log_output_dir(out_dir)
     logging.info(f"{cfg.env.task=}")
-    logging.info(f"{cfg.offline_steps=} ({format_big_number(cfg.offline_steps)})")
-    logging.info(f"{cfg.online_steps=}")
+    logging.info(f"{cfg.training.offline_steps=} ({format_big_number(cfg.training.offline_steps)})")
+    logging.info(f"{cfg.training.online_steps=}")
     logging.info(f"{offline_dataset.num_samples=} ({format_big_number(offline_dataset.num_samples)})")
     logging.info(f"{offline_dataset.num_episodes=}")
     logging.info(f"{num_learnable_params=} ({format_big_number(num_learnable_params)})")
     logging.info(f"{num_total_params=} ({format_big_number(num_total_params)})")
 
     # Note: this helper will be used in offline and online training loops.
-    def _maybe_eval_and_maybe_save(step):
-        if step % cfg.eval_freq == 0:
+    def evaluate_and_checkpoint_if_needed(step):
+        if step % cfg.training.eval_freq == 0:
             logging.info(f"Eval policy at step {step}")
             eval_info = eval_policy(
-                env,
+                eval_env,
                 policy,
+                cfg.eval.n_episodes,
                 video_dir=Path(out_dir) / "eval",
                 max_episodes_rendered=4,
-                seed=cfg.seed,
+                start_seed=cfg.seed,
             )
             log_eval_info(logger, eval_info["aggregated"], step, cfg, offline_dataset, is_offline)
             if cfg.wandb.enable:
-                logger.log_video(eval_info["videos"][0], step, mode="eval")
+                logger.log_video(eval_info["video_paths"][0], step, mode="eval")
             logging.info("Resume training")
 
-        if cfg.save_model and step % cfg.save_freq == 0:
+        if cfg.training.save_model and step % cfg.training.save_freq == 0:
             logging.info(f"Checkpoint policy after step {step}")
-            logger.save_model(policy, identifier=step)
+            # Note: Save with step as the identifier, and format it to have at least 6 digits but more if
+            # needed (choose 6 as a minimum for consistency without being overkill).
+            logger.save_model(
+                policy,
+                identifier=str(step).zfill(
+                    max(6, len(str(cfg.training.offline_steps + cfg.training.online_steps)))
+                ),
+            )
             logging.info("Resume training")
 
     # create dataloader for offline training
     dataloader = torch.utils.data.DataLoader(
         offline_dataset,
         num_workers=4,
-        batch_size=cfg.policy.batch_size,
+        batch_size=cfg.training.batch_size,
         shuffle=True,
         pin_memory=cfg.device != "cpu",
         drop_last=False,
     )
     dl_iter = cycle(dataloader)
 
+    policy.train()
     step = 0  # number of policy update (forward + backward + optim)
     is_offline = True
-    for offline_step in range(cfg.offline_steps):
+    for offline_step in range(cfg.training.offline_steps):
         if offline_step == 0:
             logging.info("Start offline training on a fixed dataset")
-        policy.train()
         batch = next(dl_iter)
 
         for key in batch:
             batch[key] = batch[key].to(cfg.device, non_blocking=True)
 
-        train_info = policy.update(batch, step=step)
+        train_info = update_policy(policy, batch, optimizer, cfg.training.grad_clip_norm, lr_scheduler)
 
         # TODO(rcadene): is it ok if step_t=0 = 0 and not 1 as previously done?
-        if step % cfg.log_freq == 0:
+        if step % cfg.training.log_freq == 0:
             log_train_info(logger, train_info, step, cfg, offline_dataset, is_offline)
 
-        # Note: _maybe_eval_and_maybe_save happens **after** the `step`th training update has completed, so we pass in
-        # step + 1.
-        _maybe_eval_and_maybe_save(step + 1)
+        # Note: evaluate_and_checkpoint_if_needed happens **after** the `step`th training update has completed,
+        # so we pass in step + 1.
+        evaluate_and_checkpoint_if_needed(step + 1)
 
         step += 1
 
     # create an env dedicated to online episodes collection from policy rollout
-    rollout_env = make_env(cfg, num_parallel_envs=1)
+    online_training_env = make_env(cfg, n_envs=1)
 
     # create an empty online dataset similar to offline dataset
     online_dataset = deepcopy(offline_dataset)
@@ -321,7 +416,7 @@ def train(cfg: dict, out_dir=None, job_name=None):
     dataloader = torch.utils.data.DataLoader(
         concat_dataset,
         num_workers=4,
-        batch_size=cfg.policy.batch_size,
+        batch_size=cfg.training.batch_size,
         sampler=sampler,
         pin_memory=cfg.device != "cpu",
         drop_last=False,
@@ -330,46 +425,51 @@ def train(cfg: dict, out_dir=None, job_name=None):
 
     online_step = 0
     is_offline = False
-    for env_step in range(cfg.online_steps):
+    for env_step in range(cfg.training.online_steps):
         if env_step == 0:
             logging.info("Start online training by interacting with environment")
 
+        policy.eval()
         with torch.no_grad():
             eval_info = eval_policy(
-                rollout_env,
+                online_training_env,
                 policy,
+                n_episodes=1,
                 return_episode_data=True,
-                seed=cfg.seed,
+                start_seed=cfg.training.online_env_seed,
+                enable_progbar=True,
             )
 
-            add_episodes_inplace(
-                online_dataset,
-                concat_dataset,
-                sampler,
-                hf_dataset=eval_info["episodes"]["hf_dataset"],
-                episode_data_index=eval_info["episodes"]["episode_data_index"],
-                pc_online_samples=cfg.get("demo_schedule", 0.5),
-            )
+        add_episodes_inplace(
+            online_dataset,
+            concat_dataset,
+            sampler,
+            hf_dataset=eval_info["episodes"]["hf_dataset"],
+            episode_data_index=eval_info["episodes"]["episode_data_index"],
+            pc_online_samples=cfg.training.online_sampling_ratio,
+        )
 
-        for _ in range(cfg.policy.utd):
-            policy.train()
+        policy.train()
+        for _ in range(cfg.training.online_steps_between_rollouts):
             batch = next(dl_iter)
 
             for key in batch:
                 batch[key] = batch[key].to(cfg.device, non_blocking=True)
 
-            train_info = policy.update(batch, step)
+            train_info = update_policy(policy, batch, optimizer, cfg.training.grad_clip_norm, lr_scheduler)
 
-            if step % cfg.log_freq == 0:
+            if step % cfg.training.log_freq == 0:
                 log_train_info(logger, train_info, step, cfg, online_dataset, is_offline)
 
-            # Note: _maybe_eval_and_maybe_save happens **after** the `step`th training update has completed, so we pass
-            # in step + 1.
-            _maybe_eval_and_maybe_save(step + 1)
+            # Note: evaluate_and_checkpoint_if_needed happens **after** the `step`th training update has completed,
+            # so we pass in step + 1.
+            evaluate_and_checkpoint_if_needed(step + 1)
 
             step += 1
             online_step += 1
 
+    eval_env.close()
+    online_training_env.close()
     logging.info("End of training")
 
 

lerobot/scripts/visualize_dataset.py

@@ -1,116 +1,245 @@
+""" Visualize data of **all** frames of any episode of a dataset of type LeRobotDataset.
+
+Note: The last frame of the episode doesnt always correspond to a final state.
+That's because our datasets are composed of transition from state to state up to
+the antepenultimate state associated to the ultimate action to arrive in the final state.
+However, there might not be a transition from a final state to another state.
+
+Note: This script aims to visualize the data used to train the neural networks.
+~What you see is what you get~. When visualizing image modality, it is often expected to observe
+lossly compression artifacts since these images have been decoded from compressed mp4 videos to
+save disk space. The compression factor applied has been tuned to not affect success rate.
+
+Examples:
+
+- Visualize data stored on a local machine:
+```
+local$ python lerobot/scripts/visualize_dataset.py \
+    --repo-id lerobot/pusht \
+    --episode-index 0
+```
+
+- Visualize data stored on a distant machine with a local viewer:
+```
+distant$ python lerobot/scripts/visualize_dataset.py \
+    --repo-id lerobot/pusht \
+    --episode-index 0 \
+    --save 1 \
+    --output-dir path/to/directory
+
+local$ scp distant:path/to/directory/lerobot_pusht_episode_0.rrd .
+local$ rerun lerobot_pusht_episode_0.rrd
+```
+
+- Visualize data stored on a distant machine through streaming:
+(You need to forward the websocket port to the distant machine, with
+`ssh -L 9087:localhost:9087 username@remote-host`)
+```
+distant$ python lerobot/scripts/visualize_dataset.py \
+    --repo-id lerobot/pusht \
+    --episode-index 0 \
+    --mode distant \
+    --ws-port 9087
+
+local$ rerun ws://localhost:9087
+```
+
+"""
+
+import argparse
 import logging
-import threading
+import time
 from pathlib import Path
 
-import einops
-import hydra
-import imageio
+import rerun as rr
 import torch
+import tqdm
 
-from lerobot.common.datasets.factory import make_dataset
-from lerobot.common.logger import log_output_dir
-from lerobot.common.utils.utils import init_logging
-
-NUM_EPISODES_TO_RENDER = 50
-MAX_NUM_STEPS = 1000
-FIRST_FRAME = 0
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
 
-@hydra.main(version_base="1.2", config_name="default", config_path="../configs")
-def visualize_dataset_cli(cfg: dict):
-    visualize_dataset(cfg, out_dir=hydra.core.hydra_config.HydraConfig.get().runtime.output_dir)
+class EpisodeSampler(torch.utils.data.Sampler):
+    def __init__(self, dataset, episode_index):
+        from_idx = dataset.episode_data_index["from"][episode_index].item()
+        to_idx = dataset.episode_data_index["to"][episode_index].item()
+        self.frame_ids = range(from_idx, to_idx)
+
+    def __iter__(self):
+        return iter(self.frame_ids)
+
+    def __len__(self):
+        return len(self.frame_ids)
 
 
-def cat_and_write_video(video_path, frames, fps):
-    frames = torch.cat(frames)
-
-    # Expects images in [0, 1].
-    frame = frames[0]
-    if frame.ndim == 4:
-        raise NotImplementedError("We currently dont support multiple timestamps.")
-    c, h, w = frame.shape
-    assert c < h and c < w, f"expect channel first images, but instead {frame.shape}"
-
-    # sanity check that images are float32 in range [0,1]
-    assert frame.dtype == torch.float32, f"expect torch.float32, but instead {frame.dtype=}"
-    assert frame.max() <= 1, f"expect pixels lower than 1, but instead {frame.max()=}"
-    assert frame.min() >= 0, f"expect pixels greater than 1, but instead {frame.min()=}"
-
-    # convert to channel last uint8 [0, 255]
-    frames = einops.rearrange(frames, "b c h w -> b h w c")
-    frames = (frames * 255).type(torch.uint8)
-    imageio.mimsave(video_path, frames.numpy(), fps=fps)
+def to_hwc_uint8_numpy(chw_float32_torch):
+    assert chw_float32_torch.dtype == torch.float32
+    assert chw_float32_torch.ndim == 3
+    c, h, w = chw_float32_torch.shape
+    assert c < h and c < w, f"expect channel first images, but instead {chw_float32_torch.shape}"
+    hwc_uint8_numpy = (chw_float32_torch * 255).type(torch.uint8).permute(1, 2, 0).numpy()
+    return hwc_uint8_numpy
 
 
-def visualize_dataset(cfg: dict, out_dir=None):
-    if out_dir is None:
-        raise NotImplementedError()
+def visualize_dataset(
+    repo_id: str,
+    episode_index: int,
+    batch_size: int = 32,
+    num_workers: int = 0,
+    mode: str = "local",
+    web_port: int = 9090,
+    ws_port: int = 9087,
+    save: bool = False,
+    output_dir: Path | None = None,
+) -> Path | None:
+    if save:
+        assert (
+            output_dir is not None
+        ), "Set an output directory where to write .rrd files with `--output-dir path/to/directory`."
 
-    init_logging()
-    log_output_dir(out_dir)
-
-    logging.info("make_dataset")
-    dataset = make_dataset(cfg)
-
-    logging.info("Start rendering episodes from offline buffer")
-    video_paths = render_dataset(dataset, out_dir, MAX_NUM_STEPS * NUM_EPISODES_TO_RENDER)
-    for video_path in video_paths:
-        logging.info(video_path)
-    return video_paths
-
-
-def render_dataset(dataset, out_dir, max_num_episodes):
-    out_dir = Path(out_dir)
-    video_paths = []
-    threads = []
+    logging.info("Loading dataset")
+    dataset = LeRobotDataset(repo_id)
 
+    logging.info("Loading dataloader")
+    episode_sampler = EpisodeSampler(dataset, episode_index)
     dataloader = torch.utils.data.DataLoader(
         dataset,
-        num_workers=4,
-        batch_size=1,
-        shuffle=False,
+        num_workers=num_workers,
+        batch_size=batch_size,
+        sampler=episode_sampler,
     )
-    dl_iter = iter(dataloader)
 
-    for ep_id in range(min(max_num_episodes, dataset.num_episodes)):
-        logging.info(f"Rendering episode {ep_id}")
+    logging.info("Starting Rerun")
 
-        frames = {}
-        end_of_episode = False
-        while not end_of_episode:
-            item = next(dl_iter)
+    if mode not in ["local", "distant"]:
+        raise ValueError(mode)
 
-            for im_key in dataset.image_keys:
-                # when first frame of episode, initialize frames dict
-                if im_key not in frames:
-                    frames[im_key] = []
-                # add current frame to list of frames to render
-                frames[im_key].append(item[im_key])
+    spawn_local_viewer = mode == "local" and not save
+    rr.init(f"{repo_id}/episode_{episode_index}", spawn=spawn_local_viewer)
+    if mode == "distant":
+        rr.serve(open_browser=False, web_port=web_port, ws_port=ws_port)
 
-            end_of_episode = item["index"].item() == dataset.episode_data_index["to"][ep_id] - 1
+    logging.info("Logging to Rerun")
 
-        out_dir.mkdir(parents=True, exist_ok=True)
-        for im_key in dataset.image_keys:
-            if len(dataset.image_keys) > 1:
-                im_name = im_key.replace("observation.images.", "")
-                video_path = out_dir / f"episode_{ep_id}_{im_name}.mp4"
-            else:
-                video_path = out_dir / f"episode_{ep_id}.mp4"
-            video_paths.append(video_path)
+    if num_workers > 0:
+        # TODO(rcadene): fix data workers hanging when `rr.init` is called
+        logging.warning("If data loader is hanging, try `--num-workers 0`.")
 
-            thread = threading.Thread(
-                target=cat_and_write_video,
-                args=(str(video_path), frames[im_key], dataset.fps),
-            )
-            thread.start()
-            threads.append(thread)
+    for batch in tqdm.tqdm(dataloader, total=len(dataloader)):
+        # iterate over the batch
+        for i in range(len(batch["index"])):
+            rr.set_time_sequence("frame_index", batch["frame_index"][i].item())
+            rr.set_time_seconds("timestamp", batch["timestamp"][i].item())
 
-    for thread in threads:
-        thread.join()
+            # display each camera image
+            for key in dataset.camera_keys:
+                # TODO(rcadene): add `.compress()`? is it lossless?
+                rr.log(key, rr.Image(to_hwc_uint8_numpy(batch[key][i])))
 
-    logging.info("End of visualize_dataset")
-    return video_paths
+            # display each dimension of action space (e.g. actuators command)
+            if "action" in batch:
+                for dim_idx, val in enumerate(batch["action"][i]):
+                    rr.log(f"action/{dim_idx}", rr.Scalar(val.item()))
+
+            # display each dimension of observed state space (e.g. agent position in joint space)
+            if "observation.state" in batch:
+                for dim_idx, val in enumerate(batch["observation.state"][i]):
+                    rr.log(f"state/{dim_idx}", rr.Scalar(val.item()))
+
+            if "next.done" in batch:
+                rr.log("next.done", rr.Scalar(batch["next.done"][i].item()))
+
+            if "next.reward" in batch:
+                rr.log("next.reward", rr.Scalar(batch["next.reward"][i].item()))
+
+            if "next.success" in batch:
+                rr.log("next.success", rr.Scalar(batch["next.success"][i].item()))
+
+    if mode == "local" and save:
+        # save .rrd locally
+        output_dir = Path(output_dir)
+        output_dir.mkdir(parents=True, exist_ok=True)
+        repo_id_str = repo_id.replace("/", "_")
+        rrd_path = output_dir / f"{repo_id_str}_episode_{episode_index}.rrd"
+        rr.save(rrd_path)
+        return rrd_path
+
+    elif mode == "distant":
+        # stop the process from exiting since it is serving the websocket connection
+        try:
+            while True:
+                time.sleep(1)
+        except KeyboardInterrupt:
+            print("Ctrl-C received. Exiting.")
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--repo-id",
+        type=str,
+        required=True,
+        help="Name of hugging face repositery containing a LeRobotDataset dataset (e.g. `lerobot/pusht`).",
+    )
+    parser.add_argument(
+        "--episode-index",
+        type=int,
+        required=True,
+        help="Episode to visualize.",
+    )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=32,
+        help="Batch size loaded by DataLoader.",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=0,
+        help="Number of processes of Dataloader for loading the data.",
+    )
+    parser.add_argument(
+        "--mode",
+        type=str,
+        default="local",
+        help=(
+            "Mode of viewing between 'local' or 'distant'. "
+            "'local' requires data to be on a local machine. It spawns a viewer to visualize the data locally. "
+            "'distant' creates a server on the distant machine where the data is stored. Visualize the data by connecting to the server with `rerun ws://localhost:PORT` on the local machine."
+        ),
+    )
+    parser.add_argument(
+        "--web-port",
+        type=int,
+        default=9090,
+        help="Web port for rerun.io when `--mode distant` is set.",
+    )
+    parser.add_argument(
+        "--ws-port",
+        type=int,
+        default=9087,
+        help="Web socket port for rerun.io when `--mode distant` is set.",
+    )
+    parser.add_argument(
+        "--save",
+        type=int,
+        default=0,
+        help=(
+            "Save a .rrd file in the directory provided by `--output-dir`. "
+            "It also deactivates the spawning of a viewer. ",
+            "Visualize the data by running `rerun path/to/file.rrd` on your local machine.",
+        ),
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        help="Directory path to write a .rrd file when `--save 1` is set.",
+    )
+
+    args = parser.parse_args()
+    visualize_dataset(**vars(args))
 
 
 if __name__ == "__main__":
-    visualize_dataset_cli()
+    main()

pyproject.toml

@@ -4,9 +4,9 @@ version = "0.1.0"
 description = "🤗 LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch"
 authors = [
     "Rémi Cadène <re.cadene@gmail.com>",
+    "Simon Alibert <alibert.sim@gmail.com>",
     "Alexander Soare <alexander.soare159@gmail.com>",
     "Quentin Gallouédec <quentin.gallouedec@ec-lyon.fr>",
-    "Simon Alibert <alibert.sim@gmail.com>",
     "Adil Zouitine <adilzouitinegm@gmail.com>",
     "Thomas Wolf <thomaswolfcontact@gmail.com>",
 ]
@@ -27,7 +27,7 @@ packages = [{include = "lerobot"}]
 
 
 [tool.poetry.dependencies]
-python = "^3.10"
+python = ">=3.10,<3.13"
 termcolor = "^2.4.0"
 omegaconf = "^2.3.0"
 wandb = "^0.16.3"
@@ -47,16 +47,18 @@ huggingface-hub = "^0.21.4"
 robomimic = "0.2.0"
 gymnasium = "^0.29.1"
 cmake = "^3.29.0.1"
-gym-pusht = { git = "git@github.com:huggingface/gym-pusht.git", optional = true}
-gym-xarm = { git = "git@github.com:huggingface/gym-xarm.git", optional = true}
-gym-aloha = { git = "git@github.com:huggingface/gym-aloha.git", optional = true}
+gym-pusht = { version = "^0.1.1", optional = true}
+gym-xarm = { version = "^0.1.0", optional = true}
+gym-aloha = { version = "^0.1.0", optional = true}
 pre-commit = {version = "^3.7.0", optional = true}
 debugpy = {version = "^1.8.1", optional = true}
 pytest = {version = "^8.1.0", optional = true}
 pytest-cov = {version = "^5.0.0", optional = true}
 datasets = "^2.19.0"
 imagecodecs = { version = "^2024.1.1", optional = true }
-torchaudio = "^2.3.0"
+pyav = "^12.0.5"
+moviepy = "^1.0.3"
+rerun-sdk = "^0.15.1"
 
 
 [tool.poetry.extras]

tests/data/lerobot/aloha_sim_insertion_human/meta_data/info.json

@@ -1,3 +1,4 @@
 {
-    "fps": 50
+    "fps": 50,
+    "video": 1
 }

tests/data/lerobot/aloha_sim_insertion_human/train/dataset_info.json

@@ -3,7 +3,7 @@
   "description": "",
   "features": {
     "observation.images.top": {
-      "_type": "Image"
+      "_type": "VideoFrame"
     },
     "observation.state": {
       "feature": {

tests/data/lerobot/aloha_sim_insertion_human/train/state.json

@@ -4,10 +4,10 @@
       "filename": "data-00000-of-00001.arrow"
     }
   ],
-  "_fingerprint": "22eeca7a3f4725ee",
+  "_fingerprint": "eb913a2b1a68aa74",
   "_format_columns": null,
   "_format_kwargs": {},
-  "_format_type": "torch",
+  "_format_type": null,
   "_output_all_columns": false,
   "_split": null
 }

tests/data/lerobot/aloha_sim_insertion_scripted/meta_data/info.json

@@ -1,3 +1,4 @@
 {
-    "fps": 50
+    "fps": 50,
+    "video": 1
 }

tests/data/lerobot/aloha_sim_insertion_scripted/train/dataset_info.json

@@ -3,7 +3,7 @@
   "description": "",
   "features": {
     "observation.images.top": {
-      "_type": "Image"
+      "_type": "VideoFrame"
     },
     "observation.state": {
       "feature": {