WIP faster act

Co-authored-by: Alexander Soare <alexander.soare159@gmail.com>

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

@@ -10,7 +10,6 @@ on:
 
 env:
   PYTHON_VERSION: "3.10"
-#   CI_SLACK_CHANNEL: ${{ secrets.CI_DOCKER_CHANNEL }}
 
 jobs:
   latest-cpu:
@@ -35,6 +34,8 @@ jobs:
 
       - name: Check out code
         uses: actions/checkout@v4
+        with:
+          lfs: true
 
       - name: Login to DockerHub
         uses: docker/login-action@v3
@@ -51,34 +52,50 @@ jobs:
           tags: huggingface/lerobot-cpu
           build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
 
-    #   - name: Post to a Slack channel
-    #     id: slack
-    #     #uses: slackapi/slack-github-action@v1.25.0
-    #     uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
-    #     with:
-    #       # Slack channel id, channel name, or user id to post message.
-    #       # See also: https://api.slack.com/methods/chat.postMessage#channels
-    #       channel-id: ${{ env.CI_SLACK_CHANNEL }}
-    #       # For posting a rich message using Block Kit
-    #       payload: |
-    #         {
-    #           "text": "lerobot-cpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}",
-    #           "blocks": [
-    #             {
-    #               "type": "section",
-    #               "text": {
-    #                 "type": "mrkdwn",
-    #                 "text": "lerobot-cpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}"
-    #               }
-    #             }
-    #           ]
-    #         }
-    #     env:
-    #       SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
 
   latest-cuda:
     name: GPU
     runs-on: ubuntu-latest
+    steps:
+      - name: Cleanup disk
+        run: |
+          sudo df -h
+          # sudo ls -l /usr/local/lib/
+          # sudo ls -l /usr/share/
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /usr/share/dotnet
+          sudo du -sh /usr/local/lib/
+          sudo du -sh /usr/share/
+          sudo df -h
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Check out code
+        uses: actions/checkout@v4
+        with:
+          lfs: true
+
+      - name: Login to DockerHub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_PASSWORD }}
+
+      - name: Build and Push GPU
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: ./docker/lerobot-gpu/Dockerfile
+          push: true
+          tags: huggingface/lerobot-gpu
+          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
+
+
+  latest-cuda-dev:
+    name: GPU Dev
+    runs-on: ubuntu-latest
     steps:
       - name: Cleanup disk
         run: |
@@ -104,36 +121,11 @@ jobs:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_PASSWORD }}
 
-      - name: Build and Push GPU
+      - name: Build and Push GPU dev
         uses: docker/build-push-action@v5
         with:
           context: .
-          file: ./docker/lerobot-gpu/Dockerfile
+          file: ./docker/lerobot-gpu-dev/Dockerfile
           push: true
-          tags: huggingface/lerobot-gpu
+          tags: huggingface/lerobot-gpu:dev
           build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
-
-      # - name: Post to a Slack channel
-      #   id: slack
-      #   #uses: slackapi/slack-github-action@v1.25.0
-      #   uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
-      #   with:
-      #     # Slack channel id, channel name, or user id to post message.
-      #     # See also: https://api.slack.com/methods/chat.postMessage#channels
-      #     channel-id: ${{ env.CI_SLACK_CHANNEL }}
-      #     # For posting a rich message using Block Kit
-      #     payload: |
-      #       {
-      #         "text": "lerobot-gpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}",
-      #         "blocks": [
-      #           {
-      #             "type": "section",
-      #             "text": {
-      #               "type": "mrkdwn",
-      #               "text": "lerobot-gpu Docker Image build result: ${{ job.status }}\n${{ github.event.pull_request.html_url || github.event.head_commit.url }}"
-      #             }
-      #           }
-      #         ]
-      #       }
-      #   env:
-      #     SLACK_BOT_TOKEN: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}

.github/workflows/nightly-tests.yml

@@ -70,6 +70,8 @@ jobs:
       #     files: ./coverage.xml
       #     verbose: true
       - name: Tests end-to-end
+        env:
+          DEVICE: cuda
         run: make test-end-to-end
 
     #   - name: Generate Report

.gitignore

@@ -2,12 +2,17 @@
 logs
 tmp
 wandb
+
+# Data
 data
 outputs
-.vscode
-rl
+
+# Apple
 .DS_Store
 
+# VS Code
+.vscode
+
 # HPC
 nautilus/*.yaml
 *.key
@@ -90,6 +95,7 @@ instance/
 docs/_build/
 
 # PyBuilder
+.pybuilder/
 target/
 
 # Jupyter Notebook
@@ -102,13 +108,6 @@ ipython_config.py
 # pyenv
 .python-version
 
-# pipenv
-#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
-#   However, in case of collaboration, if having platform-specific dependencies or dependencies
-#   having no cross-platform support, pipenv may install dependencies that don't work, or not
-#   install all needed dependencies.
-#Pipfile.lock
-
 # PEP 582; used by e.g. github.com/David-OConnor/pyflow
 __pypackages__/
 
@@ -119,6 +118,14 @@ celerybeat.pid
 # SageMath parsed files
 *.sage.py
 
+# Environments
+.env
+.venv
+venv/
+ENV/
+env.bak/
+venv.bak/
+
 # Spyder project settings
 .spyderproject
 .spyproject
@@ -136,3 +143,9 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/

Makefile

@@ -10,6 +10,7 @@ endif
 
 export PATH := $(dir $(PYTHON_PATH)):$(PATH)
 
+DEVICE ?= cpu
 
 build-cpu:
 	docker build -t lerobot:latest -f docker/lerobot-cpu/Dockerfile .
@@ -18,25 +19,29 @@ build-gpu:
 	docker build -t lerobot:latest -f docker/lerobot-gpu/Dockerfile .
 
 test-end-to-end:
-	${MAKE} test-act-ete-train
-	${MAKE} test-act-ete-eval
-	${MAKE} test-diffusion-ete-train
-	${MAKE} test-diffusion-ete-eval
-	${MAKE} test-tdmpc-ete-train
-	${MAKE} test-tdmpc-ete-eval
-	${MAKE} test-default-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-train-amp
+	${MAKE} DEVICE=$(DEVICE) test-act-ete-eval-amp
+	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train
+	${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-default-ete-eval
+	${MAKE} DEVICE=$(DEVICE) test-act-pusht-tutorial
 
 test-act-ete-train:
 	python lerobot/scripts/train.py \
 		policy=act \
+		policy.dim_model=64 \
 		env=aloha \
 		wandb.enable=False \
 		training.offline_steps=2 \
 		training.online_steps=0 \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
-		device=cpu \
-		training.save_model=true \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
 		training.save_freq=2 \
 		policy.n_action_steps=20 \
 		policy.chunk_size=20 \
@@ -45,35 +50,67 @@ test-act-ete-train:
 
 test-act-ete-eval:
 	python lerobot/scripts/eval.py \
-		-p tests/outputs/act/checkpoints/000002 \
+		-p tests/outputs/act/checkpoints/000002/pretrained_model \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
+		device=$(DEVICE) \
+
+test-act-ete-train-amp:
+	python lerobot/scripts/train.py \
+		policy=act \
+		policy.dim_model=64 \
+		env=aloha \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		training.online_steps=0 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
+		training.save_freq=2 \
+		policy.n_action_steps=20 \
+		policy.chunk_size=20 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/act_amp/ \
+		use_amp=true
+
+test-act-ete-eval-amp:
+	python lerobot/scripts/eval.py \
+		-p tests/outputs/act_amp/checkpoints/000002/pretrained_model \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=8 \
+		device=$(DEVICE) \
+		use_amp=true
 
 test-diffusion-ete-train:
 	python lerobot/scripts/train.py \
 		policy=diffusion \
+		policy.down_dims=\[64,128,256\] \
+		policy.diffusion_step_embed_dim=32 \
+		policy.num_inference_steps=10 \
 		env=pusht \
 		wandb.enable=False \
 		training.offline_steps=2 \
 		training.online_steps=0 \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
-		device=cpu \
-		training.save_model=true \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
 		training.save_freq=2 \
 		training.batch_size=2 \
 		hydra.run.dir=tests/outputs/diffusion/
 
 test-diffusion-ete-eval:
 	python lerobot/scripts/eval.py \
-		-p tests/outputs/diffusion/checkpoints/000002 \
+		-p tests/outputs/diffusion/checkpoints/000002/pretrained_model \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
+		device=$(DEVICE) \
 
+# TODO(alexander-soare): Restore online_steps to 2 when it is reinstated.
 test-tdmpc-ete-train:
 	python lerobot/scripts/train.py \
 		policy=tdmpc \
@@ -82,24 +119,23 @@ test-tdmpc-ete-train:
 		dataset_repo_id=lerobot/xarm_lift_medium \
 		wandb.enable=False \
 		training.offline_steps=2 \
-		training.online_steps=2 \
+		training.online_steps=0 \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=2 \
-		device=cpu \
-		training.save_model=true \
+		device=$(DEVICE) \
+		training.save_checkpoint=true \
 		training.save_freq=2 \
 		training.batch_size=2 \
 		hydra.run.dir=tests/outputs/tdmpc/
 
 test-tdmpc-ete-eval:
 	python lerobot/scripts/eval.py \
-		-p tests/outputs/tdmpc/checkpoints/000002 \
+		-p tests/outputs/tdmpc/checkpoints/000002/pretrained_model \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
-
+		device=$(DEVICE) \
 
 test-default-ete-eval:
 	python lerobot/scripts/eval.py \
@@ -107,4 +143,21 @@ test-default-ete-eval:
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
+		device=$(DEVICE) \
+
+test-act-pusht-tutorial:
+	cp examples/advanced/1_train_act_pusht/act_pusht.yaml lerobot/configs/policy/created_by_Makefile.yaml
+	python lerobot/scripts/train.py \
+		policy=created_by_Makefile.yaml \
+		env=pusht \
+		wandb.enable=False \
+		training.offline_steps=2 \
+		eval.n_episodes=1 \
+		eval.batch_size=1 \
+		env.episode_length=2 \
+		device=$(DEVICE) \
+		training.save_model=true \
+		training.save_freq=2 \
+		training.batch_size=2 \
+		hydra.run.dir=tests/outputs/act_pusht/
+	rm lerobot/configs/policy/created_by_Makefile.yaml

README.md

@@ -77,6 +77,10 @@ Install 🤗 LeRobot:
 pip install .
 ```
 
+> **NOTE:** Depending on your platform, If you encounter any build errors during this step
+you may need to install `cmake` and `build-essential` for building some of our dependencies.
+On linux: `sudo apt-get install cmake build-essential`
+
 For simulations, 🤗 LeRobot comes with gymnasium environments that can be installed as extras:
 - [aloha](https://github.com/huggingface/gym-aloha)
 - [xarm](https://github.com/huggingface/gym-xarm)
@@ -99,6 +103,7 @@ wandb login
 ```
 .
 ├── examples             # contains demonstration examples, start here to learn about LeRobot
+|   └── advanced         # contains even more examples for those who have mastered the basics
 ├── lerobot
 |   ├── configs          # contains hydra yaml files with all options that you can override in the command line
 |   |   ├── default.yaml   # selected by default, it loads pusht environment and diffusion policy
@@ -149,18 +154,19 @@ python lerobot/scripts/eval.py \
 ```
 
 Note: After training your own policy, you can re-evaluate the checkpoints with:
+
 ```bash
-python lerobot/scripts/eval.py \
-    -p PATH/TO/TRAIN/OUTPUT/FOLDER
+python lerobot/scripts/eval.py -p {OUTPUT_DIR}/checkpoints/last/pretrained_model
 ```
 
 See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-Check out [example 3](./examples/3_train_policy.py) that illustrates how to start training a model.
+Check out [example 3](./examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
 
 In general, you can use our training script to easily train any policy. Here is an example of training the ACT policy on trajectories collected by humans on the Aloha simulation environment for the insertion task:
+
 ```bash
 python lerobot/scripts/train.py \
     policy=act \
@@ -174,6 +180,19 @@ The experiment directory is automatically generated and will show up in yellow i
     hydra.run.dir=your/new/experiment/dir
 ```
 
+In the experiment directory there will be a folder called `checkpoints` which will have the following structure:
+
+```bash
+checkpoints
+├── 000250  # checkpoint_dir for training step 250
+│   ├── pretrained_model  # Hugging Face pretrained model dir
+│   │   ├── config.json  # Hugging Face pretrained model config
+│   │   ├── config.yaml  # consolidated Hydra config
+│   │   ├── model.safetensors  # model weights
+│   │   └── README.md  # Hugging Face model card
+│   └── training_state.pth  # optimizer/scheduler/rng state and training step
+```
+
 To use wandb for logging training and evaluation curves, make sure you've run `wandb login` as a one-time setup step. Then, when running the training command above, enable WandB in the configuration by adding:
 
 ```bash
@@ -184,7 +203,19 @@ A link to the wandb logs for the run will also show up in yellow in your termina
 
 ![](media/wandb.png)
 
-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.
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. You may use `eval.n_episodes=500` to evaluate on more episodes than the default. Or, after training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
+
+#### Reproduce state-of-the-art (SOTA)
+
+We have organized our configuration files (found under [`lerobot/configs`](./lerobot/configs)) such that they reproduce SOTA results from a given model variant in their respective original works. Simply running:
+
+```bash
+python lerobot/scripts/train.py policy=diffusion env=pusht
+```
+
+reproduces SOTA results for Diffusion Policy on the PushT task.
+
+Pretrained policies, along with reproduction details, can be found under the "Models" section of https://huggingface.co/lerobot.
 
 ## Contribute
 
@@ -215,14 +246,14 @@ If your dataset format is not supported, implement your own in `lerobot/common/d
 
 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)).
 
-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:
+You first need to find the checkpoint folder located inside your experiment directory (e.g. `outputs/train/2024-05-05/20-21-12_aloha_act_default/checkpoints/002500`). Within that there is a `pretrained_model` directory which should contain:
 - `config.json`: A serialized version of the policy configuration (following the policy's dataclass config).
 - `model.safetensors`: A set of `torch.nn.Module` parameters, saved in [Hugging Face Safetensors](https://huggingface.co/docs/safetensors/index) format.
 - `config.yaml`: A consolidated Hydra training configuration containing the policy, environment, and dataset configs. The policy configuration should match `config.json` exactly. The environment config is useful for anyone who wants to evaluate your policy. The dataset config just serves as a paper trail for reproducibility.
 
 To upload these to the hub, run the following:
 ```bash
-huggingface-cli upload ${hf_user}/${repo_name} path/to/checkpoint/dir
+huggingface-cli upload ${hf_user}/${repo_name} path/to/pretrained_model
 ```
 
 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.

docker/lerobot-gpu-dev/Dockerfile

@@ -0,0 +1,40 @@
+FROM nvidia/cuda:12.4.1-base-ubuntu22.04
+
+# Configure image
+ARG PYTHON_VERSION=3.10
+ARG DEBIAN_FRONTEND=noninteractive
+
+# Install apt dependencies
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake \
+    git git-lfs openssh-client \
+    nano vim less util-linux \
+    htop atop nvtop \
+    sed gawk grep curl wget \
+    tcpdump sysstat screen tmux \
+    libglib2.0-0 libgl1-mesa-glx libegl1-mesa ffmpeg \
+    python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Install gh cli tool
+RUN (type -p wget >/dev/null || (apt update && apt-get install wget -y)) \
+    && mkdir -p -m 755 /etc/apt/keyrings \
+    && wget -qO- https://cli.github.com/packages/githubcli-archive-keyring.gpg | tee /etc/apt/keyrings/githubcli-archive-keyring.gpg > /dev/null \
+    && chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg \
+    && echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" | tee /etc/apt/sources.list.d/github-cli.list > /dev/null \
+    && apt update \
+    && apt install gh -y \
+    && apt clean && rm -rf /var/lib/apt/lists/*
+
+# Setup `python`
+RUN ln -s /usr/bin/python3 /usr/bin/python
+
+# Install poetry
+RUN curl -sSL https://install.python-poetry.org | python -
+ENV PATH="/root/.local/bin:$PATH"
+RUN echo 'if [ "$HOME" != "/root" ]; then ln -sf /root/.local/bin/poetry $HOME/.local/bin/poetry; fi' >> /root/.bashrc
+RUN poetry config virtualenvs.create false
+RUN poetry config virtualenvs.in-project true
+
+# Set EGL as the rendering backend for MuJoCo
+ENV MUJOCO_GL="egl"

docker/lerobot-gpu/Dockerfile

@@ -4,18 +4,15 @@ FROM nvidia/cuda:12.4.1-base-ubuntu22.04
 ARG PYTHON_VERSION=3.10
 ARG DEBIAN_FRONTEND=noninteractive
 
+
 # Install apt dependencies
 RUN apt-get update && apt-get install -y --no-install-recommends \
     build-essential cmake \
-    git git-lfs openssh-client \
-    nano vim ffmpeg \
-    htop atop nvtop \
-    sed gawk grep curl wget \
-    tcpdump sysstat screen \
     libglib2.0-0 libgl1-mesa-glx libegl1-mesa \
     python${PYTHON_VERSION} python${PYTHON_VERSION}-venv \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
+
 # Create virtual environment
 RUN ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python
 RUN python -m venv /opt/venv
@@ -23,8 +20,7 @@ ENV PATH="/opt/venv/bin:$PATH"
 RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 
 # Install LeRobot
-RUN git lfs install
-RUN git clone https://github.com/huggingface/lerobot.git
+COPY . /lerobot
 WORKDIR /lerobot
 RUN pip install --upgrade --no-cache-dir pip
 RUN pip install --no-cache-dir ".[test, aloha, xarm, pusht]"

examples/4_train_policy_with_script.md

@@ -0,0 +1,183 @@
+This tutorial will explain the training script, how to use it, and particularly the use of Hydra to configure everything needed for the training run.
+
+## The training script
+
+LeRobot offers a training script at [`lerobot/scripts/train.py`](../../lerobot/scripts/train.py). At a high level it does the following:
+
+- Loads a Hydra configuration file for the following steps (more on Hydra in a moment).
+- Makes a simulation environment.
+- Makes a dataset corresponding to that simulation environment.
+- Makes a policy.
+- Runs a standard training loop with forward pass, backward pass, optimization step, and occasional logging, evaluation (of the policy on the environment), and checkpointing.
+
+## Basics of how we use Hydra
+
+Explaining the ins and outs of [Hydra](https://hydra.cc/docs/intro/) is beyond the scope of this document, but here we'll share the main points you need to know.
+
+First, `lerobot/configs` has a directory structure like this:
+
+```
+.
+├── default.yaml
+├── env
+│   ├── aloha.yaml
+│   ├── pusht.yaml
+│   └── xarm.yaml
+└── policy
+    ├── act.yaml
+    ├── diffusion.yaml
+    └── tdmpc.yaml
+```
+
+**_For brevity, in the rest of this document we'll drop the leading `lerobot/configs` path. So `default.yaml` really refers to `lerobot/configs/default.yaml`._**
+
+When you run the training script with
+
+```python
+python lerobot/scripts/train.py
+```
+
+Hydra is set up to read `default.yaml` (via the `@hydra.main` decorator). If you take a look at the `@hydra.main`'s arguments you will see `config_path="../configs", config_name="default"`. At the top of `default.yaml`, is a `defaults` section which looks likes this:
+
+```yaml
+defaults:
+  - _self_
+  - env: pusht
+  - policy: diffusion
+```
+
+This logic tells Hydra to incorporate configuration parameters from `env/pusht.yaml` and `policy/diffusion.yaml`. _Note: Be aware of the order as any configuration parameters with the same name will be overidden. Thus, `default.yaml` is overriden by `env/pusht.yaml`  which is overidden by `policy/diffusion.yaml`_.
+
+Then, `default.yaml` also contains common configuration parameters such as `device: cuda` or `use_amp: false` (for enabling fp16 training). Some other parameters are set to `???` which indicates that they are expected to be set in additional yaml files. For instance, `training.offline_steps: ???` in `default.yaml` is set to `200000` in `diffusion.yaml`.
+
+Thanks to this `defaults` section in `default.yaml`, if you want to train Diffusion Policy with PushT, you really only need to run:
+
+```bash
+python lerobot/scripts/train.py
+```
+
+However, you can be more explicit and launch the exact same Diffusion Policy training on PushT with:
+
+```bash
+python lerobot/scripts/train.py policy=diffusion env=pusht
+```
+
+This way of overriding defaults via the CLI is especially useful when you want to change the policy and/or environment. For instance, you can train ACT on the default Aloha environment with:
+
+```bash
+python lerobot/scripts/train.py policy=act env=aloha
+```
+
+There are two things to note here:
+- Config overrides are passed as `param_name=param_value`.
+- Here we have overridden the defaults section. `policy=act` tells Hydra to use `policy/act.yaml`, and `env=aloha` tells Hydra to use `env/aloha.yaml`.
+
+_As an aside: we've set up all of our configurations so that they reproduce state-of-the-art results from papers in the literature._
+
+## Overriding configuration parameters in the CLI
+
+Now let's say that we want to train on a different task in the Aloha environment. If you look in `env/aloha.yaml` you will see something like:
+
+```yaml
+# lerobot/configs/env/aloha.yaml
+env:
+  task: AlohaInsertion-v0
+```
+
+And if you look in `policy/act.yaml` you will see something like:
+
+```yaml
+# lerobot/configs/policy/act.yaml
+dataset_repo_id: lerobot/aloha_sim_insertion_human
+```
+
+But our Aloha environment actually supports a cube transfer task as well. To train for this task, you could manually modify the two yaml configuration files respectively.
+
+First, we'd need to switch to using the cube transfer task for the ALOHA environment.
+
+```diff
+# lerobot/configs/env/aloha.yaml
+env:
+-  task: AlohaInsertion-v0
++  task: AlohaTransferCube-v0
+```
+
+Then, we'd also need to switch to using the cube transfer dataset.
+
+```diff
+# lerobot/configs/policy/act.yaml
+-dataset_repo_id: lerobot/aloha_sim_insertion_human
++dataset_repo_id: lerobot/aloha_sim_transfer_cube_human
+```
+
+Then, you'd be able to run:
+
+```bash
+python lerobot/scripts/train.py policy=act env=aloha
+```
+
+and you'd be training and evaluating on the cube transfer task.
+
+An alternative approach to editing the yaml configuration files, would be to override the defaults via the command line:
+
+```bash
+python lerobot/scripts/train.py \
+    policy=act \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    env=aloha \
+    env.task=AlohaTransferCube-v0
+```
+
+There's something new here. Notice the `.` delimiter used to traverse the configuration hierarchy. _But be aware that the `defaults` section is an exception. As you saw above, we didn't need to write `defaults.policy=act` in the CLI. `policy=act` was enough._
+
+Putting all that knowledge together, here's the command that was used to train https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human.
+
+```bash
+python lerobot/scripts/train.py \
+    hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human \
+    device=cuda
+    env=aloha \
+    env.task=AlohaTransferCube-v0 \
+    dataset_repo_id=lerobot/aloha_sim_transfer_cube_human \
+    policy=act \
+    training.eval_freq=10000 \
+    training.log_freq=250 \
+    training.offline_steps=100000 \
+    training.save_model=true \
+    training.save_freq=25000 \
+    eval.n_episodes=50 \
+    eval.batch_size=50 \
+    wandb.enable=false \
+```
+
+There's one new thing here: `hydra.run.dir=outputs/train/act_aloha_sim_transfer_cube_human`, which specifies where to save the training output.
+
+## Using a configuration file not in `lerobot/configs`
+
+Above we discusses the our training script is set up such that Hydra looks for `default.yaml` in `lerobot/configs`. But, if you have a configuration file elsewhere in your filesystem you may use:
+
+```bash
+python lerobot/scripts/train.py --config-dir PARENT/PATH --config-name FILE_NAME_WITHOUT_EXTENSION
+```
+
+Note: here we use regular syntax for providing CLI arguments to a Python script, not Hydra's `param_name=param_value` syntax.
+
+As a concrete example, this becomes particularly handy when you have a folder with training outputs, and would like to re-run the training. For example, say you previously ran the training script with one of the earlier commands and have `outputs/train/my_experiment/checkpoints/pretrained_model/config.yaml`. This `config.yaml` file will have the full set of configuration parameters within it. To run the training with the same configuration again, do:
+
+```bash
+python lerobot/scripts/train.py --config-dir outputs/train/my_experiment/checkpoints/last/pretrained_model --config-name config
+```
+
+Note that you may still use the regular syntax for config parameter overrides (eg: by adding `training.offline_steps=200000`).
+
+---
+
+So far we've seen how to train Diffusion Policy for PushT and ACT for ALOHA. Now, what if we want to train ACT for PushT? Well, there are aspects of the ACT configuration that are specific to the ALOHA environments, and these happen to be incompatible with PushT. Therefore, trying to run the following will almost certainly raise an exception of sorts (eg: feature dimension mismatch):
+
+```bash
+python lerobot/scripts/train.py policy=act env=pusht dataset_repo_id=lerobot/pusht
+```
+
+Please, head on over to our [advanced tutorial on adapting policy configuration to various environments](./advanced/train_act_pusht/train_act_pusht.md) to learn more.
+
+Or in the meantime, happy coding! 🤗

examples/5_resume_training.md

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

examples/advanced/1_train_act_pusht/act_pusht.yaml

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

examples/advanced/1_train_act_pusht/train_act_pusht.md

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

gym_dora/README.md

@@ -1 +0,0 @@
-# gym_dora

gym_dora/example.py

@@ -1,17 +0,0 @@
-import gymnasium as gym
-
-import gym_dora  # noqa: F401
-
-env = gym.make("gym_dora/DoraAloha-v0", disable_env_checker=True)
-obs = env.reset()
-
-policy = ...  # make_policy
-
-done = False
-while not done:
-    actions = policy.select_action(obs)
-    observation, reward, terminated, truncated, info = env.step(actions)
-
-    done = terminated | truncated | done
-
-env.close()

gym_dora/gym_dora/__init__.py

@@ -1,17 +0,0 @@
-from gymnasium.envs.registration import register
-
-register(
-    id="gym_dora/DoraAloha-v0",
-    entry_point="gym_dora.env:DoraEnv",
-    max_episode_steps=300,
-    nondeterministic=True,
-    kwargs={"model": "aloha"},
-)
-
-register(
-    id="gym_dora/DoraKoch-v0",
-    entry_point="gym_dora.env:DoraEnv",
-    max_episode_steps=300,
-    nondeterministic=True,
-    kwargs={"model": "koch"},
-)

gym_dora/gym_dora/env.py

@@ -1,199 +0,0 @@
-import os
-
-import gymnasium as gym
-import numpy as np
-import pyarrow as pa
-from dora import Node
-from gymnasium import spaces
-
-FPS = int(os.getenv("FPS", "30"))
-IMAGE_WIDTH = int(os.getenv("IMAGE_WIDTH", "640"))
-IMAGE_HEIGHT = int(os.getenv("IMAGE_HEIGHT", "480"))
-
-ALOHA_JOINTS = [
-    # absolute joint position
-    "left_arm_waist",
-    "left_arm_shoulder",
-    "left_arm_elbow",
-    "left_arm_forearm_roll",
-    "left_arm_wrist_angle",
-    "left_arm_wrist_rotate",
-    # normalized gripper position 0: close, 1: open
-    "left_arm_gripper",
-    # absolute joint position
-    "right_arm_waist",
-    "right_arm_shoulder",
-    "right_arm_elbow",
-    "right_arm_forearm_roll",
-    "right_arm_wrist_angle",
-    "right_arm_wrist_rotate",
-    # normalized gripper position 0: close, 1: open
-    "right_arm_gripper",
-]
-ALOHA_ACTIONS = [
-    # position and quaternion for end effector
-    "left_arm_waist",
-    "left_arm_shoulder",
-    "left_arm_elbow",
-    "left_arm_forearm_roll",
-    "left_arm_wrist_angle",
-    "left_arm_wrist_rotate",
-    # normalized gripper position (0: close, 1: open)
-    "left_arm_gripper",
-    "right_arm_waist",
-    "right_arm_shoulder",
-    "right_arm_elbow",
-    "right_arm_forearm_roll",
-    "right_arm_wrist_angle",
-    "right_arm_wrist_rotate",
-    # normalized gripper position (0: close, 1: open)
-    "right_arm_gripper",
-]
-
-
-class DoraEnv(gym.Env):
-    metadata = {"render_modes": ["rgb_array"], "render_fps": FPS}
-
-    def __init__(
-        self,
-        model="aloha",
-        observation_width=IMAGE_WIDTH,
-        observation_height=IMAGE_HEIGHT,
-        cameras_names=None,
-        num_joints=None,
-        num_actions=None,
-    ):
-        """Initializes the Dora environment.
-
-        Args:
-            model (str): The model to use. Either 'aloha' or 'custom'.
-            observation_width (int): The width of the observation image.
-            observation_height (int): The height of the observation image.
-            cameras_names (list): A list of camera names to use. If not provided, the default is ['cam_high', 'cam_low', 'cam_left_wrist', 'cam_right_wrist'].
-            num_joints (int): The number of joints in the model. If not provided, the default is 14 for 'aloha' and 6 for 'fivedof'.
-            num_actions (int): The number of actions in the model. If not provided, the default is 14 for 'aloha' and 6 for 'fivedof'.
-        """
-        super().__init__()
-
-        # Initialize a new node
-        self.node = Node() if os.environ.get("DORA_NODE_CONFIG", None) is not None else None
-        self.observation = {"pixels": {}, "agent_pos": None}
-        self.terminated = False
-
-        self.observation_height = observation_height
-        self.observation_width = observation_width
-
-        # Observation space
-        if model == "aloha":
-            self.observation_space = spaces.Dict(
-                {
-                    "pixels": spaces.Dict(
-                        {
-                            "cam_high": spaces.Box(
-                                low=0,
-                                high=255,
-                                shape=(self.observation_height, self.observation_width, 3),
-                                dtype=np.uint8,
-                            ),
-                            "cam_low": spaces.Box(
-                                low=0,
-                                high=255,
-                                shape=(self.observation_height, self.observation_width, 3),
-                                dtype=np.uint8,
-                            ),
-                            "cam_left_wrist": spaces.Box(
-                                low=0,
-                                high=255,
-                                shape=(self.observation_height, self.observation_width, 3),
-                                dtype=np.uint8,
-                            ),
-                            "cam_right_wrist": spaces.Box(
-                                low=0,
-                                high=255,
-                                shape=(self.observation_height, self.observation_width, 3),
-                                dtype=np.uint8,
-                            ),
-                        }
-                    ),
-                    "agent_pos": spaces.Box(
-                        low=-1000.0,
-                        high=1000.0,
-                        shape=(len(ALOHA_JOINTS),),
-                        dtype=np.float64,
-                    ),
-                }
-            )
-        elif model == "custom":
-            pixel_dict = {}
-            for camera in cameras_names:
-                assert camera.startswith("cam"), "Camera names must start with 'cam'"
-                pixel_dict[camera] = spaces.Box(
-                    low=0,
-                    high=255,
-                    shape=(self.observation_height, self.observation_width, 3),
-                    dtype=np.uint8,
-                )
-            self.observation_space = spaces.Dict(
-                {
-                    "pixels": spaces.Dict(pixel_dict),
-                    "agent_pos": spaces.Box(
-                        low=-1000.0,
-                        high=1000.0,
-                        shape=(num_joints,),
-                        dtype=np.float64,
-                    ),
-                }
-            )
-        else:
-            raise ValueError("Model must be either 'aloha' or 'custom'.")
-
-        # Action space
-        if model == "aloha":
-            self.action_space = spaces.Box(low=-1, high=1, shape=(len(ALOHA_ACTIONS),), dtype=np.float32)
-        elif model == "custom":
-            self.action_space = spaces.Box(low=-1, high=1, shape=(num_actions,), dtype=np.float32)
-
-    def _get_obs(self):
-        while True:
-            event = self.node.next(timeout=0.001)
-
-            ## If event is None, the node event stream is closed and we should terminate the env
-            if event is None:
-                self.terminated = True
-                break
-
-            if event["type"] == "INPUT":
-                # Map Image input into pixels key within Aloha environment
-                if "cam" in event["id"]:
-                    self.observation["pixels"][event["id"]] = (
-                        event["value"].to_numpy().reshape(self.observation_height, self.observation_width, 3)
-                    )
-                else:
-                    # Map other inputs into the observation dictionary using the event id as key
-                    self.observation[event["id"]] = event["value"].to_numpy()
-
-            # If the event is a timeout error break the update loop.
-            elif event["type"] == "ERROR":
-                break
-
-    def reset(self, seed: int | None = None):
-        self.node.send_output("reset")
-        self._get_obs()
-        self.terminated = False
-        info = {}
-        return self.observation, info
-
-    def step(self, action: np.ndarray):
-        # Send the action to the dataflow as action key.
-        self.node.send_output("action", pa.array(action))
-        self._get_obs()
-        reward = 0
-        terminated = truncated = self.terminated
-        info = {}
-        return self.observation, reward, terminated, truncated, info
-
-    def render(self): ...
-
-    def close(self):
-        # Drop the node
-        del self.node

gym_dora/poetry.lock

@@ -1,182 +0,0 @@
-# This file is automatically @generated by Poetry 1.8.2 and should not be changed by hand.
-
-[[package]]
-name = "cloudpickle"
-version = "3.0.0"
-description = "Pickler class to extend the standard pickle.Pickler functionality"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "cloudpickle-3.0.0-py3-none-any.whl", hash = "sha256:246ee7d0c295602a036e86369c77fecda4ab17b506496730f2f576d9016fd9c7"},
-    {file = "cloudpickle-3.0.0.tar.gz", hash = "sha256:996d9a482c6fb4f33c1a35335cf8afd065d2a56e973270364840712d9131a882"},
-]
-
-[[package]]
-name = "dora-rs"
-version = "0.3.4"
-description = "`dora` goal is to be a low latency, composable, and distributed data flow."
-optional = false
-python-versions = "*"
-files = [
-    {file = "dora_rs-0.3.4-cp37-abi3-macosx_10_12_x86_64.whl", hash = "sha256:d1b738eea5a4966d731c26c6b6a0a50a491a24f7e9e335475f983cfc6f0da19e"},
-    {file = "dora_rs-0.3.4-cp37-abi3-macosx_11_0_arm64.whl", hash = "sha256:80b724871618c78a4e5863938fa66724176cc40352771087aebe1e62a8141157"},
-    {file = "dora_rs-0.3.4-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3a3919e157b47dc1dbc74c040a73087a4485f0d1bee99b6adcdbc36559400fe2"},
-    {file = "dora_rs-0.3.4-cp37-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f7c95f6e5858fd651d6cd220e4f052e99db2944b9c37fb0b5402d60ac4b41a63"},
-    {file = "dora_rs-0.3.4-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:37d915fbbca282446235c98a9ca08389aa3ef3155d4e88c6c136326e9a830042"},
-    {file = "dora_rs-0.3.4-cp37-abi3-win32.whl", hash = "sha256:c9f7f22f65c884ec9bee0245ce98d0c7fad25dec0f982e566f844b5e8e58818f"},
-    {file = "dora_rs-0.3.4-cp37-abi3-win_amd64.whl", hash = "sha256:0a6a37f96a9f6e13b58b02a6ea75af192af5fbe4f456f6a67b1f239c3cee3276"},
-    {file = "dora_rs-0.3.4.tar.gz", hash = "sha256:05c5d0db0d23d7c4669995ae34db11cd636dbf91f5705d832669bd04e7452903"},
-]
-
-[package.dependencies]
-pyarrow = "*"
-
-[[package]]
-name = "farama-notifications"
-version = "0.0.4"
-description = "Notifications for all Farama Foundation maintained libraries."
-optional = false
-python-versions = "*"
-files = [
-    {file = "Farama-Notifications-0.0.4.tar.gz", hash = "sha256:13fceff2d14314cf80703c8266462ebf3733c7d165336eee998fc58e545efd18"},
-    {file = "Farama_Notifications-0.0.4-py3-none-any.whl", hash = "sha256:14de931035a41961f7c056361dc7f980762a143d05791ef5794a751a2caf05ae"},
-]
-
-[[package]]
-name = "gymnasium"
-version = "0.29.1"
-description = "A standard API for reinforcement learning and a diverse set of reference environments (formerly Gym)."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "gymnasium-0.29.1-py3-none-any.whl", hash = "sha256:61c3384b5575985bb7f85e43213bcb40f36fcdff388cae6bc229304c71f2843e"},
-    {file = "gymnasium-0.29.1.tar.gz", hash = "sha256:1a532752efcb7590478b1cc7aa04f608eb7a2fdad5570cd217b66b6a35274bb1"},
-]
-
-[package.dependencies]
-cloudpickle = ">=1.2.0"
-farama-notifications = ">=0.0.1"
-numpy = ">=1.21.0"
-typing-extensions = ">=4.3.0"
-
-[package.extras]
-accept-rom-license = ["autorom[accept-rom-license] (>=0.4.2,<0.5.0)"]
-all = ["box2d-py (==2.3.5)", "cython (<3)", "imageio (>=2.14.1)", "jax (>=0.4.0)", "jaxlib (>=0.4.0)", "lz4 (>=3.1.0)", "matplotlib (>=3.0)", "moviepy (>=1.0.0)", "mujoco (>=2.3.3)", "mujoco-py (>=2.1,<2.2)", "opencv-python (>=3.0)", "pygame (>=2.1.3)", "shimmy[atari] (>=0.1.0,<1.0)", "swig (==4.*)", "torch (>=1.0.0)"]
-atari = ["shimmy[atari] (>=0.1.0,<1.0)"]
-box2d = ["box2d-py (==2.3.5)", "pygame (>=2.1.3)", "swig (==4.*)"]
-classic-control = ["pygame (>=2.1.3)", "pygame (>=2.1.3)"]
-jax = ["jax (>=0.4.0)", "jaxlib (>=0.4.0)"]
-mujoco = ["imageio (>=2.14.1)", "mujoco (>=2.3.3)"]
-mujoco-py = ["cython (<3)", "cython (<3)", "mujoco-py (>=2.1,<2.2)", "mujoco-py (>=2.1,<2.2)"]
-other = ["lz4 (>=3.1.0)", "matplotlib (>=3.0)", "moviepy (>=1.0.0)", "opencv-python (>=3.0)", "torch (>=1.0.0)"]
-testing = ["pytest (==7.1.3)", "scipy (>=1.7.3)"]
-toy-text = ["pygame (>=2.1.3)", "pygame (>=2.1.3)"]
-
-[[package]]
-name = "numpy"
-version = "1.26.4"
-description = "Fundamental package for array computing in Python"
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "numpy-1.26.4-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:9ff0f4f29c51e2803569d7a51c2304de5554655a60c5d776e35b4a41413830d0"},
-    {file = "numpy-1.26.4-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:2e4ee3380d6de9c9ec04745830fd9e2eccb3e6cf790d39d7b98ffd19b0dd754a"},
-    {file = "numpy-1.26.4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d209d8969599b27ad20994c8e41936ee0964e6da07478d6c35016bc386b66ad4"},
-    {file = "numpy-1.26.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ffa75af20b44f8dba823498024771d5ac50620e6915abac414251bd971b4529f"},
-    {file = "numpy-1.26.4-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:62b8e4b1e28009ef2846b4c7852046736bab361f7aeadeb6a5b89ebec3c7055a"},
-    {file = "numpy-1.26.4-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:a4abb4f9001ad2858e7ac189089c42178fcce737e4169dc61321660f1a96c7d2"},
-    {file = "numpy-1.26.4-cp310-cp310-win32.whl", hash = "sha256:bfe25acf8b437eb2a8b2d49d443800a5f18508cd811fea3181723922a8a82b07"},
-    {file = "numpy-1.26.4-cp310-cp310-win_amd64.whl", hash = "sha256:b97fe8060236edf3662adfc2c633f56a08ae30560c56310562cb4f95500022d5"},
-    {file = "numpy-1.26.4-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:4c66707fabe114439db9068ee468c26bbdf909cac0fb58686a42a24de1760c71"},
-    {file = "numpy-1.26.4-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:edd8b5fe47dab091176d21bb6de568acdd906d1887a4584a15a9a96a1dca06ef"},
-    {file = "numpy-1.26.4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:7ab55401287bfec946ced39700c053796e7cc0e3acbef09993a9ad2adba6ca6e"},
-    {file = "numpy-1.26.4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:666dbfb6ec68962c033a450943ded891bed2d54e6755e35e5835d63f4f6931d5"},
-    {file = "numpy-1.26.4-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:96ff0b2ad353d8f990b63294c8986f1ec3cb19d749234014f4e7eb0112ceba5a"},
-    {file = "numpy-1.26.4-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:60dedbb91afcbfdc9bc0b1f3f402804070deed7392c23eb7a7f07fa857868e8a"},
-    {file = "numpy-1.26.4-cp311-cp311-win32.whl", hash = "sha256:1af303d6b2210eb850fcf03064d364652b7120803a0b872f5211f5234b399f20"},
-    {file = "numpy-1.26.4-cp311-cp311-win_amd64.whl", hash = "sha256:cd25bcecc4974d09257ffcd1f098ee778f7834c3ad767fe5db785be9a4aa9cb2"},
-    {file = "numpy-1.26.4-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:b3ce300f3644fb06443ee2222c2201dd3a89ea6040541412b8fa189341847218"},
-    {file = "numpy-1.26.4-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:03a8c78d01d9781b28a6989f6fa1bb2c4f2d51201cf99d3dd875df6fbd96b23b"},
-    {file = "numpy-1.26.4-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:9fad7dcb1aac3c7f0584a5a8133e3a43eeb2fe127f47e3632d43d677c66c102b"},
-    {file = "numpy-1.26.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:675d61ffbfa78604709862923189bad94014bef562cc35cf61d3a07bba02a7ed"},
-    {file = "numpy-1.26.4-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:ab47dbe5cc8210f55aa58e4805fe224dac469cde56b9f731a4c098b91917159a"},
-    {file = "numpy-1.26.4-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:1dda2e7b4ec9dd512f84935c5f126c8bd8b9f2fc001e9f54af255e8c5f16b0e0"},
-    {file = "numpy-1.26.4-cp312-cp312-win32.whl", hash = "sha256:50193e430acfc1346175fcbdaa28ffec49947a06918b7b92130744e81e640110"},
-    {file = "numpy-1.26.4-cp312-cp312-win_amd64.whl", hash = "sha256:08beddf13648eb95f8d867350f6a018a4be2e5ad54c8d8caed89ebca558b2818"},
-    {file = "numpy-1.26.4-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:7349ab0fa0c429c82442a27a9673fc802ffdb7c7775fad780226cb234965e53c"},
-    {file = "numpy-1.26.4-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:52b8b60467cd7dd1e9ed082188b4e6bb35aa5cdd01777621a1658910745b90be"},
-    {file = "numpy-1.26.4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d5241e0a80d808d70546c697135da2c613f30e28251ff8307eb72ba696945764"},
-    {file = "numpy-1.26.4-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f870204a840a60da0b12273ef34f7051e98c3b5961b61b0c2c1be6dfd64fbcd3"},
-    {file = "numpy-1.26.4-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:679b0076f67ecc0138fd2ede3a8fd196dddc2ad3254069bcb9faf9a79b1cebcd"},
-    {file = "numpy-1.26.4-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:47711010ad8555514b434df65f7d7b076bb8261df1ca9bb78f53d3b2db02e95c"},
-    {file = "numpy-1.26.4-cp39-cp39-win32.whl", hash = "sha256:a354325ee03388678242a4d7ebcd08b5c727033fcff3b2f536aea978e15ee9e6"},
-    {file = "numpy-1.26.4-cp39-cp39-win_amd64.whl", hash = "sha256:3373d5d70a5fe74a2c1bb6d2cfd9609ecf686d47a2d7b1d37a8f3b6bf6003aea"},
-    {file = "numpy-1.26.4-pp39-pypy39_pp73-macosx_10_9_x86_64.whl", hash = "sha256:afedb719a9dcfc7eaf2287b839d8198e06dcd4cb5d276a3df279231138e83d30"},
-    {file = "numpy-1.26.4-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:95a7476c59002f2f6c590b9b7b998306fba6a5aa646b1e22ddfeaf8f78c3a29c"},
-    {file = "numpy-1.26.4-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:7e50d0a0cc3189f9cb0aeb3a6a6af18c16f59f004b866cd2be1c14b36134a4a0"},
-    {file = "numpy-1.26.4.tar.gz", hash = "sha256:2a02aba9ed12e4ac4eb3ea9421c420301a0c6460d9830d74a9df87efa4912010"},
-]
-
-[[package]]
-name = "pyarrow"
-version = "16.1.0"
-description = "Python library for Apache Arrow"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "pyarrow-16.1.0-cp310-cp310-macosx_10_15_x86_64.whl", hash = "sha256:17e23b9a65a70cc733d8b738baa6ad3722298fa0c81d88f63ff94bf25eaa77b9"},
-    {file = "pyarrow-16.1.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:4740cc41e2ba5d641071d0ab5e9ef9b5e6e8c7611351a5cb7c1d175eaf43674a"},
-    {file = "pyarrow-16.1.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:98100e0268d04e0eec47b73f20b39c45b4006f3c4233719c3848aa27a03c1aef"},
-    {file = "pyarrow-16.1.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f68f409e7b283c085f2da014f9ef81e885d90dcd733bd648cfba3ef265961848"},
-    {file = "pyarrow-16.1.0-cp310-cp310-manylinux_2_28_aarch64.whl", hash = "sha256:a8914cd176f448e09746037b0c6b3a9d7688cef451ec5735094055116857580c"},
-    {file = "pyarrow-16.1.0-cp310-cp310-manylinux_2_28_x86_64.whl", hash = "sha256:48be160782c0556156d91adbdd5a4a7e719f8d407cb46ae3bb4eaee09b3111bd"},
-    {file = "pyarrow-16.1.0-cp310-cp310-win_amd64.whl", hash = "sha256:9cf389d444b0f41d9fe1444b70650fea31e9d52cfcb5f818b7888b91b586efff"},
-    {file = "pyarrow-16.1.0-cp311-cp311-macosx_10_15_x86_64.whl", hash = "sha256:d0ebea336b535b37eee9eee31761813086d33ed06de9ab6fc6aaa0bace7b250c"},
-    {file = "pyarrow-16.1.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:2e73cfc4a99e796727919c5541c65bb88b973377501e39b9842ea71401ca6c1c"},
-    {file = "pyarrow-16.1.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:bf9251264247ecfe93e5f5a0cd43b8ae834f1e61d1abca22da55b20c788417f6"},
-    {file = "pyarrow-16.1.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ddf5aace92d520d3d2a20031d8b0ec27b4395cab9f74e07cc95edf42a5cc0147"},
-    {file = "pyarrow-16.1.0-cp311-cp311-manylinux_2_28_aarch64.whl", hash = "sha256:25233642583bf658f629eb230b9bb79d9af4d9f9229890b3c878699c82f7d11e"},
-    {file = "pyarrow-16.1.0-cp311-cp311-manylinux_2_28_x86_64.whl", hash = "sha256:a33a64576fddfbec0a44112eaf844c20853647ca833e9a647bfae0582b2ff94b"},
-    {file = "pyarrow-16.1.0-cp311-cp311-win_amd64.whl", hash = "sha256:185d121b50836379fe012753cf15c4ba9638bda9645183ab36246923875f8d1b"},
-    {file = "pyarrow-16.1.0-cp312-cp312-macosx_10_15_x86_64.whl", hash = "sha256:2e51ca1d6ed7f2e9d5c3c83decf27b0d17bb207a7dea986e8dc3e24f80ff7d6f"},
-    {file = "pyarrow-16.1.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:06ebccb6f8cb7357de85f60d5da50e83507954af617d7b05f48af1621d331c9a"},
-    {file = "pyarrow-16.1.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b04707f1979815f5e49824ce52d1dceb46e2f12909a48a6a753fe7cafbc44a0c"},
-    {file = "pyarrow-16.1.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0d32000693deff8dc5df444b032b5985a48592c0697cb6e3071a5d59888714e2"},
-    {file = "pyarrow-16.1.0-cp312-cp312-manylinux_2_28_aarch64.whl", hash = "sha256:8785bb10d5d6fd5e15d718ee1d1f914fe768bf8b4d1e5e9bf253de8a26cb1628"},
-    {file = "pyarrow-16.1.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:e1369af39587b794873b8a307cc6623a3b1194e69399af0efd05bb202195a5a7"},
-    {file = "pyarrow-16.1.0-cp312-cp312-win_amd64.whl", hash = "sha256:febde33305f1498f6df85e8020bca496d0e9ebf2093bab9e0f65e2b4ae2b3444"},
-    {file = "pyarrow-16.1.0-cp38-cp38-macosx_10_15_x86_64.whl", hash = "sha256:b5f5705ab977947a43ac83b52ade3b881eb6e95fcc02d76f501d549a210ba77f"},
-    {file = "pyarrow-16.1.0-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:0d27bf89dfc2576f6206e9cd6cf7a107c9c06dc13d53bbc25b0bd4556f19cf5f"},
-    {file = "pyarrow-16.1.0-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0d07de3ee730647a600037bc1d7b7994067ed64d0eba797ac74b2bc77384f4c2"},
-    {file = "pyarrow-16.1.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fbef391b63f708e103df99fbaa3acf9f671d77a183a07546ba2f2c297b361e83"},
-    {file = "pyarrow-16.1.0-cp38-cp38-manylinux_2_28_aarch64.whl", hash = "sha256:19741c4dbbbc986d38856ee7ddfdd6a00fc3b0fc2d928795b95410d38bb97d15"},
-    {file = "pyarrow-16.1.0-cp38-cp38-manylinux_2_28_x86_64.whl", hash = "sha256:f2c5fb249caa17b94e2b9278b36a05ce03d3180e6da0c4c3b3ce5b2788f30eed"},
-    {file = "pyarrow-16.1.0-cp38-cp38-win_amd64.whl", hash = "sha256:e6b6d3cd35fbb93b70ade1336022cc1147b95ec6af7d36906ca7fe432eb09710"},
-    {file = "pyarrow-16.1.0-cp39-cp39-macosx_10_15_x86_64.whl", hash = "sha256:18da9b76a36a954665ccca8aa6bd9f46c1145f79c0bb8f4f244f5f8e799bca55"},
-    {file = "pyarrow-16.1.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:99f7549779b6e434467d2aa43ab2b7224dd9e41bdde486020bae198978c9e05e"},
-    {file = "pyarrow-16.1.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f07fdffe4fd5b15f5ec15c8b64584868d063bc22b86b46c9695624ca3505b7b4"},
-    {file = "pyarrow-16.1.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ddfe389a08ea374972bd4065d5f25d14e36b43ebc22fc75f7b951f24378bf0b5"},
-    {file = "pyarrow-16.1.0-cp39-cp39-manylinux_2_28_aarch64.whl", hash = "sha256:3b20bd67c94b3a2ea0a749d2a5712fc845a69cb5d52e78e6449bbd295611f3aa"},
-    {file = "pyarrow-16.1.0-cp39-cp39-manylinux_2_28_x86_64.whl", hash = "sha256:ba8ac20693c0bb0bf4b238751d4409e62852004a8cf031c73b0e0962b03e45e3"},
-    {file = "pyarrow-16.1.0-cp39-cp39-win_amd64.whl", hash = "sha256:31a1851751433d89a986616015841977e0a188662fcffd1a5677453f1df2de0a"},
-    {file = "pyarrow-16.1.0.tar.gz", hash = "sha256:15fbb22ea96d11f0b5768504a3f961edab25eaf4197c341720c4a387f6c60315"},
-]
-
-[package.dependencies]
-numpy = ">=1.16.6"
-
-[[package]]
-name = "typing-extensions"
-version = "4.11.0"
-description = "Backported and Experimental Type Hints for Python 3.8+"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "typing_extensions-4.11.0-py3-none-any.whl", hash = "sha256:c1f94d72897edaf4ce775bb7558d5b79d8126906a14ea5ed1635921406c0387a"},
-    {file = "typing_extensions-4.11.0.tar.gz", hash = "sha256:83f085bd5ca59c80295fc2a82ab5dac679cbe02b9f33f7d83af68e241bea51b0"},
-]
-
-[metadata]
-lock-version = "2.0"
-python-versions = "^3.10"
-content-hash = "7e437b5c547ebe11095f1ce4ff1851d636f8e707ad7de8a6224b0f9ad978240f"

gym_dora/pyproject.toml

@@ -1,17 +0,0 @@
-[tool.poetry]
-name = "gym-dora"
-version = "0.1.0"
-description = ""
-authors = ["Simon Alibert <alibert.sim@gmail.com>"]
-readme = "README.md"
-packages = [{ include = "gym_dora" }]
-
-[tool.poetry.dependencies]
-python = "^3.10"
-gymnasium = ">=0.29.1"
-dora-rs = ">=0.3.4"
-pyarrow = ">=12.0.0"
-
-[build-system]
-requires = ["poetry-core"]
-build-backend = "poetry.core.masonry.api"

lerobot/__init__.py

@@ -45,6 +45,9 @@ import itertools
 
 from lerobot.__version__ import __version__  # noqa: F401
 
+# TODO(rcadene): Improve policies and envs. As of now, an item in `available_policies`
+# refers to a yaml file AND a modeling name. Same for `available_envs` which refers to
+# a yaml file AND a environment name. The difference should be more obvious.
 available_tasks_per_env = {
     "aloha": [
         "AlohaInsertion-v0",
@@ -52,6 +55,7 @@ available_tasks_per_env = {
     ],
     "pusht": ["PushT-v0"],
     "xarm": ["XarmLift-v0"],
+    "dora_aloha_real": ["DoraAloha-v0", "DoraKoch-v0", "DoraReachy2-v0"],
 }
 available_envs = list(available_tasks_per_env.keys())
 
@@ -77,6 +81,23 @@ available_datasets_per_env = {
         "lerobot/xarm_push_medium_image",
         "lerobot/xarm_push_medium_replay_image",
     ],
+    "dora_aloha_real": [
+        "lerobot/aloha_static_battery",
+        "lerobot/aloha_static_candy",
+        "lerobot/aloha_static_coffee",
+        "lerobot/aloha_static_coffee_new",
+        "lerobot/aloha_static_cups_open",
+        "lerobot/aloha_static_fork_pick_up",
+        "lerobot/aloha_static_pingpong_test",
+        "lerobot/aloha_static_pro_pencil",
+        "lerobot/aloha_static_screw_driver",
+        "lerobot/aloha_static_tape",
+        "lerobot/aloha_static_thread_velcro",
+        "lerobot/aloha_static_towel",
+        "lerobot/aloha_static_vinh_cup",
+        "lerobot/aloha_static_vinh_cup_left",
+        "lerobot/aloha_static_ziploc_slide",
+    ],
 }
 
 available_real_world_datasets = [
@@ -108,16 +129,19 @@ available_datasets = list(
     itertools.chain(*available_datasets_per_env.values(), available_real_world_datasets)
 )
 
+# lists all available policies from `lerobot/common/policies` by their class attribute: `name`.
 available_policies = [
     "act",
     "diffusion",
     "tdmpc",
 ]
 
+# keys and values refer to yaml files
 available_policies_per_env = {
     "aloha": ["act"],
     "pusht": ["diffusion"],
     "xarm": ["tdmpc"],
+    "dora_aloha_real": ["act_real"],
 }
 
 env_task_pairs = [(env, task) for env, tasks in available_tasks_per_env.items() for task in tasks]

lerobot/common/datasets/compute_stats.py

@@ -16,17 +16,15 @@
 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):
+def get_stats_einops_patterns(dataset, num_workers=0):
     """These einops patterns will be used to aggregate batches and compute statistics.
 
     Note: We assume the images are in channel first format
@@ -66,9 +64,8 @@ def get_stats_einops_patterns(dataset: LeRobotDataset | datasets.Dataset, num_wo
     return stats_patterns
 
 
-def compute_stats(
-    dataset: LeRobotDataset | datasets.Dataset, batch_size=32, num_workers=16, max_num_samples=None
-):
+def compute_stats(dataset, batch_size=32, num_workers=16, max_num_samples=None):
+    """Compute mean/std and min/max statistics of all data keys in a LeRobotDataset."""
     if max_num_samples is None:
         max_num_samples = len(dataset)
 
@@ -159,3 +156,54 @@ def compute_stats(
             "min": min[key],
         }
     return stats
+
+
+def aggregate_stats(ls_datasets) -> dict[str, torch.Tensor]:
+    """Aggregate stats of multiple LeRobot datasets into one set of stats without recomputing from scratch.
+
+    The final stats will have the union of all data keys from each of the datasets.
+
+    The final stats will have the union of all data keys from each of the datasets. For instance:
+    - new_max = max(max_dataset_0, max_dataset_1, ...)
+    - new_min = min(min_dataset_0, min_dataset_1, ...)
+    - new_mean = (mean of all data)
+    - new_std = (std of all data)
+    """
+    data_keys = set()
+    for dataset in ls_datasets:
+        data_keys.update(dataset.stats.keys())
+    stats = {k: {} for k in data_keys}
+    for data_key in data_keys:
+        for stat_key in ["min", "max"]:
+            # compute `max(dataset_0["max"], dataset_1["max"], ...)`
+            stats[data_key][stat_key] = einops.reduce(
+                torch.stack([d.stats[data_key][stat_key] for d in ls_datasets if data_key in d.stats], dim=0),
+                "n ... -> ...",
+                stat_key,
+            )
+        total_samples = sum(d.num_samples for d in ls_datasets if data_key in d.stats)
+        # Compute the "sum" statistic by multiplying each mean by the number of samples in the respective
+        # dataset, then divide by total_samples to get the overall "mean".
+        # NOTE: the brackets around (d.num_samples / total_samples) are needed tor minimize the risk of
+        # numerical overflow!
+        stats[data_key]["mean"] = sum(
+            d.stats[data_key]["mean"] * (d.num_samples / total_samples)
+            for d in ls_datasets
+            if data_key in d.stats
+        )
+        # The derivation for standard deviation is a little more involved but is much in the same spirit as
+        # the computation of the mean.
+        # Given two sets of data where the statistics are known:
+        # σ_combined = sqrt[ (n1 * (σ1^2 + d1^2) + n2 * (σ2^2 + d2^2)) / (n1 + n2) ]
+        # where d1 = μ1 - μ_combined, d2 = μ2 - μ_combined
+        # NOTE: the brackets around (d.num_samples / total_samples) are needed tor minimize the risk of
+        # numerical overflow!
+        stats[data_key]["std"] = torch.sqrt(
+            sum(
+                (d.stats[data_key]["std"] ** 2 + (d.stats[data_key]["mean"] - stats[data_key]["mean"]) ** 2)
+                * (d.num_samples / total_samples)
+                for d in ls_datasets
+                if data_key in d.stats
+            )
+        )
+    return stats

lerobot/common/datasets/factory.py

@@ -16,34 +16,73 @@
 import logging
 
 import torch
-from omegaconf import OmegaConf
+from omegaconf import ListConfig, OmegaConf
 
-from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
 
 
-def make_dataset(
-    cfg,
-    split="train",
-):
-    if cfg.env.name not in cfg.dataset_repo_id:
-        logging.warning(
-            f"There might be a mismatch between your training dataset ({cfg.dataset_repo_id=}) and your "
-            f"environment ({cfg.env.name=})."
-        )
+def resolve_delta_timestamps(cfg):
+    """Resolves delta_timestamps config key (in-place) by using `eval`.
 
+    Doesn't do anything if delta_timestamps is not specified or has already been resolve (as evidenced by
+    the data type of its values).
+    """
     delta_timestamps = cfg.training.get("delta_timestamps")
     if delta_timestamps is not None:
         for key in delta_timestamps:
             if isinstance(delta_timestamps[key], str):
-                delta_timestamps[key] = eval(delta_timestamps[key])
+                # TODO(rcadene, alexander-soare): remove `eval` to avoid exploit
+                cfg.training.delta_timestamps[key] = eval(delta_timestamps[key])
+
+
+def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotDataset:
+    """
+    Args:
+        cfg: A Hydra config as per the LeRobot config scheme.
+        split: Select the data subset used to create an instance of LeRobotDataset.
+            All datasets hosted on [lerobot](https://huggingface.co/lerobot) contain only one subset: "train".
+            Thus, by default, `split="train"` selects all the available data. `split` aims to work like the
+            slicer in the hugging face datasets:
+            https://huggingface.co/docs/datasets/v2.19.0/loading#slice-splits
+            As of now, it only supports `split="train[:n]"` to load the first n frames of the dataset or
+            `split="train[n:]"` to load the last n frames. For instance `split="train[:1000]"`.
+    Returns:
+        The LeRobotDataset.
+    """
+    if not isinstance(cfg.dataset_repo_id, (str, ListConfig)):
+        raise ValueError(
+            "Expected cfg.dataset_repo_id to be either a single string to load one dataset or a list of "
+            "strings to load multiple datasets."
+        )
+
+    # A soft check to warn if the environment matches the dataset. Don't check if we are using a real world env (dora).
+    if cfg.env.name != "dora":
+        if isinstance(cfg.dataset_repo_id, str):
+            dataset_repo_ids = [cfg.dataset_repo_id]  # single dataset
+        else:
+            dataset_repo_ids = cfg.dataset_repo_id  # multiple datasets
+
+        for dataset_repo_id in dataset_repo_ids:
+            if cfg.env.name not in dataset_repo_id:
+                logging.warning(
+                    f"There might be a mismatch between your training dataset ({dataset_repo_id=}) and your "
+                    f"environment ({cfg.env.name=})."
+                )
+
+    resolve_delta_timestamps(cfg)
 
     # TODO(rcadene): add data augmentations
 
-    dataset = LeRobotDataset(
-        cfg.dataset_repo_id,
-        split=split,
-        delta_timestamps=delta_timestamps,
-    )
+    if isinstance(cfg.dataset_repo_id, str):
+        dataset = LeRobotDataset(
+            cfg.dataset_repo_id,
+            split=split,
+            delta_timestamps=cfg.training.get("delta_timestamps"),
+        )
+    else:
+        dataset = MultiLeRobotDataset(
+            cfg.dataset_repo_id, split=split, delta_timestamps=cfg.training.get("delta_timestamps")
+        )
 
     if cfg.get("override_dataset_stats"):
         for key, stats_dict in cfg.override_dataset_stats.items():

lerobot/common/datasets/lerobot_dataset.py

@@ -13,12 +13,16 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import logging
 import os
 from pathlib import Path
+from typing import Callable
 
 import datasets
 import torch
+import torch.utils
 
+from lerobot.common.datasets.compute_stats import aggregate_stats
 from lerobot.common.datasets.utils import (
     calculate_episode_data_index,
     load_episode_data_index,
@@ -42,7 +46,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         version: str | None = CODEBASE_VERSION,
         root: Path | None = DATA_DIR,
         split: str = "train",
-        transform: callable = None,
+        transform: Callable | None = None,
         delta_timestamps: dict[list[float]] | None = None,
     ):
         super().__init__()
@@ -171,7 +175,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
     @classmethod
     def from_preloaded(
         cls,
-        repo_id: str,
+        repo_id: str = "from_preloaded",
         version: str | None = CODEBASE_VERSION,
         root: Path | None = None,
         split: str = "train",
@@ -183,7 +187,15 @@ class LeRobotDataset(torch.utils.data.Dataset):
         stats=None,
         info=None,
         videos_dir=None,
-    ):
+    ) -> "LeRobotDataset":
+        """Create a LeRobot Dataset from existing data and attributes instead of loading from the filesystem.
+
+        It is especially useful when converting raw data into LeRobotDataset before saving the dataset
+        on the filesystem or uploading to the hub.
+
+        Note: Meta-data attributes like `repo_id`, `version`, `root`, etc are optional and potentially
+        meaningless depending on the downstream usage of the return dataset.
+        """
         # create an empty object of type LeRobotDataset
         obj = cls.__new__(cls)
         obj.repo_id = repo_id
@@ -195,6 +207,193 @@ class LeRobotDataset(torch.utils.data.Dataset):
         obj.hf_dataset = hf_dataset
         obj.episode_data_index = episode_data_index
         obj.stats = stats
-        obj.info = info
+        obj.info = info if info is not None else {}
         obj.videos_dir = videos_dir
         return obj
+
+
+class MultiLeRobotDataset(torch.utils.data.Dataset):
+    """A dataset consisting of multiple underlying `LeRobotDataset`s.
+
+    The underlying `LeRobotDataset`s are effectively concatenated, and this class adopts much of the API
+    structure of `LeRobotDataset`.
+    """
+
+    def __init__(
+        self,
+        repo_ids: list[str],
+        version: str | None = CODEBASE_VERSION,
+        root: Path | None = DATA_DIR,
+        split: str = "train",
+        transform: Callable | None = None,
+        delta_timestamps: dict[list[float]] | None = None,
+    ):
+        super().__init__()
+        self.repo_ids = repo_ids
+        # Construct the underlying datasets passing everything but `transform` and `delta_timestamps` which
+        # are handled by this class.
+        self._datasets = [
+            LeRobotDataset(
+                repo_id,
+                version=version,
+                root=root,
+                split=split,
+                delta_timestamps=delta_timestamps,
+                transform=transform,
+            )
+            for repo_id in repo_ids
+        ]
+        # Check that some properties are consistent across datasets. Note: We may relax some of these
+        # consistency requirements in future iterations of this class.
+        for repo_id, dataset in zip(self.repo_ids, self._datasets, strict=True):
+            if dataset.info != self._datasets[0].info:
+                raise ValueError(
+                    f"Detected a mismatch in dataset info between {self.repo_ids[0]} and {repo_id}. This is "
+                    "not yet supported."
+                )
+        # Disable any data keys that are not common across all of the datasets. Note: we may relax this
+        # restriction in future iterations of this class. For now, this is necessary at least for being able
+        # to use PyTorch's default DataLoader collate function.
+        self.disabled_data_keys = set()
+        intersection_data_keys = set(self._datasets[0].hf_dataset.features)
+        for dataset in self._datasets:
+            intersection_data_keys.intersection_update(dataset.hf_dataset.features)
+        if len(intersection_data_keys) == 0:
+            raise RuntimeError(
+                "Multiple datasets were provided but they had no keys common to all of them. The "
+                "multi-dataset functionality currently only keeps common keys."
+            )
+        for repo_id, dataset in zip(self.repo_ids, self._datasets, strict=True):
+            extra_keys = set(dataset.hf_dataset.features).difference(intersection_data_keys)
+            logging.warning(
+                f"keys {extra_keys} of {repo_id} were disabled as they are not contained in all the "
+                "other datasets."
+            )
+            self.disabled_data_keys.update(extra_keys)
+
+        self.version = version
+        self.root = root
+        self.split = split
+        self.transform = transform
+        self.delta_timestamps = delta_timestamps
+        self.stats = aggregate_stats(self._datasets)
+
+    @property
+    def repo_id_to_index(self):
+        """Return a mapping from dataset repo_id to a dataset index automatically created by this class.
+
+        This index is incorporated as a data key in the dictionary returned by `__getitem__`.
+        """
+        return {repo_id: i for i, repo_id in enumerate(self.repo_ids)}
+
+    @property
+    def repo_index_to_id(self):
+        """Return the inverse mapping if repo_id_to_index."""
+        return {v: k for k, v in self.repo_id_to_index}
+
+    @property
+    def fps(self) -> int:
+        """Frames per second used during data collection.
+
+        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
+        """
+        return self._datasets[0].info["fps"]
+
+    @property
+    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.
+
+        NOTE: Fow now, this relies on a check in __init__ to make sure all sub-datasets have the same info.
+        """
+        return self._datasets[0].info.get("video", False)
+
+    @property
+    def features(self) -> datasets.Features:
+        features = {}
+        for dataset in self._datasets:
+            features.update({k: v for k, v in dataset.features.items() if k not in self.disabled_data_keys})
+        return features
+
+    @property
+    def camera_keys(self) -> list[str]:
+        """Keys to access image and video stream from cameras."""
+        keys = []
+        for key, feats in self.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.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 sum(d.num_samples for d in self._datasets)
+
+    @property
+    def num_episodes(self) -> int:
+        """Number of episodes."""
+        return sum(d.num_episodes for d in self._datasets)
+
+    @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
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        if idx >= len(self):
+            raise IndexError(f"Index {idx} out of bounds.")
+        # Determine which dataset to get an item from based on the index.
+        start_idx = 0
+        dataset_idx = 0
+        for dataset in self._datasets:
+            if idx >= start_idx + dataset.num_samples:
+                start_idx += dataset.num_samples
+                dataset_idx += 1
+                continue
+            break
+        else:
+            raise AssertionError("We expect the loop to break out as long as the index is within bounds.")
+        item = self._datasets[dataset_idx][idx - start_idx]
+        item["dataset_index"] = torch.tensor(dataset_idx)
+        for data_key in self.disabled_data_keys:
+            if data_key in item:
+                del item[data_key]
+        return item
+
+    def __repr__(self):
+        return (
+            f"{self.__class__.__name__}(\n"
+            f"  Repository IDs: '{self.repo_ids}',\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")"
+        )

lerobot/common/datasets/push_dataset_to_hub/aloha_dora_format.py

@@ -18,6 +18,7 @@ Contains utilities to process raw data format from dora-record
 """
 
 import logging
+import re
 from pathlib import Path
 
 import pandas as pd
@@ -40,7 +41,7 @@ def check_format(raw_dir) -> bool:
     return True
 
 
-def load_from_raw(raw_dir: Path, out_dir: Path):
+def load_from_raw(raw_dir: Path, out_dir: Path, fps: int):
     # Load data stream that will be used as reference for the timestamps synchronization
     reference_files = list(raw_dir.glob("observation.images.cam_*.parquet"))
     if len(reference_files) == 0:
@@ -56,24 +57,45 @@ def load_from_raw(raw_dir: Path, out_dir: Path):
         key = path.stem  # action or observation.state or ...
         if key == reference_key:
             continue
+        if "failed_episode_index" in key:
+            # TODO(rcadene): add support for removing episodes that are tagged as "failed"
+            continue
         modality_df = pd.read_parquet(path)
         modality_df = modality_df[["timestamp_utc", key]]
         df = pd.merge_asof(
             df,
             modality_df,
             on="timestamp_utc",
-            direction="backward",
+            # "nearest" is the best option over "backward", since the latter can desynchronizes camera timestamps by
+            # matching timestamps that are too far appart, in order to fit the backward constraints. It's not the case for "nearest".
+            # However, note that "nearest" might synchronize the reference camera with other cameras on slightly future timestamps.
+            # are too far appart.
+            direction="nearest",
+            tolerance=pd.Timedelta(f"{1/fps} seconds"),
         )
-
-    # Remove rows with a NaN in any column. It can happened during the first frames of an episode,
-    # because some cameras didnt start recording yet.
-    df = df.dropna(axis=0)
-
-    # Remove rows with episode_index -1 which indicates a failed episode
+    # Remove rows with episode_index -1 which indicates data that correspond to in-between episodes
     df = df[df["episode_index"] != -1]
 
+    image_keys = [key for key in df if "observation.images." in key]
+
+    def get_episode_index(row):
+        episode_index_per_cam = {}
+        for key in image_keys:
+            path = row[key][0]["path"]
+            match = re.search(r"_(\d{6}).mp4", path)
+            if not match:
+                raise ValueError(path)
+            episode_index = int(match.group(1))
+            episode_index_per_cam[key] = episode_index
+        if len(set(episode_index_per_cam.values())) != 1:
+            raise ValueError(
+                f"All cameras are expected to belong to the same episode, but getting {episode_index_per_cam}"
+            )
+        return episode_index
+
+    df["episode_index"] = df.apply(get_episode_index, axis=1)
+
     # dora only use arrays, so single values are encapsulated into a list
-    df["episode_index"] = df["episode_index"].map(lambda x: x[0])
     df["frame_index"] = df.groupby("episode_index").cumcount()
     df = df.reset_index()
     df["index"] = df.index
@@ -88,10 +110,16 @@ def load_from_raw(raw_dir: Path, out_dir: Path):
 
     del df["timestamp_utc"]
 
+    # sanity check
+    has_nan = df.isna().any().any()
+    if has_nan:
+        raise ValueError("Dataset contains Nan values.")
+
     # sanity check episode indices go from 0 to n-1
     ep_ids = [ep_idx for ep_idx, _ in df.groupby("episode_index")]
     expected_ep_ids = list(range(df["episode_index"].max() + 1))
-    assert ep_ids == expected_ep_ids, f"Episodes indices go from {ep_ids} instead of {expected_ep_ids}"
+    if ep_ids != expected_ep_ids:
+        raise ValueError(f"Episodes indices go from {ep_ids} instead of {expected_ep_ids}")
 
     # Create symlink to raw videos directory (that needs to be absolute not relative)
     out_dir.mkdir(parents=True, exist_ok=True)
@@ -104,7 +132,8 @@ def load_from_raw(raw_dir: Path, out_dir: Path):
             continue
         for ep_idx in ep_ids:
             video_path = videos_dir / f"{key}_episode_{ep_idx:06d}.mp4"
-            assert video_path.exists(), f"Video file not found in {video_path}"
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
 
     data_dict = {}
     for key in df:
@@ -116,7 +145,8 @@ def load_from_raw(raw_dir: Path, out_dir: Path):
 
             # sanity check the video path is well formated
             video_path = videos_dir.parent / data_dict[key][0]["path"]
-            assert video_path.exists(), f"Video file not found in {video_path}"
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
         # is number
         elif df[key].iloc[0].ndim == 0 or df[key].iloc[0].shape[0] == 1:
             data_dict[key] = torch.from_numpy(df[key].values)
@@ -190,7 +220,7 @@ def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=Tru
     if not video:
         raise NotImplementedError()
 
-    data_df, episode_data_index = load_from_raw(raw_dir, out_dir)
+    data_df, episode_data_index = load_from_raw(raw_dir, out_dir, fps)
     hf_dataset = to_hf_dataset(data_df, video)
 
     info = {

lerobot/common/datasets/push_dataset_to_hub/reachy2_hdf5_format.py

@@ -0,0 +1,189 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Contains utilities to process raw data format of HDF5 files like in: https://github.com/tonyzhaozh/act
+"""
+
+import gc
+import re
+import shutil
+from pathlib import Path
+
+import h5py
+import torch
+import tqdm
+from datasets import Dataset, Features, Image, Sequence, Value
+
+from lerobot.common.datasets.push_dataset_to_hub.utils import concatenate_episodes
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame
+
+
+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_ids"].keys() if "depth" not in key]  # noqa: SIM118
+    return rgb_cameras
+
+
+def check_format(raw_dir) -> bool:
+    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_ids/{camera}"].shape[0]
+                assert (raw_dir / hdf5_path.name.replace(".hdf5", f"_{camera}.mp4")).exists()
+
+                # assert data[f"/observations/images_ids/{camera}"].ndim == 4
+                # b, h, w, c = data[f"/observations/images_ids/{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):
+    hdf5_files = list(raw_dir.glob("*.hdf5"))
+    ep_dicts = []
+    episode_data_index = {"from": [], "to": []}
+
+    id_from = 0
+    for ep_idx, ep_path in tqdm.tqdm(enumerate(hdf5_files), total=len(hdf5_files)):
+        match = re.search(r"_(\d+).hdf5", ep_path.name)
+        if not match:
+            raise ValueError(ep_path.name)
+        raw_ep_idx = int(match.group(1))
+
+        with h5py.File(ep_path, "r") as ep:
+            num_frames = ep["/action"].shape[0]
+
+            # last step of demonstration is considered done
+            done = torch.zeros(num_frames, dtype=torch.bool)
+            done[-1] = True
+
+            state = torch.from_numpy(ep["/observations/qpos"][:])
+            action = torch.from_numpy(ep["/action"][:])
+            if "/observations/qvel" in ep:
+                velocity = torch.from_numpy(ep["/observations/qvel"][:])
+            if "/observations/effort" in ep:
+                effort = torch.from_numpy(ep["/observations/effort"][:])
+
+            ep_dict = {}
+
+            videos_dir = out_dir / "videos"
+            videos_dir.mkdir(parents=True, exist_ok=True)
+
+            for camera in get_cameras(ep):
+                img_key = f"observation.images.{camera}"
+
+                raw_fname = f"episode_{raw_ep_idx}_{camera}.mp4"
+                new_fname = f"{img_key}_episode_{ep_idx:06d}.mp4"
+                shutil.copy(str(raw_dir / raw_fname), str(videos_dir / new_fname))
+
+                # store the reference to the video frame
+                ep_dict[img_key] = [
+                    {"path": f"videos/{new_fname}", "timestamp": i / fps} for i in range(num_frames)
+                ]
+
+            ep_dict["observation.state"] = state
+            if "/observations/velocity" in ep:
+                ep_dict["observation.velocity"] = velocity
+            if "/observations/effort" in ep:
+                ep_dict["observation.effort"] = effort
+            ep_dict["action"] = action
+            ep_dict["episode_index"] = torch.tensor([ep_idx] * num_frames)
+            ep_dict["frame_index"] = torch.arange(0, num_frames, 1)
+            ep_dict["timestamp"] = torch.arange(0, num_frames, 1) / fps
+            ep_dict["next.done"] = done
+            # TODO(rcadene): add reward and success by computing them in sim
+
+            assert isinstance(ep_idx, int)
+            ep_dicts.append(ep_dict)
+
+            episode_data_index["from"].append(id_from)
+            episode_data_index["to"].append(id_from + num_frames)
+
+        id_from += num_frames
+
+        gc.collect()
+
+        # 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)
+    )
+    if "observation.velocity" in data_dict:
+        features["observation.velocity"] = Sequence(
+            length=data_dict["observation.velocity"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    if "observation.effort" in data_dict:
+        features["observation.effort"] = Sequence(
+            length=data_dict["observation.effort"].shape[1], feature=Value(dtype="float32", id=None)
+        )
+    features["action"] = Sequence(
+        length=data_dict["action"].shape[1], feature=Value(dtype="float32", id=None)
+    )
+    features["episode_index"] = Value(dtype="int64", id=None)
+    features["frame_index"] = Value(dtype="int64", id=None)
+    features["timestamp"] = Value(dtype="float32", id=None)
+    features["next.done"] = Value(dtype="bool", id=None)
+    features["index"] = Value(dtype="int64", id=None)
+
+    hf_dataset = Dataset.from_dict(data_dict, features=Features(features))
+    hf_dataset.set_transform(hf_transform_to_torch)
+    return hf_dataset
+
+
+def from_raw_to_lerobot_format(raw_dir: Path, out_dir: Path, fps=None, video=True, debug=False):
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+
+    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/sampler.py

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

lerobot/common/datasets/utils.py

@@ -59,7 +59,7 @@ def unflatten_dict(d, sep="/"):
     return outdict
 
 
-def hf_transform_to_torch(items_dict):
+def hf_transform_to_torch(items_dict: dict[torch.Tensor | None]):
     """Get a transform function that convert items from Hugging Face dataset (pyarrow)
     to torch tensors. Importantly, images are converted from PIL, which corresponds to
     a channel last representation (h w c) of uint8 type, to a torch image representation
@@ -73,6 +73,8 @@ def hf_transform_to_torch(items_dict):
         elif isinstance(first_item, dict) and "path" in first_item and "timestamp" in first_item:
             # video frame will be processed downstream
             pass
+        elif first_item is None:
+            pass
         else:
             items_dict[key] = [torch.tensor(x) for x in items_dict[key]]
     return items_dict
@@ -318,8 +320,7 @@ def calculate_episode_data_index(hf_dataset: datasets.Dataset) -> Dict[str, torc
 
 
 def reset_episode_index(hf_dataset: datasets.Dataset) -> datasets.Dataset:
-    """
-    Reset the `episode_index` of the provided HuggingFace Dataset.
+    """Reset the `episode_index` of the provided HuggingFace Dataset.
 
     `episode_data_index` (and related functionality such as `load_previous_and_future_frames`) requires the
     `episode_index` to be sorted, continuous (1,1,1 and not 1,2,1) and start at 0.
@@ -338,6 +339,7 @@ def reset_episode_index(hf_dataset: datasets.Dataset) -> datasets.Dataset:
         return example
 
     hf_dataset = hf_dataset.map(modify_ep_idx_func)
+
     return hf_dataset
 
 

lerobot/common/envs/factory.py

@@ -27,14 +27,6 @@ def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv
     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",
-        "max_episode_steps": cfg.env.episode_length,
-        # "visualization_width": 384,
-        # "visualization_height": 384,
-    }
-
     package_name = f"gym_{cfg.env.name}"
 
     try:
@@ -46,12 +38,16 @@ def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv
         raise e
 
     gym_handle = f"{package_name}/{cfg.env.task}"
+    gym_kwgs = dict(cfg.env.get("gym", {}))
+
+    if cfg.env.get("episode_length"):
+        gym_kwgs["max_episode_steps"] = cfg.env.episode_length
 
     # batched version of the env that returns an observation of shape (b, c)
     env_cls = gym.vector.AsyncVectorEnv if cfg.eval.use_async_envs else gym.vector.SyncVectorEnv
     env = env_cls(
         [
-            lambda: gym.make(gym_handle, disable_env_checker=True, **kwargs)
+            lambda: gym.make(gym_handle, disable_env_checker=True, **gym_kwgs)
             for _ in range(n_envs if n_envs is not None else cfg.eval.batch_size)
         ]
     )

lerobot/common/logger.py

@@ -13,25 +13,33 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+"""Borrowed from https://github.com/fyhMer/fowm/blob/main/src/logger.py
+
 # TODO(rcadene, alexander-soare): clean this file
-"""Borrowed from https://github.com/fyhMer/fowm/blob/main/src/logger.py"""
+"""
 
 import logging
 import os
+import re
+from glob import glob
 from pathlib import Path
 
+import torch
 from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE
-from omegaconf import OmegaConf
+from omegaconf import DictConfig, OmegaConf
 from termcolor import colored
+from torch.optim import Optimizer
+from torch.optim.lr_scheduler import LRScheduler
 
 from lerobot.common.policies.policy_protocol import Policy
+from lerobot.common.utils.utils import get_global_random_state, set_global_random_state
 
 
 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):
+def cfg_to_group(cfg: DictConfig, return_list: bool = False) -> list[str] | str:
     """Return a group name for logging. Optionally returns group name as list."""
     lst = [
         f"policy:{cfg.policy.name}",
@@ -42,22 +50,54 @@ def cfg_to_group(cfg, return_list=False):
     return lst if return_list else "-".join(lst)
 
 
-class Logger:
-    """Primary logger object. Logs either locally or using wandb."""
+def get_wandb_run_id_from_filesystem(checkpoint_dir: Path) -> str:
+    # Get the WandB run ID.
+    paths = glob(str(checkpoint_dir / "../wandb/latest-run/run-*"))
+    if len(paths) != 1:
+        raise RuntimeError("Couldn't get the previous WandB run ID for run resumption.")
+    match = re.search(r"run-([^\.]+).wandb", paths[0].split("/")[-1])
+    if match is None:
+        raise RuntimeError("Couldn't get the previous WandB run ID for run resumption.")
+    wandb_run_id = match.groups(0)[0]
+    return wandb_run_id
 
-    def __init__(self, log_dir, job_name, cfg):
-        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 / "checkpoints"
-        self._buffer_dir = self._log_dir / "buffers"
-        self._save_model = cfg.training.save_model
-        self._disable_wandb_artifact = cfg.wandb.disable_artifact
-        self._save_buffer = cfg.training.get("save_buffer", False)
-        self._group = cfg_to_group(cfg)
-        self._seed = cfg.seed
+
+class Logger:
+    """Primary logger object. Logs either locally or using wandb.
+
+    The logger creates the following directory structure:
+
+    provided_log_dir
+    ├── .hydra  # hydra's configuration cache
+    ├── checkpoints
+    │   ├── specific_checkpoint_name
+    │   │   ├── pretrained_model  # Hugging Face pretrained model directory
+    │   │   │   ├── ...
+    │   │   └── training_state.pth  # optimizer, scheduler, and random states + training step
+    |   ├── another_specific_checkpoint_name
+    │   │   ├── ...
+    |   ├── ...
+    │   └── last  # a softlink to the last logged checkpoint
+    """
+
+    pretrained_model_dir_name = "pretrained_model"
+    training_state_file_name = "training_state.pth"
+
+    def __init__(self, cfg: DictConfig, log_dir: str, wandb_job_name: str | None = None):
+        """
+        Args:
+            log_dir: The directory to save all logs and training outputs to.
+            job_name: The WandB job name.
+        """
         self._cfg = cfg
-        self._eval = []
+        self.log_dir = Path(log_dir)
+        self.log_dir.mkdir(parents=True, exist_ok=True)
+        self.checkpoints_dir = self.get_checkpoints_dir(log_dir)
+        self.last_checkpoint_dir = self.get_last_checkpoint_dir(log_dir)
+        self.last_pretrained_model_dir = self.get_last_pretrained_model_dir(log_dir)
+
+        # Set up WandB.
+        self._group = cfg_to_group(cfg)
         project = cfg.get("wandb", {}).get("project")
         entity = cfg.get("wandb", {}).get("entity")
         enable_wandb = cfg.get("wandb", {}).get("enable", False)
@@ -69,71 +109,130 @@ class Logger:
             os.environ["WANDB_SILENT"] = "true"
             import wandb
 
+            wandb_run_id = None
+            if cfg.resume:
+                wandb_run_id = get_wandb_run_id_from_filesystem(self.checkpoints_dir)
+
             wandb.init(
+                id=wandb_run_id,
                 project=project,
                 entity=entity,
-                name=job_name,
+                name=wandb_job_name,
                 notes=cfg.get("wandb", {}).get("notes"),
-                # group=self._group,
                 tags=cfg_to_group(cfg, return_list=True),
-                dir=self._log_dir,
+                dir=log_dir,
                 config=OmegaConf.to_container(cfg, resolve=True),
                 # TODO(rcadene): try set to True
                 save_code=False,
                 # TODO(rcadene): split train and eval, and run async eval with job_type="eval"
                 job_type="train_eval",
-                # TODO(rcadene): add resume option
-                resume=None,
+                resume="must" if cfg.resume else None,
             )
             print(colored("Logs will be synced with wandb.", "blue", attrs=["bold"]))
             logging.info(f"Track this run --> {colored(wandb.run.get_url(), 'yellow', attrs=['bold'])}")
             self._wandb = wandb
 
-    def save_model(self, policy: Policy, identifier):
-        if self._save_model:
-            self._model_dir.mkdir(parents=True, exist_ok=True)
-            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 ":" or "/" in its name
-                artifact = self._wandb.Artifact(
-                    f"{self._group.replace(':', '_').replace('/', '_')}-{self._seed}-{identifier}",
-                    type="model",
-                )
-                artifact.add_file(save_dir / SAFETENSORS_SINGLE_FILE)
-                self._wandb.log_artifact(artifact)
+    @classmethod
+    def get_checkpoints_dir(cls, log_dir: str | Path) -> Path:
+        """Given the log directory, get the sub-directory in which checkpoints will be saved."""
+        return Path(log_dir) / "checkpoints"
 
-    def save_buffer(self, buffer, identifier):
-        self._buffer_dir.mkdir(parents=True, exist_ok=True)
-        fp = self._buffer_dir / f"{str(identifier)}.pkl"
-        buffer.save(fp)
-        if self._wandb and not self._disable_wandb_artifact:
+    @classmethod
+    def get_last_checkpoint_dir(cls, log_dir: str | Path) -> Path:
+        """Given the log directory, get the sub-directory in which the last checkpoint will be saved."""
+        return cls.get_checkpoints_dir(log_dir) / "last"
+
+    @classmethod
+    def get_last_pretrained_model_dir(cls, log_dir: str | Path) -> Path:
+        """
+        Given the log directory, get the sub-directory in which the last checkpoint's pretrained weights will
+        be saved.
+        """
+        return cls.get_last_checkpoint_dir(log_dir) / cls.pretrained_model_dir_name
+
+    def save_model(self, save_dir: Path, policy: Policy, wandb_artifact_name: str | None = None):
+        """Save the weights of the Policy model using PyTorchModelHubMixin.
+
+        The weights are saved in a folder called "pretrained_model" under the checkpoint directory.
+
+        Optionally also upload the model to WandB.
+        """
+        self.checkpoints_dir.mkdir(parents=True, exist_ok=True)
+        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._cfg.wandb.disable_artifact:
             # note wandb artifact does not accept ":" or "/" in its name
-            artifact = self._wandb.Artifact(
-                f"{self._group.replace(':', '_').replace('/', '_')}-{self._seed}-{identifier}",
-                type="buffer",
-            )
-            artifact.add_file(fp)
+            artifact = self._wandb.Artifact(wandb_artifact_name, type="model")
+            artifact.add_file(save_dir / SAFETENSORS_SINGLE_FILE)
             self._wandb.log_artifact(artifact)
+        if self.last_checkpoint_dir.exists():
+            os.remove(self.last_checkpoint_dir)
 
-    def finish(self, agent, buffer):
-        if self._save_model:
-            self.save_model(agent, identifier="final")
-        if self._save_buffer:
-            self.save_buffer(buffer, identifier="buffer")
-        if self._wandb:
-            self._wandb.finish()
+    def save_training_state(
+        self,
+        save_dir: Path,
+        train_step: int,
+        optimizer: Optimizer,
+        scheduler: LRScheduler | None,
+    ):
+        """Checkpoint the global training_step, optimizer state, scheduler state, and random state.
+
+        All of these are saved as "training_state.pth" under the checkpoint directory.
+        """
+        training_state = {
+            "step": train_step,
+            "optimizer": optimizer.state_dict(),
+            **get_global_random_state(),
+        }
+        if scheduler is not None:
+            training_state["scheduler"] = scheduler.state_dict()
+        torch.save(training_state, save_dir / self.training_state_file_name)
+
+    def save_checkpont(
+        self,
+        train_step: int,
+        policy: Policy,
+        optimizer: Optimizer,
+        scheduler: LRScheduler | None,
+        identifier: str,
+    ):
+        """Checkpoint the model weights and the training state."""
+        checkpoint_dir = self.checkpoints_dir / str(identifier)
+        wandb_artifact_name = (
+            None
+            if self._wandb is None
+            else f"{self._group.replace(':', '_').replace('/', '_')}-{self._cfg.seed}-{identifier}"
+        )
+        self.save_model(
+            checkpoint_dir / self.pretrained_model_dir_name, policy, wandb_artifact_name=wandb_artifact_name
+        )
+        self.save_training_state(checkpoint_dir, train_step, optimizer, scheduler)
+        os.symlink(checkpoint_dir.absolute(), self.last_checkpoint_dir)
+
+    def load_last_training_state(self, optimizer: Optimizer, scheduler: LRScheduler | None) -> int:
+        """
+        Given the last checkpoint in the logging directory, load the optimizer state, scheduler state, and
+        random state, and return the global training step.
+        """
+        training_state = torch.load(self.last_checkpoint_dir / self.training_state_file_name)
+        optimizer.load_state_dict(training_state["optimizer"])
+        if scheduler is not None:
+            scheduler.load_state_dict(training_state["scheduler"])
+        elif "scheduler" in training_state:
+            raise ValueError(
+                "The checkpoint contains a scheduler state_dict, but no LRScheduler was provided."
+            )
+        # Small hack to get the expected keys: use `get_global_random_state`.
+        set_global_random_state({k: training_state[k] for k in get_global_random_state()})
+        return training_state["step"]
 
     def log_dict(self, d, step, mode="train"):
         assert mode in {"train", "eval"}
+        # TODO(alexander-soare): Add local text log.
         if self._wandb is not None:
             for k, v in d.items():
                 if not isinstance(v, (int, float, str)):
-                    logging.warning(
-                        f'WandB logging of key "{k}" was ignored as its type is not handled by this wrapper.'
-                    )
                     continue
                 self._wandb.log({f"{mode}/{k}": v}, step=step)
 

lerobot/common/policies/act/configuration_act.py

@@ -25,6 +25,13 @@ class ACTConfig:
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
     Those are: `input_shapes` and 'output_shapes`.
 
+    Notes on the inputs and outputs:
+        - At least one key starting with "observation.image is required as an input.
+        - If there are multiple keys beginning with "observation.images." they are treated as multiple camera
+          views. Right now we only support all images having the same shape.
+        - May optionally work without an "observation.state" key for the proprioceptive robot state.
+        - "action" is required as an output key.
+
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
@@ -33,15 +40,15 @@ class ACTConfig:
             This should be no greater than the chunk size. For example, if the chunk size size 100, you may
             set this to 50. This would mean that the model predicts 100 steps worth of actions, runs 50 in the
             environment, and throws the other 50 out.
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-            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 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 doesn't include batch dimension or temporal dimension.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_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, `output_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

lerobot/common/policies/act/modeling_act.py

@@ -139,25 +139,26 @@ class ACTPolicy(nn.Module, PyTorchModelHubMixin):
         batch = self.normalize_targets(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()
+        bsize = actions_hat.shape[0]
+        l1_loss = F.l1_loss(batch["action"], actions_hat, reduction="none")
+        l1_loss = l1_loss * ~batch["action_is_pad"].unsqueeze(-1)
+        l1_loss = l1_loss.view(bsize, -1).mean(dim=1)
+
+        out_dict = {}
+        out_dict["l1_loss"] = l1_loss
 
-        loss_dict = {"l1_loss": l1_loss.item()}
         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.item()
-            loss_dict["loss"] = l1_loss + mean_kld * self.config.kl_weight
+            kld_loss = (-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1)
+            out_dict["loss"] = l1_loss + kld_loss * self.config.kl_weight
         else:
-            loss_dict["loss"] = l1_loss
+            out_dict["loss"] = l1_loss
 
-        return loss_dict
+        out_dict["action"] = self.unnormalize_outputs({"action": actions_hat})["action"]
+        return out_dict
 
 
 class ACT(nn.Module):
@@ -198,27 +199,31 @@ class ACT(nn.Module):
     def __init__(self, config: ACTConfig):
         super().__init__()
         self.config = config
-        # BERT style VAE encoder with input [cls, *joint_space_configuration, *action_sequence].
+        # BERT style VAE encoder with input tokens [cls, robot_state, *action_sequence].
         # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
+        self.use_input_state = "observation.state" in config.input_shapes
         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(
-                config.input_shapes["observation.state"][0], config.dim_model
-            )
+            if self.use_input_state:
+                self.vae_encoder_robot_state_input_proj = nn.Linear(
+                    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(
-                config.input_shapes["observation.state"][0], config.dim_model
+                config.output_shapes["action"][0], config.dim_model
             )
-            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(config.dim_model, self.latent_dim * 2)
-            # Fixed sinusoidal positional embedding the whole input to the VAE encoder. Unsqueeze for batch
+            self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, config.latent_dim * 2)
+            # Fixed sinusoidal positional embedding for the input to the VAE encoder. Unsqueeze for batch
             # dimension.
+            num_input_token_encoder = 1 + config.chunk_size
+            if self.use_input_state:
+                num_input_token_encoder += 1
             self.register_buffer(
                 "vae_encoder_pos_enc",
-                create_sinusoidal_pos_embedding(1 + 1 + config.chunk_size, config.dim_model).unsqueeze(0),
+                create_sinusoidal_pos_embedding(num_input_token_encoder, config.dim_model).unsqueeze(0),
             )
 
         # Backbone for image feature extraction.
@@ -238,15 +243,17 @@ class ACT(nn.Module):
 
         # 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(
-            config.input_shapes["observation.state"][0], config.dim_model
-        )
-        self.encoder_latent_input_proj = nn.Linear(self.latent_dim, config.dim_model)
+        if self.use_input_state:
+            self.encoder_robot_state_input_proj = nn.Linear(
+                config.input_shapes["observation.state"][0], config.dim_model
+            )
+        self.encoder_latent_input_proj = nn.Linear(config.latent_dim, config.dim_model)
         self.encoder_img_feat_input_proj = nn.Conv2d(
             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, config.dim_model)
+        num_input_token_decoder = 2 if self.use_input_state else 1
+        self.encoder_robot_and_latent_pos_embed = nn.Embedding(num_input_token_decoder, config.dim_model)
         self.encoder_cam_feat_pos_embed = ACTSinusoidalPositionEmbedding2d(config.dim_model // 2)
 
         # Transformer decoder.
@@ -258,6 +265,16 @@ class ACT(nn.Module):
 
         self._reset_parameters()
 
+        self.register_buffer(
+            "latent_sample",
+            torch.zeros(1, config.latent_dim, dtype=torch.float32),
+        )
+
+        self.register_buffer(
+            "decoder_in",
+            torch.zeros(config.chunk_size, 1, config.dim_model, dtype=torch.float32),
+        )
+
     def _reset_parameters(self):
         """Xavier-uniform initialization of the transformer parameters as in the original code."""
         for p in chain(self.encoder.parameters(), self.decoder.parameters()):
@@ -285,7 +302,7 @@ class ACT(nn.Module):
                 "action" in batch
             ), "actions must be provided when using the variational objective in training mode."
 
-        batch_size = batch["observation.state"].shape[0]
+        batch_size = batch["observation.images"].shape[0]
 
         # Prepare the latent for input to the transformer encoder.
         if self.config.use_vae and "action" in batch:
@@ -293,11 +310,16 @@ class ACT(nn.Module):
             cls_embed = einops.repeat(
                 self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
             )  # (B, 1, D)
-            robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"]).unsqueeze(
-                1
-            )  # (B, 1, D)
+            if self.use_input_state:
+                robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
+                robot_state_embed = robot_state_embed.unsqueeze(1)  # (B, 1, D)
             action_embed = self.vae_encoder_action_input_proj(batch["action"])  # (B, S, D)
-            vae_encoder_input = torch.cat([cls_embed, robot_state_embed, action_embed], axis=1)  # (B, S+2, D)
+
+            if self.use_input_state:
+                vae_encoder_input = [cls_embed, robot_state_embed, action_embed]  # (B, S+2, D)
+            else:
+                vae_encoder_input = [cls_embed, action_embed]
+            vae_encoder_input = torch.cat(vae_encoder_input, axis=1)
 
             # Prepare fixed positional embedding.
             # Note: detach() shouldn't be necessary but leaving it the same as the original code just in case.
@@ -308,27 +330,26 @@ class ACT(nn.Module):
                 vae_encoder_input.permute(1, 0, 2), pos_embed=pos_embed.permute(1, 0, 2)
             )[0]  # select the class token, with shape (B, D)
             latent_pdf_params = self.vae_encoder_latent_output_proj(cls_token_out)
-            mu = latent_pdf_params[:, : self.latent_dim]
+            mu = latent_pdf_params[:, : self.config.latent_dim]
             # This is 2log(sigma). Done this way to match the original implementation.
-            log_sigma_x2 = latent_pdf_params[:, self.latent_dim :]
+            log_sigma_x2 = latent_pdf_params[:, self.config.latent_dim :]
 
             # Sample the latent with the reparameterization trick.
             latent_sample = mu + log_sigma_x2.div(2).exp() * torch.randn_like(mu)
         else:
             # When not using the VAE encoder, we set the latent to be all zeros.
             mu = log_sigma_x2 = None
-            latent_sample = torch.zeros([batch_size, self.latent_dim], dtype=torch.float32).to(
-                batch["observation.state"].device
-            )
+            latent_sample = self.latent_sample
 
         # Prepare all other transformer encoder inputs.
         # Camera observation features and positional embeddings.
         all_cam_features = []
         all_cam_pos_embeds = []
         images = batch["observation.images"]
+
         for cam_index in range(images.shape[-4]):
             cam_features = self.backbone(images[:, cam_index])["feature_map"]
-            cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
+            cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features)
             cam_features = self.encoder_img_feat_input_proj(cam_features)  # (B, C, h, w)
             all_cam_features.append(cam_features)
             all_cam_pos_embeds.append(cam_pos_embed)
@@ -337,13 +358,15 @@ class ACT(nn.Module):
         cam_pos_embed = torch.cat(all_cam_pos_embeds, axis=-1)
 
         # Get positional embeddings for robot state and latent.
-        robot_state_embed = self.encoder_robot_state_input_proj(batch["observation.state"])  # (B, C)
+        if self.use_input_state:
+            robot_state_embed = self.encoder_robot_state_input_proj(batch["observation.state"])  # (B, C)
         latent_embed = self.encoder_latent_input_proj(latent_sample)  # (B, C)
 
         # Stack encoder input and positional embeddings moving to (S, B, C).
+        encoder_in_feats = [latent_embed, robot_state_embed] if self.use_input_state else [latent_embed]
         encoder_in = torch.cat(
             [
-                torch.stack([latent_embed, robot_state_embed], axis=0),
+                torch.stack(encoder_in_feats, axis=0),
                 einops.rearrange(encoder_in, "b c h w -> (h w) b c"),
             ]
         )
@@ -357,11 +380,7 @@ class ACT(nn.Module):
 
         # Forward pass through the transformer modules.
         encoder_out = self.encoder(encoder_in, pos_embed=pos_embed)
-        decoder_in = torch.zeros(
-            (self.config.chunk_size, batch_size, self.config.dim_model),
-            dtype=pos_embed.dtype,
-            device=pos_embed.device,
-        )
+        decoder_in = self.decoder_in
         decoder_out = self.decoder(
             decoder_in,
             encoder_out,
@@ -561,6 +580,10 @@ class ACTSinusoidalPositionEmbedding2d(nn.Module):
         self._eps = 1e-6
         # Inverse "common ratio" for the geometric progression in sinusoid frequencies.
         self._temperature = 10000
+        self.register_buffer(
+            "inverse_frequency", 
+            self._temperature ** (2 * (torch.arange(self.dimension, dtype=torch.float32) // 2) / self.dimension),
+        )
 
     def forward(self, x: Tensor) -> Tensor:
         """
@@ -572,8 +595,8 @@ class ACTSinusoidalPositionEmbedding2d(nn.Module):
         not_mask = torch.ones_like(x[0, :1])  # (1, H, W)
         # Note: These are like range(1, H+1) and range(1, W+1) respectively, but in most implementations
         # they would be range(0, H) and range(0, W). Keeping it at as is to match the original code.
-        y_range = not_mask.cumsum(1, dtype=torch.float32)
-        x_range = not_mask.cumsum(2, dtype=torch.float32)
+        y_range = not_mask.cumsum(1, dtype=x.dtype)
+        x_range = not_mask.cumsum(2, dtype=x.dtype)
 
         # "Normalize" the position index such that it ranges in [0, 2π].
         # Note: Adding epsilon on the denominator should not be needed as all values of y_embed and x_range
@@ -581,9 +604,7 @@ class ACTSinusoidalPositionEmbedding2d(nn.Module):
         y_range = y_range / (y_range[:, -1:, :] + self._eps) * self._two_pi
         x_range = x_range / (x_range[:, :, -1:] + self._eps) * self._two_pi
 
-        inverse_frequency = self._temperature ** (
-            2 * (torch.arange(self.dimension, dtype=torch.float32, device=x.device) // 2) / self.dimension
-        )
+        inverse_frequency = self.inverse_frequency
 
         x_range = x_range.unsqueeze(-1) / inverse_frequency  # (1, H, W, 1)
         y_range = y_range.unsqueeze(-1) / inverse_frequency  # (1, H, W, 1)

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -26,21 +26,26 @@ class DiffusionConfig:
     The parameters you will most likely need to change are the ones which depend on the environment / sensors.
     Those are: `input_shapes` and `output_shapes`.
 
+    Notes on the inputs and outputs:
+        - "observation.state" is required as an input key.
+        - A key starting with "observation.image is required as an input.
+        - "action" is required as an output key.
+
     Args:
         n_obs_steps: Number of environment steps worth of observations to pass to the policy (takes the
             current step and additional steps going back).
         horizon: Diffusion model action prediction size as detailed in `DiffusionPolicy.select_action`.
         n_action_steps: The number of action steps to run in the environment for one invocation of the policy.
             See `DiffusionPolicy.select_action` for more details.
-        input_shapes: A dictionary defining the shapes of the input data for the policy.
-            The key represents the input data name, and the value is a list indicating the dimensions
-            of the corresponding data. For example, "observation.image" refers to an input from
-            a camera with dimensions [3, 96, 96], indicating it has three color channels and 96x96 resolution.
-            Importantly, shapes doesnt 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.
+        input_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_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, `output_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
@@ -155,7 +160,7 @@ class DiffusionConfig:
                 f"{self.__class__.__name__} only handles one image for now. Got image keys {image_keys}."
             )
         image_key = next(iter(image_keys))
-        if (
+        if self.crop_shape is not None and (
             self.crop_shape[0] > self.input_shapes[image_key][1]
             or self.crop_shape[1] > self.input_shapes[image_key][2]
         ):

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -239,10 +239,8 @@ class DiffusionModel(nn.Module):
         global_cond = torch.cat([batch["observation.state"], img_features], dim=-1).flatten(start_dim=1)
 
         # run sampling
-        sample = self.conditional_sample(batch_size, global_cond=global_cond)
+        actions = self.conditional_sample(batch_size, global_cond=global_cond)
 
-        # `horizon` steps worth of actions (from the first observation).
-        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.config.n_action_steps
@@ -304,7 +302,11 @@ class DiffusionModel(nn.Module):
         loss = F.mse_loss(pred, target, reduction="none")
 
         # Mask loss wherever the action is padded with copies (edges of the dataset trajectory).
-        if self.config.do_mask_loss_for_padding and "action_is_pad" in batch:
+        if self.config.do_mask_loss_for_padding:
+            if "action_is_pad" not in batch:
+                raise ValueError(
+                    f"You need to provide 'action_is_pad' in the batch when {self.config.do_mask_loss_for_padding=}."
+                )
             in_episode_bound = ~batch["action_is_pad"]
             loss = loss * in_episode_bound.unsqueeze(-1)
 
@@ -423,11 +425,15 @@ class DiffusionRgbEncoder(nn.Module):
         # Set up pooling and final layers.
         # Use a dry run to get the feature map shape.
         # The dummy input should take the number of image channels from `config.input_shapes` and it should
-        # use the height and width from `config.crop_shape`.
+        # use the height and width from `config.crop_shape` if it is provided, otherwise it should use the
+        # height and width from `config.input_shapes`.
         image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
         assert len(image_keys) == 1
         image_key = image_keys[0]
-        dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *config.crop_shape))
+        dummy_input_h_w = (
+            config.crop_shape if config.crop_shape is not None else config.input_shapes[image_key][1:]
+        )
+        dummy_input = torch.zeros(size=(1, config.input_shapes[image_key][0], *dummy_input_h_w))
         with torch.inference_mode():
             dummy_feature_map = self.backbone(dummy_input)
         feature_map_shape = tuple(dummy_feature_map.shape[1:])

lerobot/common/policies/tdmpc/configuration_tdmpc.py

@@ -31,6 +31,15 @@ class TDMPCConfig:
         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_shapes: A dictionary defining the shapes of the input data for the policy. The key represents
+            the input data name, and the value is a list indicating the dimensions of the corresponding data.
+            For example, "observation.image" refers to an input from a camera with dimensions [3, 96, 96],
+            indicating it has three color channels and 96x96 resolution. Importantly, `input_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, `output_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

lerobot/common/utils/utils.py

@@ -19,7 +19,7 @@ import random
 from contextlib import contextmanager
 from datetime import datetime
 from pathlib import Path
-from typing import Generator
+from typing import Any, Generator
 
 import hydra
 import numpy as np
@@ -48,12 +48,38 @@ def get_safe_torch_device(cfg_device: str, log: bool = False) -> torch.device:
     return device
 
 
+def get_global_random_state() -> dict[str, Any]:
+    """Get the random state for `random`, `numpy`, and `torch`."""
+    random_state_dict = {
+        "random_state": random.getstate(),
+        "numpy_random_state": np.random.get_state(),
+        "torch_random_state": torch.random.get_rng_state(),
+    }
+    if torch.cuda.is_available():
+        random_state_dict["torch_cuda_random_state"] = torch.cuda.random.get_rng_state()
+    return random_state_dict
+
+
+def set_global_random_state(random_state_dict: dict[str, Any]):
+    """Set the random state for `random`, `numpy`, and `torch`.
+
+    Args:
+        random_state_dict: A dictionary of the form returned by `get_global_random_state`.
+    """
+    random.setstate(random_state_dict["random_state"])
+    np.random.set_state(random_state_dict["numpy_random_state"])
+    torch.random.set_rng_state(random_state_dict["torch_random_state"])
+    if torch.cuda.is_available():
+        torch.cuda.random.set_rng_state(random_state_dict["torch_cuda_random_state"])
+
+
 def set_global_seed(seed):
     """Set seed for reproducibility."""
     random.seed(seed)
     np.random.seed(seed)
     torch.manual_seed(seed)
-    torch.cuda.manual_seed_all(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
 
 
 @contextmanager
@@ -69,16 +95,10 @@ def seeded_context(seed: int) -> Generator[None, None, None]:
     c = random.random()  # produces yet another random number, but the same it would have if we never made `b`
     ```
     """
-    random_state = random.getstate()
-    np_random_state = np.random.get_state()
-    torch_random_state = torch.random.get_rng_state()
-    torch_cuda_random_state = torch.cuda.random.get_rng_state()
+    random_state_dict = get_global_random_state()
     set_global_seed(seed)
     yield None
-    random.setstate(random_state)
-    np.random.set_state(np_random_state)
-    torch.random.set_rng_state(torch_random_state)
-    torch.cuda.random.set_rng_state(torch_cuda_random_state)
+    set_global_random_state(random_state_dict)
 
 
 def init_logging():
@@ -100,13 +120,13 @@ def init_logging():
     logging.getLogger().addHandler(console_handler)
 
 
-def format_big_number(num):
+def format_big_number(num, precision=0):
     suffixes = ["", "K", "M", "B", "T", "Q"]
     divisor = 1000.0
 
     for suffix in suffixes:
         if abs(num) < divisor:
-            return f"{num:.0f}{suffix}"
+            return f"{num:.{precision}f}{suffix}"
         num /= divisor
 
     return num

lerobot/configs/default.yaml

@@ -5,18 +5,33 @@ defaults:
 
 hydra:
   run:
+    # Set `dir` to where you would like to save all of the run outputs. If you run another training session
+    # with the same value for `dir` its contents will be overwritten unless you set `resume` to true.
     dir: outputs/train/${now:%Y-%m-%d}/${now:%H-%M-%S}_${env.name}_${policy.name}_${hydra.job.name}
   job:
     name: default
 
+# Set `resume` to true to resume a previous run. In order for this to work, you will need to make sure
+# `hydra.run.dir` is the directory of an existing run with at least one checkpoint in it.
+# Note that when resuming a run, the default behavior is to use the configuration from the checkpoint,
+# regardless of what's provided with the training command at the time of resumption.
+resume: false
 device: cuda  # cpu
+# `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
+# automatic gradient scaling is used.
+use_amp: false
 # `seed` is used for training (eg: model initialization, dataset shuffling)
 # AND for the evaluation environments.
 seed: ???
+# You may provide a list of datasets here. `train.py` creates them all and concatenates them. Note: only data
+# keys common between the datasets are kept. Each dataset gets and additional transform that inserts the
+# "dataset_index" into the returned item. The index mapping is made according to the order in which the
+# datsets are provided.
 dataset_repo_id: lerobot/pusht
 
 training:
   offline_steps: ???
+  # NOTE: `online_steps` is not implemented yet. It's here as a placeholder.
   online_steps: ???
   online_steps_between_rollouts: ???
   online_sampling_ratio: 0.5
@@ -25,7 +40,9 @@ training:
   eval_freq: ???
   save_freq: ???
   log_freq: 250
-  save_model: true
+  save_checkpoint: true
+  num_workers: 4
+  batch_size: ???
 
 eval:
   n_episodes: 1
@@ -36,7 +53,7 @@ eval:
 
 wandb:
   enable: false
-  # Set to true to disable saving an artifact despite save_model == True
+  # Set to true to disable saving an artifact despite save_checkpoint == True
   disable_artifact: false
   project: lerobot
   notes: ""

lerobot/configs/env/aloha.yaml

@@ -5,10 +5,10 @@ fps: 50
 env:
   name: aloha
   task: AlohaInsertion-v0
-  from_pixels: True
-  pixels_only: False
-  image_size: [3, 480, 640]
-  episode_length: 400
-  fps: ${fps}
   state_dim: 14
   action_dim: 14
+  fps: ${fps}
+  episode_length: 400
+  gym:
+    obs_type: pixels_agent_pos
+    render_mode: rgb_array

lerobot/configs/env/dora.yaml

@@ -1,14 +0,0 @@
-# @package _global_
-
-fps: 30
-
-env:
-  name: dora
-  task: DoraAloha-v0
-  # from_pixels: True
-  # pixels_only: False
-  # image_size: [3, 480, 640]
-  episode_length: 400
-  # fps: ${fps}
-  # state_dim: 14
-  # action_dim: 14

lerobot/configs/env/dora_aloha_real.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+fps: 30
+
+env:
+  name: dora
+  task: DoraAloha-v0
+  state_dim: 14
+  action_dim: 14
+  fps: ${fps}
+  episode_length: 400
+  gym:
+    fps: ${fps}

lerobot/configs/env/dora_reachy2_real.yaml

@@ -0,0 +1,13 @@
+# @package _global_
+
+fps: 30
+
+env:
+  name: dora
+  task: DoraReachy2-v0
+  state_dim: 22
+  action_dim: 22
+  fps: ${fps}
+  episode_length: 400
+  gym:
+    fps: ${fps}

lerobot/configs/env/pusht.yaml

@@ -5,10 +5,13 @@ fps: 10
 env:
   name: pusht
   task: PushT-v0
-  from_pixels: True
-  pixels_only: False
   image_size: 96
-  episode_length: 300
-  fps: ${fps}
   state_dim: 2
   action_dim: 2
+  fps: ${fps}
+  episode_length: 300
+  gym:
+    obs_type: pixels_agent_pos
+    render_mode: rgb_array
+    visualization_width: 384
+    visualization_height: 384

lerobot/configs/env/xarm.yaml

@@ -5,10 +5,13 @@ fps: 15
 env:
   name: xarm
   task: XarmLift-v0
-  from_pixels: True
-  pixels_only: False
   image_size: 84
-  episode_length: 25
-  fps: ${fps}
   state_dim: 4
   action_dim: 4
+  fps: ${fps}
+  episode_length: 25
+  gym:
+    obs_type: pixels_agent_pos
+    render_mode: rgb_array
+    visualization_width: 384
+    visualization_height: 384

lerobot/configs/policy/act.yaml

@@ -15,7 +15,7 @@ training:
   eval_freq: 10000
   save_freq: 100000
   log_freq: 250
-  save_model: true
+  save_checkpoint: true
 
   batch_size: 8
   lr: 1e-5
@@ -25,7 +25,7 @@ training:
   online_steps_between_rollouts: 1
 
   delta_timestamps:
-    action: "[i / ${fps} for i in range(${policy.chunk_size})]"
+    action: "[i / ${fps} for i in range(1, ${policy.chunk_size} + 1)]"
 
 eval:
   n_episodes: 50

lerobot/configs/policy/act_reachy2_real.yaml

@@ -0,0 +1,97 @@
+# @package _global_
+
+# Use `act_real.yaml` to train on real-world Aloha/Aloha2 datasets.
+# Compared to `act.yaml`, it contains 4 cameras (i.e. cam_right_wrist, cam_left_wrist, images,
+# cam_low) instead of 1 camera (i.e. top). Also, `training.eval_freq` is set to -1. This config is used
+# to evaluate checkpoints at a certain frequency of training steps. When it is set to -1, it deactivates evaluation.
+# This is because real-world evaluation is done through [dora-lerobot](https://github.com/dora-rs/dora-lerobot).
+# Look at its README for more information on how to evaluate a checkpoint in the real-world.
+#
+# Example of usage for training:
+# ```bash
+# python lerobot/scripts/train.py \
+#   policy=act_real \
+#   env=dora_aloha_real
+# ```
+
+seed: 1000
+dataset_repo_id: cadene/reachy2_teleop_remi
+
+override_dataset_stats:
+  observation.images.cam_trunk:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+training:
+  offline_steps: 80000
+  online_steps: 0
+  eval_freq: -1
+  save_freq: 10000
+  log_freq: 100
+  save_checkpoint: true
+
+  batch_size: 8
+  lr: 1e-5
+  lr_backbone: 1e-5
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
+
+  delta_timestamps:
+    action: "[i / ${fps} for i in range(1, ${policy.chunk_size} + 1)]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.cam_trunk: [3, 800, 1280]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.cam_trunk: mean_std
+    observation.state: mean_std
+  output_normalization_modes:
+    action: mean_std
+
+  # Architecture.
+  # Vision backbone.
+  vision_backbone: resnet18
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
+  replace_final_stride_with_dilation: false
+  # Transformer layers.
+  pre_norm: false
+  dim_model: 512
+  n_heads: 8
+  dim_feedforward: 3200
+  feedforward_activation: relu
+  n_encoder_layers: 4
+  # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
+  # that means only the first layer is used. Here we match the original implementation by setting this to 1.
+  # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
+  n_decoder_layers: 1
+  # VAE.
+  use_vae: true
+  latent_dim: 32
+  n_vae_encoder_layers: 4
+
+  # Inference.
+  temporal_ensemble_momentum: null
+
+  # Training and loss computation.
+  dropout: 0.1
+  kl_weight: 10.0

lerobot/configs/policy/act_real.yaml

@@ -0,0 +1,115 @@
+# @package _global_
+
+# Use `act_real.yaml` to train on real-world Aloha/Aloha2 datasets.
+# Compared to `act.yaml`, it contains 4 cameras (i.e. cam_right_wrist, cam_left_wrist, images,
+# cam_low) instead of 1 camera (i.e. top). Also, `training.eval_freq` is set to -1. This config is used
+# to evaluate checkpoints at a certain frequency of training steps. When it is set to -1, it deactivates evaluation.
+# This is because real-world evaluation is done through [dora-lerobot](https://github.com/dora-rs/dora-lerobot).
+# Look at its README for more information on how to evaluate a checkpoint in the real-world.
+#
+# Example of usage for training:
+# ```bash
+# python lerobot/scripts/train.py \
+#   policy=act_real \
+#   env=dora_aloha_real
+# ```
+
+seed: 1000
+dataset_repo_id: lerobot/aloha_static_vinh_cup
+
+override_dataset_stats:
+  observation.images.cam_right_wrist:
+    # 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)
+  observation.images.cam_left_wrist:
+    # 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)
+  observation.images.cam_high:
+    # 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)
+  observation.images.cam_low:
+    # stats from imagenet, since we use a pretrained vision model
+    mean: [[[0.485]], [[0.456]], [[0.406]]]  # (c,1,1)
+    std: [[[0.229]], [[0.224]], [[0.225]]]  # (c,1,1)
+
+training:
+  offline_steps: 80000
+  online_steps: 0
+  eval_freq: -1
+  save_freq: 10000
+  log_freq: 100
+  save_checkpoint: true
+
+  batch_size: 8
+  lr: 1e-5
+  lr_backbone: 1e-5
+  weight_decay: 1e-4
+  grad_clip_norm: 10
+  online_steps_between_rollouts: 1
+
+  delta_timestamps:
+    action: "[i / ${fps} for i in range(1, ${policy.chunk_size} + 1)]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    # TODO(rcadene, alexander-soare): add variables for height and width from the dataset/env?
+    observation.images.cam_right_wrist: [3, 480, 640]
+    observation.images.cam_left_wrist: [3, 480, 640]
+    observation.images.cam_high: [3, 480, 640]
+    observation.images.cam_low: [3, 480, 640]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.images.cam_right_wrist: mean_std
+    observation.images.cam_left_wrist: mean_std
+    observation.images.cam_high: mean_std
+    observation.images.cam_low: mean_std
+    observation.state: mean_std
+  output_normalization_modes:
+    action: mean_std
+
+  # Architecture.
+  # Vision backbone.
+  vision_backbone: resnet18
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
+  replace_final_stride_with_dilation: false
+  # Transformer layers.
+  pre_norm: false
+  dim_model: 512
+  n_heads: 8
+  dim_feedforward: 3200
+  feedforward_activation: relu
+  n_encoder_layers: 4
+  # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
+  # that means only the first layer is used. Here we match the original implementation by setting this to 1.
+  # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
+  n_decoder_layers: 1
+  # VAE.
+  use_vae: true
+  latent_dim: 32
+  n_vae_encoder_layers: 4
+
+  # Inference.
+  temporal_ensemble_momentum: null
+
+  # Training and loss computation.
+  dropout: 0.1
+  kl_weight: 10.0

lerobot/configs/policy/act_real_no_state.yaml

@@ -1,7 +1,19 @@
 # @package _global_
 
+# Use `act_real_no_state.yaml` to train on real-world Aloha/Aloha2 datasets when cameras are moving (e.g. wrist cameras)
+# Compared to `act_real.yaml`, it is camera only and does not use the state as input which is vector of robot joint positions.
+# We validated experimentaly that not using state reaches better success rate. Our hypothesis is that `act_real.yaml` might
+# overfits to the state, because the images are more complex to learn from since they are moving.
+#
+# Example of usage for training:
+# ```bash
+# python lerobot/scripts/train.py \
+#   policy=act_real_no_state \
+#   env=dora_aloha_real
+# ```
+
 seed: 1000
-dataset_repo_id: cadene/aloha_v2_static_dora_test
+dataset_repo_id: lerobot/aloha_static_vinh_cup
 
 override_dataset_stats:
   observation.images.cam_right_wrist:
@@ -24,10 +36,10 @@ override_dataset_stats:
 training:
   offline_steps: 80000
   online_steps: 0
-  eval_freq: 99999999999999
-  save_freq: 1000
+  eval_freq: -1
+  save_freq: 10000
   log_freq: 100
-  save_model: true
+  save_checkpoint: true
 
   batch_size: 8
   lr: 1e-5
@@ -37,7 +49,7 @@ training:
   online_steps_between_rollouts: 1
 
   delta_timestamps:
-    action: "[i / ${fps} for i in range(${policy.chunk_size})]"
+    action: "[i / ${fps} for i in range(1, ${policy.chunk_size} + 1)]"
 
 eval:
   n_episodes: 50
@@ -58,7 +70,6 @@ policy:
     observation.images.cam_left_wrist: [3, 480, 640]
     observation.images.cam_high: [3, 480, 640]
     observation.images.cam_low: [3, 480, 640]
-    observation.state: ["${env.state_dim}"]
   output_shapes:
     action: ["${env.action_dim}"]
 
@@ -68,7 +79,6 @@ policy:
     observation.images.cam_left_wrist: mean_std
     observation.images.cam_high: mean_std
     observation.images.cam_low: mean_std
-    observation.state: mean_std
   output_normalization_modes:
     action: mean_std
 

lerobot/configs/policy/diffusion.yaml

@@ -27,7 +27,7 @@ training:
   eval_freq: 5000
   save_freq: 5000
   log_freq: 250
-  save_model: true
+  save_checkpoint: true
 
   batch_size: 64
   grad_clip_norm: 10
@@ -44,6 +44,10 @@ training:
     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})]"
 
+  # The original implementation doesn't sample frames for the last 7 steps,
+  # which avoids excessive padding and leads to improved training results.
+  drop_n_last_frames: 7  # ${policy.horizon} - ${policy.n_action_steps} - ${policy.n_obs_steps} + 1
+
 eval:
   n_episodes: 50
   batch_size: 50

lerobot/configs/policy/tdmpc.yaml

@@ -5,7 +5,8 @@ dataset_repo_id: lerobot/xarm_lift_medium
 
 training:
   offline_steps: 25000
-  online_steps: 25000
+  # TODO(alexander-soare): uncomment when online training gets reinstated
+  online_steps: 0  # 25000 not implemented yet
   eval_freq: 5000
   online_steps_between_rollouts: 1
   online_sampling_ratio: 0.5

lerobot/scripts/eval.py

@@ -28,7 +28,7 @@ OR, you want to evaluate a model checkpoint from the LeRobot training script for
 
 ```
 python lerobot/scripts/eval.py \
-    -p outputs/train/diffusion_pusht/checkpoints/005000 \
+    -p outputs/train/diffusion_pusht/checkpoints/005000/pretrained_model \
     eval.n_episodes=10
 ```
 
@@ -46,6 +46,7 @@ import json
 import logging
 import threading
 import time
+from contextlib import nullcontext
 from copy import deepcopy
 from datetime import datetime as dt
 from pathlib import Path
@@ -520,7 +521,7 @@ def eval(
         raise NotImplementedError()
 
     # Check device is available
-    get_safe_torch_device(hydra_cfg.device, log=True)
+    device = get_safe_torch_device(hydra_cfg.device, log=True)
 
     torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
@@ -539,16 +540,17 @@ def eval(
         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",
-        start_seed=hydra_cfg.seed,
-        enable_progbar=True,
-        enable_inner_progbar=True,
-    )
+    with torch.no_grad(), torch.autocast(device_type=device.type) if hydra_cfg.use_amp else nullcontext():
+        info = eval_policy(
+            env,
+            policy,
+            hydra_cfg.eval.n_episodes,
+            max_episodes_rendered=10,
+            video_dir=Path(out_dir) / "eval",
+            start_seed=hydra_cfg.seed,
+            enable_progbar=True,
+            enable_inner_progbar=True,
+        )
     print(info["aggregated"])
 
     # Save info

lerobot/scripts/push_dataset_to_hub.py

@@ -71,9 +71,9 @@ import torch
 from huggingface_hub import HfApi
 from safetensors.torch import save_file
 
+from lerobot.common.datasets.compute_stats import compute_stats
 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.compute_stats import compute_stats
 from lerobot.common.datasets.utils import flatten_dict
 
 
@@ -86,6 +86,8 @@ def get_from_raw_to_lerobot_format_fn(raw_format):
         from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import from_raw_to_lerobot_format
     elif raw_format == "aloha_dora":
         from lerobot.common.datasets.push_dataset_to_hub.aloha_dora_format import from_raw_to_lerobot_format
+    elif raw_format == "reachy2_hdf5":
+        from lerobot.common.datasets.push_dataset_to_hub.reachy2_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:
@@ -144,8 +146,7 @@ def push_videos_to_hub(repo_id, videos_dir, revision):
 
 
 def push_dataset_to_hub(
-    input_data_dir: Path,
-    output_data_dir: Path,
+    data_dir: Path,
     dataset_id: str,
     raw_format: str | None,
     community_id: str,
@@ -162,33 +163,34 @@ def push_dataset_to_hub(
 ):
     repo_id = f"{community_id}/{dataset_id}"
 
-    meta_data_dir = output_data_dir / "meta_data"
-    videos_dir = output_data_dir / "videos"
+    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"
 
     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 output_data_dir.exists():
-        shutil.rmtree(output_data_dir)
+    if out_dir.exists():
+        shutil.rmtree(out_dir)
 
     if tests_out_dir.exists() and save_tests_to_disk:
         shutil.rmtree(tests_out_dir)
 
-    if not input_data_dir.exists():
-        download_raw(input_data_dir, dataset_id)
+    if not raw_dir.exists():
+        download_raw(raw_dir, dataset_id)
 
     if raw_format is None:
         # TODO(rcadene, adilzouitine): implement auto_find_raw_format
         raise NotImplementedError()
-        # raw_format = auto_find_raw_format(input_data_dir)
+        # raw_format = auto_find_raw_format(raw_dir)
 
     from_raw_to_lerobot_format = get_from_raw_to_lerobot_format_fn(raw_format)
 
     # convert dataset from original raw format to LeRobot format
-    hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(
-        input_data_dir, output_data_dir, fps, video, debug
-    )
+    hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(raw_dir, out_dir, fps, video, debug)
 
     lerobot_dataset = LeRobotDataset.from_preloaded(
         repo_id=repo_id,
@@ -202,7 +204,7 @@ def push_dataset_to_hub(
 
     if save_to_disk:
         hf_dataset = hf_dataset.with_format(None)  # to remove transforms that cant be saved
-        hf_dataset.save_to_disk(str(output_data_dir / "train"))
+        hf_dataset.save_to_disk(str(out_dir / "train"))
 
     if not dry_run or save_to_disk:
         # mandatory for upload
@@ -236,25 +238,19 @@ def push_dataset_to_hub(
             fname = f"{key}_episode_{episode_index:06d}.mp4"
             shutil.copy(videos_dir / fname, tests_videos_dir / fname)
 
-    if not save_to_disk and output_data_dir.exists():
+    if not save_to_disk and out_dir.exists():
         # remove possible temporary files remaining in the output directory
-        shutil.rmtree(output_data_dir)
+        shutil.rmtree(out_dir)
 
 
 def main():
     parser = argparse.ArgumentParser()
 
     parser.add_argument(
-        "--input-data-dir",
+        "--data-dir",
         type=Path,
         required=True,
-        help="Directory containing input raw datasets (e.g. `data/aloha_mobile_chair_raw` or `data/pusht_raw`).",
-    )
-    parser.add_argument(
-        "--output-data-dir",
-        type=Path,
-        required=True,
-        help="Root directory containing output dataset (e.g. `data/lerobot/aloha_mobile_chair` or `data/lerobot/pusht`).",
+        help="Root directory containing datasets (e.g. `data` or `tmp/data` or `/tmp/lerobot/data`).",
     )
     parser.add_argument(
         "--dataset-id",

lerobot/scripts/train.py

@@ -15,25 +15,30 @@
 # limitations under the License.
 import logging
 import time
-from copy import deepcopy
+from contextlib import nullcontext
 from pathlib import Path
+from pprint import pformat
 
-import datasets
 import hydra
 import torch
-from datasets import concatenate_datasets
-from datasets.utils import disable_progress_bars, enable_progress_bars
-from omegaconf import DictConfig
+from deepdiff import DeepDiff
+from omegaconf import DictConfig, OmegaConf
+from termcolor import colored
+from torch.cuda.amp import GradScaler
 
-from lerobot.common.datasets.factory import make_dataset
+from lerobot.common.datasets.factory import make_dataset, resolve_delta_timestamps
+from lerobot.common.datasets.lerobot_dataset import MultiLeRobotDataset
+from lerobot.common.datasets.sampler import EpisodeAwareSampler
 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.policies.utils import get_device_from_parameters
 from lerobot.common.utils.utils import (
     format_big_number,
     get_safe_torch_device,
+    init_hydra_config,
     init_logging,
     set_global_seed,
 )
@@ -69,7 +74,6 @@ def make_optimizer_and_scheduler(cfg, policy):
             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(
@@ -87,21 +91,40 @@ def make_optimizer_and_scheduler(cfg, policy):
     return optimizer, lr_scheduler
 
 
-def update_policy(policy, batch, optimizer, grad_clip_norm, lr_scheduler=None):
+def update_policy(
+    policy,
+    batch,
+    optimizer,
+    grad_clip_norm,
+    grad_scaler: GradScaler,
+    lr_scheduler=None,
+    use_amp: bool = False,
+):
     """Returns a dictionary of items for logging."""
-    start_time = time.time()
+    start_time = time.perf_counter()
+    device = get_device_from_parameters(policy)
     policy.train()
-    output_dict = policy.forward(batch)
-    # TODO(rcadene): policy.unnormalize_outputs(out_dict)
-    loss = output_dict["loss"]
-    loss.backward()
+    with torch.autocast(device_type=device.type) if use_amp else nullcontext():
+        output_dict = policy.forward(batch)
+        # TODO(rcadene): policy.unnormalize_outputs(out_dict)
+        loss = output_dict["loss"].mean()
+    grad_scaler.scale(loss).backward()
+
+    # Unscale the graident of the optimzer's assigned params in-place **prior to gradient clipping**.
+    grad_scaler.unscale_(optimizer)
+
     grad_norm = torch.nn.utils.clip_grad_norm_(
         policy.parameters(),
         grad_clip_norm,
         error_if_nonfinite=False,
     )
 
-    optimizer.step()
+    # Optimizer's gradients are already unscaled, so scaler.step does not unscale them,
+    # although it still skips optimizer.step() if the gradients contain infs or NaNs.
+    grad_scaler.step(optimizer)
+    # Updates the scale for next iteration.
+    grad_scaler.update()
+
     optimizer.zero_grad()
 
     if lr_scheduler is not None:
@@ -115,31 +138,13 @@ def update_policy(policy, batch, optimizer, grad_clip_norm, lr_scheduler=None):
         "loss": loss.item(),
         "grad_norm": float(grad_norm),
         "lr": optimizer.param_groups[0]["lr"],
-        "update_s": time.time() - start_time,
+        "update_s": time.perf_counter() - start_time,
         **{k: v for k, v in output_dict.items() if k != "loss"},
     }
 
     return info
 
 
-@hydra.main(version_base="1.2", config_name="default", config_path="../configs")
-def train_cli(cfg: dict):
-    train(
-        cfg,
-        out_dir=hydra.core.hydra_config.HydraConfig.get().run.dir,
-        job_name=hydra.core.hydra_config.HydraConfig.get().job.name,
-    )
-
-
-def train_notebook(out_dir=None, job_name=None, config_name="default", config_path="../configs"):
-    from hydra import compose, initialize
-
-    hydra.core.global_hydra.GlobalHydra.instance().clear()
-    initialize(config_path=config_path)
-    cfg = compose(config_name=config_name)
-    train(cfg, out_dir=out_dir, job_name=job_name)
-
-
 def log_train_info(logger: Logger, info, step, cfg, dataset, is_offline):
     loss = info["loss"]
     grad_norm = info["grad_norm"]
@@ -211,103 +216,6 @@ def log_eval_info(logger, info, step, cfg, dataset, is_offline):
     logger.log_dict(info, step, mode="eval")
 
 
-def calculate_online_sample_weight(n_off: int, n_on: int, pc_on: float):
-    """
-    Calculate the sampling weight to be assigned to samples so that a specified percentage of the batch comes from online dataset (on average).
-
-    Parameters:
-    - n_off (int): Number of offline samples, each with a sampling weight of 1.
-    - n_on (int): Number of online samples.
-    - pc_on (float): Desired percentage of online samples in decimal form (e.g., 50% as 0.5).
-
-    The total weight of offline samples is n_off * 1.0.
-    The total weight of offline samples is n_on * w.
-    The total combined weight of all samples is n_off + n_on * w.
-    The fraction of the weight that is online is n_on * w / (n_off + n_on * w).
-    We want this fraction to equal pc_on, so we set up the equation n_on * w / (n_off + n_on * w) = pc_on.
-    The solution is w = - (n_off * pc_on) / (n_on * (pc_on - 1))
-    """
-    assert 0.0 <= pc_on <= 1.0
-    return -(n_off * pc_on) / (n_on * (pc_on - 1))
-
-
-def add_episodes_inplace(
-    online_dataset: torch.utils.data.Dataset,
-    concat_dataset: torch.utils.data.ConcatDataset,
-    sampler: torch.utils.data.WeightedRandomSampler,
-    hf_dataset: datasets.Dataset,
-    episode_data_index: dict[str, torch.Tensor],
-    pc_online_samples: float,
-):
-    """
-    Modifies the online_dataset, concat_dataset, and sampler in place by integrating
-    new episodes from hf_dataset into the online_dataset, updating the concatenated
-    dataset's structure and adjusting the sampling strategy based on the specified
-    percentage of online samples.
-
-    Parameters:
-    - online_dataset (torch.utils.data.Dataset): The existing online dataset to be updated.
-    - concat_dataset (torch.utils.data.ConcatDataset): The concatenated dataset that combines
-      offline and online datasets, used for sampling purposes.
-    - sampler (torch.utils.data.WeightedRandomSampler): A sampler that will be updated to
-      reflect changes in the dataset sizes and specified sampling weights.
-    - hf_dataset (datasets.Dataset): A Hugging Face dataset containing the new episodes to be added.
-    - episode_data_index (dict): A dictionary containing two keys ("from" and "to") associated to dataset indices.
-      They indicate the start index and end index of each episode in the dataset.
-    - pc_online_samples (float): The target percentage of samples that should come from
-      the online dataset during sampling operations.
-
-    Raises:
-    - AssertionError: If the first episode_id or index in hf_dataset is not 0
-    """
-    first_episode_idx = hf_dataset.select_columns("episode_index")[0]["episode_index"].item()
-    last_episode_idx = hf_dataset.select_columns("episode_index")[-1]["episode_index"].item()
-    first_index = hf_dataset.select_columns("index")[0]["index"].item()
-    last_index = hf_dataset.select_columns("index")[-1]["index"].item()
-    # sanity check
-    assert first_episode_idx == 0, f"{first_episode_idx=} is not 0"
-    assert first_index == 0, f"{first_index=} is not 0"
-    assert first_index == episode_data_index["from"][first_episode_idx].item()
-    assert last_index == episode_data_index["to"][last_episode_idx].item() - 1
-
-    if len(online_dataset) == 0:
-        # initialize online dataset
-        online_dataset.hf_dataset = hf_dataset
-        online_dataset.episode_data_index = episode_data_index
-    else:
-        # get the starting indices of the new episodes and frames to be added
-        start_episode_idx = last_episode_idx + 1
-        start_index = last_index + 1
-
-        def shift_indices(episode_index, index):
-            # note: we dont shift "frame_index" since it represents the index of the frame in the episode it belongs to
-            example = {"episode_index": episode_index + start_episode_idx, "index": index + start_index}
-            return example
-
-        disable_progress_bars()  # map has a tqdm progress bar
-        hf_dataset = hf_dataset.map(shift_indices, input_columns=["episode_index", "index"])
-        enable_progress_bars()
-
-        episode_data_index["from"] += start_index
-        episode_data_index["to"] += start_index
-
-        # extend online dataset
-        online_dataset.hf_dataset = concatenate_datasets([online_dataset.hf_dataset, hf_dataset])
-
-    # update the concatenated dataset length used during sampling
-    concat_dataset.cumulative_sizes = concat_dataset.cumsum(concat_dataset.datasets)
-
-    # update the sampling weights for each frame so that online frames get sampled a certain percentage of times
-    len_online = len(online_dataset)
-    len_offline = len(concat_dataset) - len_online
-    weight_offline = 1.0
-    weight_online = calculate_online_sample_weight(len_offline, len_online, pc_online_samples)
-    sampler.weights = torch.tensor([weight_offline] * len_offline + [weight_online] * len(online_dataset))
-
-    # update the total number of samples used during sampling
-    sampler.num_samples = len(concat_dataset)
-
-
 def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = None):
     if out_dir is None:
         raise NotImplementedError()
@@ -316,35 +224,96 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
     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")
+    # If we are resuming a run, we need to check that a checkpoint exists in the log directory, and we need
+    # to check for any differences between the provided config and the checkpoint's config.
+    if cfg.resume:
+        if not Logger.get_last_checkpoint_dir(out_dir).exists():
+            raise RuntimeError(
+                "You have set resume=True, but there is no model checkpoint in "
+                f"{Logger.get_last_checkpoint_dir(out_dir)}"
+            )
+        checkpoint_cfg_path = str(Logger.get_last_pretrained_model_dir(out_dir) / "config.yaml")
+        logging.info(
+            colored(
+                "You have set resume=True, indicating that you wish to resume a run",
+                color="yellow",
+                attrs=["bold"],
+            )
+        )
+        # Get the configuration file from the last checkpoint.
+        checkpoint_cfg = init_hydra_config(checkpoint_cfg_path)
+        # Check for differences between the checkpoint configuration and provided configuration.
+        # Hack to resolve the delta_timestamps ahead of time in order to properly diff.
+        resolve_delta_timestamps(cfg)
+        diff = DeepDiff(OmegaConf.to_container(checkpoint_cfg), OmegaConf.to_container(cfg))
+        # Ignore the `resume` and parameters.
+        if "values_changed" in diff and "root['resume']" in diff["values_changed"]:
+            del diff["values_changed"]["root['resume']"]
+        # Log a warning about differences between the checkpoint configuration and the provided
+        # configuration.
+        if len(diff) > 0:
+            logging.warning(
+                "At least one difference was detected between the checkpoint configuration and "
+                f"the provided configuration: \n{pformat(diff)}\nNote that the checkpoint configuration "
+                "takes precedence.",
+            )
+        # Use the checkpoint config instead of the provided config (but keep `resume` parameter).
+        cfg = checkpoint_cfg
+        cfg.resume = True
+    elif Logger.get_last_checkpoint_dir(out_dir).exists():
+        raise RuntimeError(
+            f"The configured output directory {Logger.get_last_checkpoint_dir(out_dir)} already exists."
+        )
+
+    # log metrics to terminal and wandb
+    logger = Logger(cfg, out_dir, wandb_job_name=job_name)
+
+    if cfg.training.online_steps > 0:
+        raise NotImplementedError("Online training is not implemented yet.")
+
+    set_global_seed(cfg.seed)
 
     # Check device is available
-    get_safe_torch_device(cfg.device, log=True)
+    device = get_safe_torch_device(cfg.device, log=True)
 
     torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
-    set_global_seed(cfg.seed)
 
     logging.info("make_dataset")
     offline_dataset = make_dataset(cfg)
+    if isinstance(offline_dataset, MultiLeRobotDataset):
+        logging.info(
+            "Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
+            f"{pformat(offline_dataset.repo_id_to_index , indent=2)}"
+        )
 
-    logging.info("make_env")
-    eval_env = make_env(cfg)
+    # Create environment used for evaluating checkpoints during training on simulation data.
+    # On real-world data, no need to create an environment as evaluations are done outside train.py,
+    # using the eval.py instead, with gym_dora environment and dora-rs.
+    if cfg.training.eval_freq > 0:
+        logging.info("make_env")
+        eval_env = make_env(cfg)
 
     logging.info("make_policy")
-    policy = make_policy(hydra_cfg=cfg, dataset_stats=offline_dataset.stats)
+    policy = make_policy(
+        hydra_cfg=cfg,
+        dataset_stats=offline_dataset.stats if not cfg.resume else None,
+        pretrained_policy_name_or_path=str(logger.last_pretrained_model_dir) if cfg.resume else None,
+    )
 
     # Create optimizer and scheduler
     # Temporary hack to move optimizer out of policy
     optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
+    grad_scaler = GradScaler(enabled=cfg.use_amp)
+
+    step = 0  # number of policy updates (forward + backward + optim)
+
+    if cfg.resume:
+        step = logger.load_last_training_state(optimizer, lr_scheduler)
 
     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())
 
-    # log metrics to terminal and wandb
-    logger = Logger(out_dir, job_name, cfg)
-
     log_output_dir(out_dir)
     logging.info(f"{cfg.env.task=}")
     logging.info(f"{cfg.training.offline_steps=} ({format_big_number(cfg.training.offline_steps)})")
@@ -356,27 +325,31 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
     # Note: this helper will be used in offline and online training loops.
     def evaluate_and_checkpoint_if_needed(step):
-        if step % cfg.training.eval_freq == 0:
+        if cfg.training.eval_freq > 0 and step % cfg.training.eval_freq == 0:
             logging.info(f"Eval policy at step {step}")
-            eval_info = eval_policy(
-                eval_env,
-                policy,
-                cfg.eval.n_episodes,
-                video_dir=Path(out_dir) / "eval",
-                max_episodes_rendered=4,
-                start_seed=cfg.seed,
-            )
-            log_eval_info(logger, eval_info["aggregated"], step, cfg, offline_dataset, is_offline)
+            with torch.no_grad(), torch.autocast(device_type=device.type) if cfg.use_amp else nullcontext():
+                eval_info = eval_policy(
+                    eval_env,
+                    policy,
+                    cfg.eval.n_episodes,
+                    video_dir=Path(out_dir) / "eval",
+                    max_episodes_rendered=4,
+                    start_seed=cfg.seed,
+                )
+            log_eval_info(logger, eval_info["aggregated"], step, cfg, offline_dataset, is_offline=True)
             if cfg.wandb.enable:
                 logger.log_video(eval_info["video_paths"][0], step, mode="eval")
             logging.info("Resume training")
 
-        if cfg.training.save_model and step % cfg.training.save_freq == 0:
+        if cfg.training.save_checkpoint and step % cfg.training.save_freq == 0:
             logging.info(f"Checkpoint policy after step {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(
+            logger.save_checkpont(
+                step,
                 policy,
+                optimizer,
+                lr_scheduler,
                 identifier=str(step).zfill(
                     max(6, len(str(cfg.training.offline_steps + cfg.training.online_steps)))
                 ),
@@ -384,32 +357,48 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
             logging.info("Resume training")
 
     # create dataloader for offline training
+    if cfg.training.get("drop_n_last_frames"):
+        shuffle = False
+        sampler = EpisodeAwareSampler(
+            offline_dataset.episode_data_index,
+            drop_n_last_frames=cfg.training.drop_n_last_frames,
+            shuffle=True,
+        )
+    else:
+        shuffle = True
+        sampler = None
     dataloader = torch.utils.data.DataLoader(
         offline_dataset,
-        num_workers=4,
+        num_workers=cfg.training.num_workers,
         batch_size=cfg.training.batch_size,
-        shuffle=True,
-        pin_memory=cfg.device != "cpu",
+        shuffle=shuffle,
+        sampler=sampler,
+        pin_memory=device.type != "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.training.offline_steps):
-        if offline_step == 0:
+    for _ in range(step, cfg.training.offline_steps):
+        if step == 0:
             logging.info("Start offline training on a fixed dataset")
         batch = next(dl_iter)
 
         for key in batch:
-            batch[key] = batch[key].to(cfg.device, non_blocking=True)
+            batch[key] = batch[key].to(device, non_blocking=True)
 
-        train_info = update_policy(policy, batch, optimizer, cfg.training.grad_clip_norm, lr_scheduler)
+        train_info = update_policy(
+            policy,
+            batch,
+            optimizer,
+            cfg.training.grad_clip_norm,
+            grad_scaler=grad_scaler,
+            lr_scheduler=lr_scheduler,
+            use_amp=cfg.use_amp,
+        )
 
-        # TODO(rcadene): is it ok if step_t=0 = 0 and not 1 as previously done?
         if step % cfg.training.log_freq == 0:
-            log_train_info(logger, train_info, step, cfg, offline_dataset, is_offline)
+            log_train_info(logger, train_info, step, cfg, offline_dataset, is_offline=True)
 
         # Note: evaluate_and_checkpoint_if_needed happens **after** the `step`th training update has completed,
         # so we pass in step + 1.
@@ -417,79 +406,27 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
         step += 1
 
-    # create an env dedicated to online episodes collection from policy rollout
-    online_training_env = make_env(cfg, n_envs=1)
-
-    # create an empty online dataset similar to offline dataset
-    online_dataset = deepcopy(offline_dataset)
-    online_dataset.hf_dataset = {}
-    online_dataset.episode_data_index = {}
-
-    # create dataloader for online training
-    concat_dataset = torch.utils.data.ConcatDataset([offline_dataset, online_dataset])
-    weights = [1.0] * len(concat_dataset)
-    sampler = torch.utils.data.WeightedRandomSampler(
-        weights, num_samples=len(concat_dataset), replacement=True
-    )
-    dataloader = torch.utils.data.DataLoader(
-        concat_dataset,
-        num_workers=4,
-        batch_size=cfg.training.batch_size,
-        sampler=sampler,
-        pin_memory=cfg.device != "cpu",
-        drop_last=False,
-    )
-    dl_iter = cycle(dataloader)
-
-    online_step = 0
-    is_offline = False
-    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(
-                online_training_env,
-                policy,
-                n_episodes=1,
-                return_episode_data=True,
-                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.training.online_sampling_ratio,
-        )
-
-        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 = update_policy(policy, batch, optimizer, cfg.training.grad_clip_norm, lr_scheduler)
-
-            if step % cfg.training.log_freq == 0:
-                log_train_info(logger, train_info, step, cfg, online_dataset, is_offline)
-
-            # 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")
 
 
+@hydra.main(version_base="1.2", config_name="default", config_path="../configs")
+def train_cli(cfg: dict):
+    train(
+        cfg,
+        out_dir=hydra.core.hydra_config.HydraConfig.get().run.dir,
+        job_name=hydra.core.hydra_config.HydraConfig.get().job.name,
+    )
+
+
+def train_notebook(out_dir=None, job_name=None, config_name="default", config_path="../configs"):
+    from hydra import compose, initialize
+
+    hydra.core.global_hydra.GlobalHydra.instance().clear()
+    initialize(config_path=config_path)
+    cfg = compose(config_name=config_name)
+    train(cfg, out_dir=out_dir, job_name=job_name)
+
+
 if __name__ == "__main__":
     train_cli()

lerobot/scripts/visualize_dataset.py

@@ -30,48 +30,46 @@ Examples:
 - Visualize data stored on a local machine:
 ```
 local$ python lerobot/scripts/visualize_dataset.py \
-    --repo-id lerobot/pusht \
-    --episode-index 0
+    --repo-id lerobot/pusht
+
+local$ open http://localhost:9090
 ```
 
 - Visualize data stored on a distant machine with a local viewer:
 ```
 distant$ python lerobot/scripts/visualize_dataset.py \
+    --repo-id lerobot/pusht
+
+local$ ssh -L 9090:localhost:9090 distant  # create a ssh tunnel
+local$ open http://localhost:9090
+```
+
+- Select episodes to visualize:
+```
+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
+    --episode-indices 7 3 5 1 4
 ```
-
-- 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 gc
+import http.server
 import logging
-import time
+import os
+import shutil
+import socketserver
 from pathlib import Path
 
-import rerun as rr
 import torch
 import tqdm
+import yaml
+from bs4 import BeautifulSoup
+from huggingface_hub import snapshot_download
+from safetensors.torch import load_file, save_file
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.policies.act.modeling_act import ACTPolicy
+from lerobot.common.utils.utils import init_logging
 
 
 class EpisodeSampler(torch.utils.data.Sampler):
@@ -87,33 +85,307 @@ class EpisodeSampler(torch.utils.data.Sampler):
         return len(self.frame_ids)
 
 
-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
+class NoCacheHTTPRequestHandler(http.server.SimpleHTTPRequestHandler):
+    def end_headers(self):
+        self.send_header("Cache-Control", "no-store, no-cache, must-revalidate")
+        self.send_header("Pragma", "no-cache")
+        self.send_header("Expires", "0")
+        super().end_headers()
 
 
-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`."
+def run_server(path, port):
+    # Change directory to serve 'index.html` as front page
+    os.chdir(path)
 
-    logging.info("Loading dataset")
-    dataset = LeRobotDataset(repo_id)
+    with socketserver.TCPServer(("", port), NoCacheHTTPRequestHandler) as httpd:
+        logging.info(f"Serving HTTP on 0.0.0.0 port {port} (http://0.0.0.0:{port}/) ...")
+        httpd.serve_forever()
+
+
+def create_html_page(page_title: str):
+    """Create a html page with beautiful soop with default doctype, meta, header and title."""
+    soup = BeautifulSoup("", "html.parser")
+
+    doctype = soup.new_tag("!DOCTYPE html")
+    soup.append(doctype)
+
+    html = soup.new_tag("html", lang="en")
+    soup.append(html)
+
+    head = soup.new_tag("head")
+    html.append(head)
+
+    meta_charset = soup.new_tag("meta", charset="UTF-8")
+    head.append(meta_charset)
+
+    meta_viewport = soup.new_tag(
+        "meta", attrs={"name": "viewport", "content": "width=device-width, initial-scale=1.0"}
+    )
+    head.append(meta_viewport)
+
+    title = soup.new_tag("title")
+    title.string = page_title
+    head.append(title)
+
+    body = soup.new_tag("body")
+    html.append(body)
+
+    main_div = soup.new_tag("div")
+    body.append(main_div)
+    return soup, head, body
+
+
+def write_episode_data_csv(output_dir, file_name, episode_index, dataset, inference_results=None):
+    """Write a csv file containg timeseries data of an episode (e.g. state and action).
+    This file will be loaded by Dygraph javascript to plot data in real time."""
+    from_idx = dataset.episode_data_index["from"][episode_index]
+    to_idx = dataset.episode_data_index["to"][episode_index]
+
+    has_state = "observation.state" in dataset.hf_dataset.features
+    has_action = "action" in dataset.hf_dataset.features
+    has_inference = inference_results is not None
+
+    # init header of csv with state and action names
+    header = ["timestamp"]
+    if has_state:
+        dim_state = len(dataset.hf_dataset["observation.state"][0])
+        header += [f"state_{i}" for i in range(dim_state)]
+    if has_action:
+        dim_action = len(dataset.hf_dataset["action"][0])
+        header += [f"action_{i}" for i in range(dim_action)]
+    if has_inference:
+        assert "actions" in inference_results
+        assert "loss" in inference_results
+        dim_pred_action = inference_results["actions"].shape[2]
+        header += [f"pred_action_{i}" for i in range(dim_pred_action)]
+        header += ["loss"]
+
+    columns = ["timestamp"]
+    if has_state:
+        columns += ["observation.state"]
+    if has_action:
+        columns += ["action"]
+
+    rows = []
+    data = dataset.hf_dataset.select_columns(columns)
+    for i in range(from_idx, to_idx):
+        row = [data[i]["timestamp"].item()]
+        if has_state:
+            row += data[i]["observation.state"].tolist()
+        if has_action:
+            row += data[i]["action"].tolist()
+        rows.append(row)
+
+    if has_inference:
+        num_frames = len(rows)
+        assert num_frames == inference_results["actions"].shape[0]
+        assert num_frames == inference_results["loss"].shape[0]
+        for i in range(num_frames):
+            rows[i] += inference_results["actions"][i, 0].tolist()
+            rows[i] += [inference_results["loss"][i].item()]
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    with open(output_dir / file_name, "w") as f:
+        f.write(",".join(header) + "\n")
+        for row in rows:
+            row_str = [str(col) for col in row]
+            f.write(",".join(row_str) + "\n")
+
+
+def write_episode_data_js(output_dir, file_name, ep_csv_fname, dataset):
+    """Write a javascript file containing logic to synchronize camera feeds and timeseries."""
+    s = ""
+    s += "document.addEventListener('DOMContentLoaded', function () {\n"
+    for i, key in enumerate(dataset.video_frame_keys):
+        s += f"  const video{i} = document.getElementById('video_{key}');\n"
+    s += "  const slider = document.getElementById('videoControl');\n"
+    s += "  const playButton = document.getElementById('playButton');\n"
+    s += f"  const dygraph = new Dygraph(document.getElementById('graph'), '{ep_csv_fname}', " + "{\n"
+    s += "    pixelsPerPoint: 0.01,\n"
+    s += "    legend: 'always',\n"
+    s += "    labelsDiv: document.getElementById('labels'),\n"
+    s += "    labelsSeparateLines: true,\n"
+    s += "    labelsKMB: true,\n"
+    s += "    highlightCircleSize: 1.5,\n"
+    s += "    highlightSeriesOpts: {\n"
+    s += "        strokeWidth: 1.5,\n"
+    s += "        strokeBorderWidth: 1,\n"
+    s += "        highlightCircleSize: 3\n"
+    s += "    }\n"
+    s += "  });\n"
+    s += "\n"
+    s += "  // Function to play both videos\n"
+    s += "  playButton.addEventListener('click', function () {\n"
+    for i in range(len(dataset.video_frame_keys)):
+        s += f"    video{i}.play();\n"
+    s += "    // playButton.disabled = true; // Optional: disable button after playing\n"
+    s += "  });\n"
+    s += "\n"
+    s += "  // Update the video time when the slider value changes\n"
+    s += "  slider.addEventListener('input', function () {\n"
+    s += "    const sliderValue = slider.value;\n"
+    for i in range(len(dataset.video_frame_keys)):
+        s += f"    const time{i} = (video{i}.duration * sliderValue) / 100;\n"
+    for i in range(len(dataset.video_frame_keys)):
+        s += f"    video{i}.currentTime = time{i};\n"
+    s += "  });\n"
+    s += "\n"
+    s += "  // Synchronize slider with the video's current time\n"
+    s += "  const syncSlider = (video) => {\n"
+    s += "    video.addEventListener('timeupdate', function () {\n"
+    s += "      if (video.duration) {\n"
+    s += "        const pc = (100 / video.duration) * video.currentTime;\n"
+    s += "        slider.value = pc;\n"
+    s += "        const index = Math.floor(pc * dygraph.numRows() / 100);\n"
+    s += "        dygraph.setSelection(index, undefined, true, true);\n"
+    s += "      }\n"
+    s += "    });\n"
+    s += "  };\n"
+    s += "\n"
+    for i in range(len(dataset.video_frame_keys)):
+        s += f"  syncSlider(video{i});\n"
+    s += "\n"
+    s += "});\n"
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    with open(output_dir / file_name, "w", encoding="utf-8") as f:
+        f.write(s)
+
+
+def write_episode_data_html(output_dir, file_name, js_fname, ep_index, dataset):
+    """Write an html file containg video feeds and timeseries associated to an episode."""
+    soup, head, body = create_html_page("")
+
+    css_style = soup.new_tag("style")
+    css_style.string = ""
+    css_style.string += "#labels > span.highlight {\n"
+    css_style.string += "  border: 1px solid grey;\n"
+    css_style.string += "}"
+    head.append(css_style)
+
+    # Add videos from camera feeds
+
+    videos_control_div = soup.new_tag("div")
+    body.append(videos_control_div)
+
+    videos_div = soup.new_tag("div")
+    videos_control_div.append(videos_div)
+
+    def create_video(id, src):
+        video = soup.new_tag("video", id=id, width="320", height="240", controls="")
+        source = soup.new_tag("source", src=src, type="video/mp4")
+        video.string = "Your browser does not support the video tag."
+        video.append(source)
+        return video
+
+    # get first frame of episode (hack to get video_path of the episode)
+    first_frame_idx = dataset.episode_data_index["from"][ep_index].item()
+
+    for key in dataset.video_frame_keys:
+        # Example of video_path: 'videos/observation.image_episode_000004.mp4'
+        video_path = dataset.hf_dataset.select_columns(key)[first_frame_idx][key]["path"]
+        videos_div.append(create_video(f"video_{key}", video_path))
+
+    # Add controls for videos and graph
+
+    control_div = soup.new_tag("div")
+    videos_control_div.append(control_div)
+
+    button_div = soup.new_tag("div")
+    control_div.append(button_div)
+
+    button = soup.new_tag("button", id="playButton")
+    button.string = "Play Videos"
+    button_div.append(button)
+
+    slider_div = soup.new_tag("div")
+    control_div.append(slider_div)
+
+    slider = soup.new_tag("input", type="range", id="videoControl", min="0", max="100", value="0", step="1")
+    control_div.append(slider)
+
+    # Add graph of states/actions, and its labels
+
+    graph_labels_div = soup.new_tag("div", style="display: flex;")
+    body.append(graph_labels_div)
+
+    graph_div = soup.new_tag("div", id="graph", style="flex: 1; width: 85%")
+    graph_labels_div.append(graph_div)
+
+    labels_div = soup.new_tag("div", id="labels", style="flex: 1; width: 15%")
+    graph_labels_div.append(labels_div)
+
+    # add dygraph library
+    script = soup.new_tag("script", type="text/javascript", src=js_fname)
+    body.append(script)
+
+    script_dygraph = soup.new_tag(
+        "script",
+        type="text/javascript",
+        src="https://cdn.jsdelivr.net/npm/dygraphs@2.1.0/dist/dygraph.min.js",
+    )
+    body.append(script_dygraph)
+
+    link_dygraph = soup.new_tag(
+        "link", rel="stylesheet", href="https://cdn.jsdelivr.net/npm/dygraphs@2.1.0/dist/dygraph.min.css"
+    )
+    body.append(link_dygraph)
+
+    # Write as a html file
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    with open(output_dir / file_name, "w", encoding="utf-8") as f:
+        f.write(soup.prettify())
+
+
+def write_episodes_list_html(output_dir, file_name, ep_indices, ep_html_fnames, dataset):
+    """Write an html file containing information related to the dataset and a list of links to
+    html pages of episodes."""
+    soup, head, body = create_html_page("TODO")
+
+    h3 = soup.new_tag("h3")
+    h3.string = "TODO"
+    body.append(h3)
+
+    ul_info = soup.new_tag("ul")
+    body.append(ul_info)
+
+    li_info = soup.new_tag("li")
+    li_info.string = f"Number of samples/frames: {dataset.num_samples}"
+    ul_info.append(li_info)
+
+    li_info = soup.new_tag("li")
+    li_info.string = f"Number of episodes: {dataset.num_episodes}"
+    ul_info.append(li_info)
+
+    li_info = soup.new_tag("li")
+    li_info.string = f"Frames per second: {dataset.fps}"
+    ul_info.append(li_info)
+
+    # li_info = soup.new_tag("li")
+    # li_info.string = f"Size: {format_big_number(dataset.hf_dataset.info.size_in_bytes)}B"
+    # ul_info.append(li_info)
+
+    ul = soup.new_tag("ul")
+    body.append(ul)
+
+    for ep_idx, ep_html_fname in zip(ep_indices, ep_html_fnames, strict=False):
+        li = soup.new_tag("li")
+        ul.append(li)
+
+        a = soup.new_tag("a", href=ep_html_fname)
+        a.string = f"Episode number {ep_idx}"
+
+        li.append(a)
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    with open(output_dir / file_name, "w", encoding="utf-8") as f:
+        f.write(soup.prettify())
+
+
+def run_inference(dataset, episode_index, policy, num_workers=4, batch_size=32, device="cuda"):
+    policy.eval()
+    policy.to(device)
 
     logging.info("Loading dataloader")
     episode_sampler = EpisodeSampler(dataset, episode_index)
@@ -124,70 +396,104 @@ def visualize_dataset(
         sampler=episode_sampler,
     )
 
-    logging.info("Starting Rerun")
-
-    if mode not in ["local", "distant"]:
-        raise ValueError(mode)
-
-    spawn_local_viewer = mode == "local" and not save
-    rr.init(f"{repo_id}/episode_{episode_index}", spawn=spawn_local_viewer)
-
-    # Manually call python garbage collector after `rr.init` to avoid hanging in a blocking flush
-    # when iterating on a dataloader with `num_workers` > 0
-    # TODO(rcadene): remove `gc.collect` when rerun version 0.16 is out, which includes a fix
-    gc.collect()
-
-    if mode == "distant":
-        rr.serve(open_browser=False, web_port=web_port, ws_port=ws_port)
-
-    logging.info("Logging to Rerun")
-
+    logging.info("Running inference")
+    inference_results = {}
     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())
+        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+        with torch.inference_mode():
+            output_dict = policy.forward(batch)
 
-            # 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])))
+        for key in output_dict:
+            if key not in inference_results:
+                inference_results[key] = []
+            inference_results[key].append(output_dict[key].to("cpu"))
 
-            # 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()))
+    for key in inference_results:
+        inference_results[key] = torch.cat(inference_results[key])
 
-            # 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()))
+    return inference_results
 
-            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()))
+def visualize_dataset(
+    repo_id: str,
+    episode_indices: list[int] = None,
+    output_dir: Path | None = None,
+    serve: bool = True,
+    port: int = 9090,
+    force_overwrite: bool = True,
+    policy_repo_id: str | None = None,
+    policy_ckpt_path: Path | None = None,
+    batch_size: int = 32,
+    num_workers: int = 4,
+) -> Path | None:
+    init_logging()
 
-            if "next.success" in batch:
-                rr.log("next.success", rr.Scalar(batch["next.success"][i].item()))
+    has_policy = policy_repo_id or policy_ckpt_path
 
-    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
+    if has_policy:
+        logging.info("Loading policy")
+        if policy_repo_id:
+            pretrained_policy_path = Path(snapshot_download(policy_repo_id))
+        elif policy_ckpt_path:
+            pretrained_policy_path = Path(policy_ckpt_path)
+        policy = ACTPolicy.from_pretrained(pretrained_policy_path)
+        with open(pretrained_policy_path / "config.yaml") as f:
+            cfg = yaml.safe_load(f)
+        delta_timestamps = cfg["training"]["delta_timestamps"]
+    else:
+        delta_timestamps = None
 
-    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.")
+    logging.info("Loading dataset")
+    dataset = LeRobotDataset(repo_id, delta_timestamps=delta_timestamps)
+
+    if not dataset.video:
+        raise NotImplementedError(f"Image datasets ({dataset.video=}) are currently not supported.")
+
+    if output_dir is None:
+        output_dir = f"outputs/visualize_dataset/{repo_id}"
+
+    output_dir = Path(output_dir)
+    if force_overwrite and output_dir.exists():
+        shutil.rmtree(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    # Create a simlink from the dataset video folder containg mp4 files to the output directory
+    # so that the http server can get access to the mp4 files.
+    ln_videos_dir = output_dir / "videos"
+    if not ln_videos_dir.exists():
+        ln_videos_dir.symlink_to(dataset.videos_dir.resolve())
+
+    if episode_indices is None:
+        episode_indices = list(range(dataset.num_episodes))
+
+    logging.info("Writing html")
+    ep_html_fnames = []
+    for episode_index in tqdm.tqdm(episode_indices):
+        inference_results = None
+        if has_policy:
+            inference_results_path = output_dir / f"episode_{episode_index}.safetensors"
+            if inference_results_path.exists():
+                inference_results = load_file(inference_results_path)
+            else:
+                inference_results = run_inference(dataset, episode_index, policy)
+                save_file(inference_results, inference_results_path)
+
+        # write states and actions in a csv
+        ep_csv_fname = f"episode_{episode_index}.csv"
+        write_episode_data_csv(output_dir, ep_csv_fname, episode_index, dataset, inference_results)
+
+        js_fname = f"episode_{episode_index}.js"
+        write_episode_data_js(output_dir, js_fname, ep_csv_fname, dataset)
+
+        # write a html page to view videos and timeseries
+        ep_html_fname = f"episode_{episode_index}.html"
+        write_episode_data_html(output_dir, ep_html_fname, js_fname, episode_index, dataset)
+        ep_html_fnames.append(ep_html_fname)
+
+    write_episodes_list_html(output_dir, "index.html", episode_indices, ep_html_fnames, dataset)
+
+    if serve:
+        run_server(output_dir, port)
 
 
 def main():
@@ -197,13 +503,51 @@ def main():
         "--repo-id",
         type=str,
         required=True,
-        help="Name of hugging face repositery containing a LeRobotDataset dataset (e.g. `lerobot/pusht`).",
+        help="Name of hugging face repositery containing a LeRobotDataset dataset (e.g. `lerobot/pusht` for https://huggingface.co/datasets/lerobot/pusht).",
     )
     parser.add_argument(
-        "--episode-index",
+        "--episode-indices",
         type=int,
-        required=True,
-        help="Episode to visualize.",
+        nargs="*",
+        default=None,
+        help="Episode indices to visualize (e.g. `0 1 5 6` to load episodes of index 0, 1, 5 and 6). By default loads all episodes.",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=str,
+        default=None,
+        help="Directory path to write html files and kickoff a web server. By default write them to 'outputs/visualize_dataset/REPO_ID'.",
+    )
+    parser.add_argument(
+        "--serve",
+        type=int,
+        default=1,
+        help="Launch web server.",
+    )
+    parser.add_argument(
+        "--port",
+        type=int,
+        default=9090,
+        help="Web port used by the http server.",
+    )
+    parser.add_argument(
+        "--force-overwrite",
+        type=int,
+        default=1,
+        help="Delete the output directory if it exists already.",
+    )
+
+    parser.add_argument(
+        "--policy-repo-id",
+        type=str,
+        default=None,
+        help="Name of hugging face repositery containing a pretrained policy (e.g. `lerobot/diffusion_pusht` for https://huggingface.co/lerobot/diffusion_pusht).",
+    )
+    parser.add_argument(
+        "--policy-ckpt-path",
+        type=str,
+        default=None,
+        help="Name of hugging face repositery containing a pretrained policy (e.g. `lerobot/diffusion_pusht` for https://huggingface.co/lerobot/diffusion_pusht).",
     )
     parser.add_argument(
         "--batch-size",
@@ -217,43 +561,6 @@ def main():
         default=4,
         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))

lerobot/scripts/visualize_dataset_rerun.py

@@ -0,0 +1,263 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" 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 gc
+import logging
+import time
+from pathlib import Path
+
+import rerun as rr
+import torch
+import tqdm
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+
+
+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 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(
+    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`."
+
+    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=num_workers,
+        batch_size=batch_size,
+        sampler=episode_sampler,
+    )
+
+    logging.info("Starting Rerun")
+
+    if mode not in ["local", "distant"]:
+        raise ValueError(mode)
+
+    spawn_local_viewer = mode == "local" and not save
+    rr.init(f"{repo_id}/episode_{episode_index}", spawn=spawn_local_viewer)
+
+    # Manually call python garbage collector after `rr.init` to avoid hanging in a blocking flush
+    # when iterating on a dataloader with `num_workers` > 0
+    # TODO(rcadene): remove `gc.collect` when rerun version 0.16 is out, which includes a fix
+    gc.collect()
+
+    if mode == "distant":
+        rr.serve(open_browser=False, web_port=web_port, ws_port=ws_port)
+
+    logging.info("Logging to Rerun")
+
+    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())
+
+            # 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])))
+
+            # 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=4,
+        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__":
+    main()

poetry.lock

@@ -1,4 +1,4 @@
-# This file is automatically @generated by Poetry 1.8.2 and should not be changed by hand.
+# This file is automatically @generated by Poetry 1.8.1 and should not be changed by hand.
 
 [[package]]
 name = "absl-py"
@@ -444,63 +444,63 @@ files = [
 
 [[package]]
 name = "coverage"
-version = "7.5.1"
+version = "7.5.3"
 description = "Code coverage measurement for Python"
 optional = true
 python-versions = ">=3.8"
 files = [
-    {file = "coverage-7.5.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:c0884920835a033b78d1c73b6d3bbcda8161a900f38a488829a83982925f6c2e"},
-    {file = "coverage-7.5.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:39afcd3d4339329c5f58de48a52f6e4e50f6578dd6099961cf22228feb25f38f"},
-    {file = "coverage-7.5.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4a7b0ceee8147444347da6a66be737c9d78f3353b0681715b668b72e79203e4a"},
-    {file = "coverage-7.5.1-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:4a9ca3f2fae0088c3c71d743d85404cec8df9be818a005ea065495bedc33da35"},
-    {file = "coverage-7.5.1-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5fd215c0c7d7aab005221608a3c2b46f58c0285a819565887ee0b718c052aa4e"},
-    {file = "coverage-7.5.1-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:4bf0655ab60d754491004a5efd7f9cccefcc1081a74c9ef2da4735d6ee4a6223"},
-    {file = "coverage-7.5.1-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:61c4bf1ba021817de12b813338c9be9f0ad5b1e781b9b340a6d29fc13e7c1b5e"},
-    {file = "coverage-7.5.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:db66fc317a046556a96b453a58eced5024af4582a8dbdc0c23ca4dbc0d5b3146"},
-    {file = "coverage-7.5.1-cp310-cp310-win32.whl", hash = "sha256:b016ea6b959d3b9556cb401c55a37547135a587db0115635a443b2ce8f1c7228"},
-    {file = "coverage-7.5.1-cp310-cp310-win_amd64.whl", hash = "sha256:df4e745a81c110e7446b1cc8131bf986157770fa405fe90e15e850aaf7619bc8"},
-    {file = "coverage-7.5.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:796a79f63eca8814ca3317a1ea443645c9ff0d18b188de470ed7ccd45ae79428"},
-    {file = "coverage-7.5.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:4fc84a37bfd98db31beae3c2748811a3fa72bf2007ff7902f68746d9757f3746"},
-    {file = "coverage-7.5.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6175d1a0559986c6ee3f7fccfc4a90ecd12ba0a383dcc2da30c2b9918d67d8a3"},
-    {file = "coverage-7.5.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:1fc81d5878cd6274ce971e0a3a18a8803c3fe25457165314271cf78e3aae3aa2"},
-    {file = "coverage-7.5.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:556cf1a7cbc8028cb60e1ff0be806be2eded2daf8129b8811c63e2b9a6c43bca"},
-    {file = "coverage-7.5.1-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:9981706d300c18d8b220995ad22627647be11a4276721c10911e0e9fa44c83e8"},
-    {file = "coverage-7.5.1-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:d7fed867ee50edf1a0b4a11e8e5d0895150e572af1cd6d315d557758bfa9c057"},
-    {file = "coverage-7.5.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:ef48e2707fb320c8f139424a596f5b69955a85b178f15af261bab871873bb987"},
-    {file = "coverage-7.5.1-cp311-cp311-win32.whl", hash = "sha256:9314d5678dcc665330df5b69c1e726a0e49b27df0461c08ca12674bcc19ef136"},
-    {file = "coverage-7.5.1-cp311-cp311-win_amd64.whl", hash = "sha256:5fa567e99765fe98f4e7d7394ce623e794d7cabb170f2ca2ac5a4174437e90dd"},
-    {file = "coverage-7.5.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:b6cf3764c030e5338e7f61f95bd21147963cf6aa16e09d2f74f1fa52013c1206"},
-    {file = "coverage-7.5.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:2ec92012fefebee89a6b9c79bc39051a6cb3891d562b9270ab10ecfdadbc0c34"},
-    {file = "coverage-7.5.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:16db7f26000a07efcf6aea00316f6ac57e7d9a96501e990a36f40c965ec7a95d"},
-    {file = "coverage-7.5.1-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:beccf7b8a10b09c4ae543582c1319c6df47d78fd732f854ac68d518ee1fb97fa"},
-    {file = "coverage-7.5.1-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8748731ad392d736cc9ccac03c9845b13bb07d020a33423fa5b3a36521ac6e4e"},
-    {file = "coverage-7.5.1-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:7352b9161b33fd0b643ccd1f21f3a3908daaddf414f1c6cb9d3a2fd618bf2572"},
-    {file = "coverage-7.5.1-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:7a588d39e0925f6a2bff87154752481273cdb1736270642aeb3635cb9b4cad07"},
-    {file = "coverage-7.5.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:68f962d9b72ce69ea8621f57551b2fa9c70509af757ee3b8105d4f51b92b41a7"},
-    {file = "coverage-7.5.1-cp312-cp312-win32.whl", hash = "sha256:f152cbf5b88aaeb836127d920dd0f5e7edff5a66f10c079157306c4343d86c19"},
-    {file = "coverage-7.5.1-cp312-cp312-win_amd64.whl", hash = "sha256:5a5740d1fb60ddf268a3811bcd353de34eb56dc24e8f52a7f05ee513b2d4f596"},
-    {file = "coverage-7.5.1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:e2213def81a50519d7cc56ed643c9e93e0247f5bbe0d1247d15fa520814a7cd7"},
-    {file = "coverage-7.5.1-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:5037f8fcc2a95b1f0e80585bd9d1ec31068a9bcb157d9750a172836e98bc7a90"},
-    {file = "coverage-7.5.1-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5c3721c2c9e4c4953a41a26c14f4cef64330392a6d2d675c8b1db3b645e31f0e"},
-    {file = "coverage-7.5.1-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:ca498687ca46a62ae590253fba634a1fe9836bc56f626852fb2720f334c9e4e5"},
-    {file = "coverage-7.5.1-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0cdcbc320b14c3e5877ee79e649677cb7d89ef588852e9583e6b24c2e5072661"},
-    {file = "coverage-7.5.1-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:57e0204b5b745594e5bc14b9b50006da722827f0b8c776949f1135677e88d0b8"},
-    {file = "coverage-7.5.1-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:8fe7502616b67b234482c3ce276ff26f39ffe88adca2acf0261df4b8454668b4"},
-    {file = "coverage-7.5.1-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:9e78295f4144f9dacfed4f92935fbe1780021247c2fabf73a819b17f0ccfff8d"},
-    {file = "coverage-7.5.1-cp38-cp38-win32.whl", hash = "sha256:1434e088b41594baa71188a17533083eabf5609e8e72f16ce8c186001e6b8c41"},
-    {file = "coverage-7.5.1-cp38-cp38-win_amd64.whl", hash = "sha256:0646599e9b139988b63704d704af8e8df7fa4cbc4a1f33df69d97f36cb0a38de"},
-    {file = "coverage-7.5.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:4cc37def103a2725bc672f84bd939a6fe4522310503207aae4d56351644682f1"},
-    {file = "coverage-7.5.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:fc0b4d8bfeabd25ea75e94632f5b6e047eef8adaed0c2161ada1e922e7f7cece"},
-    {file = "coverage-7.5.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0d0a0f5e06881ecedfe6f3dd2f56dcb057b6dbeb3327fd32d4b12854df36bf26"},
-    {file = "coverage-7.5.1-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:9735317685ba6ec7e3754798c8871c2f49aa5e687cc794a0b1d284b2389d1bd5"},
-    {file = "coverage-7.5.1-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d21918e9ef11edf36764b93101e2ae8cc82aa5efdc7c5a4e9c6c35a48496d601"},
-    {file = "coverage-7.5.1-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:c3e757949f268364b96ca894b4c342b41dc6f8f8b66c37878aacef5930db61be"},
-    {file = "coverage-7.5.1-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:79afb6197e2f7f60c4824dd4b2d4c2ec5801ceb6ba9ce5d2c3080e5660d51a4f"},
-    {file = "coverage-7.5.1-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:d1d0d98d95dd18fe29dc66808e1accf59f037d5716f86a501fc0256455219668"},
-    {file = "coverage-7.5.1-cp39-cp39-win32.whl", hash = "sha256:1cc0fe9b0b3a8364093c53b0b4c0c2dd4bb23acbec4c9240b5f284095ccf7981"},
-    {file = "coverage-7.5.1-cp39-cp39-win_amd64.whl", hash = "sha256:dde0070c40ea8bb3641e811c1cfbf18e265d024deff6de52c5950677a8fb1e0f"},
-    {file = "coverage-7.5.1-pp38.pp39.pp310-none-any.whl", hash = "sha256:6537e7c10cc47c595828b8a8be04c72144725c383c4702703ff4e42e44577312"},
-    {file = "coverage-7.5.1.tar.gz", hash = "sha256:54de9ef3a9da981f7af93eafde4ede199e0846cd819eb27c88e2b712aae9708c"},
+    {file = "coverage-7.5.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:a6519d917abb15e12380406d721e37613e2a67d166f9fb7e5a8ce0375744cd45"},
+    {file = "coverage-7.5.3-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:aea7da970f1feccf48be7335f8b2ca64baf9b589d79e05b9397a06696ce1a1ec"},
+    {file = "coverage-7.5.3-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:923b7b1c717bd0f0f92d862d1ff51d9b2b55dbbd133e05680204465f454bb286"},
+    {file = "coverage-7.5.3-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:62bda40da1e68898186f274f832ef3e759ce929da9a9fd9fcf265956de269dbc"},
+    {file = "coverage-7.5.3-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d8b7339180d00de83e930358223c617cc343dd08e1aa5ec7b06c3a121aec4e1d"},
+    {file = "coverage-7.5.3-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:25a5caf742c6195e08002d3b6c2dd6947e50efc5fc2c2205f61ecb47592d2d83"},
+    {file = "coverage-7.5.3-cp310-cp310-musllinux_1_1_i686.whl", hash = "sha256:05ac5f60faa0c704c0f7e6a5cbfd6f02101ed05e0aee4d2822637a9e672c998d"},
+    {file = "coverage-7.5.3-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:239a4e75e09c2b12ea478d28815acf83334d32e722e7433471fbf641c606344c"},
+    {file = "coverage-7.5.3-cp310-cp310-win32.whl", hash = "sha256:a5812840d1d00eafae6585aba38021f90a705a25b8216ec7f66aebe5b619fb84"},
+    {file = "coverage-7.5.3-cp310-cp310-win_amd64.whl", hash = "sha256:33ca90a0eb29225f195e30684ba4a6db05dbef03c2ccd50b9077714c48153cac"},
+    {file = "coverage-7.5.3-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:f81bc26d609bf0fbc622c7122ba6307993c83c795d2d6f6f6fd8c000a770d974"},
+    {file = "coverage-7.5.3-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:7cec2af81f9e7569280822be68bd57e51b86d42e59ea30d10ebdbb22d2cb7232"},
+    {file = "coverage-7.5.3-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:55f689f846661e3f26efa535071775d0483388a1ccfab899df72924805e9e7cd"},
+    {file = "coverage-7.5.3-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:50084d3516aa263791198913a17354bd1dc627d3c1639209640b9cac3fef5807"},
+    {file = "coverage-7.5.3-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:341dd8f61c26337c37988345ca5c8ccabeff33093a26953a1ac72e7d0103c4fb"},
+    {file = "coverage-7.5.3-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:ab0b028165eea880af12f66086694768f2c3139b2c31ad5e032c8edbafca6ffc"},
+    {file = "coverage-7.5.3-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:5bc5a8c87714b0c67cfeb4c7caa82b2d71e8864d1a46aa990b5588fa953673b8"},
+    {file = "coverage-7.5.3-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:38a3b98dae8a7c9057bd91fbf3415c05e700a5114c5f1b5b0ea5f8f429ba6614"},
+    {file = "coverage-7.5.3-cp311-cp311-win32.whl", hash = "sha256:fcf7d1d6f5da887ca04302db8e0e0cf56ce9a5e05f202720e49b3e8157ddb9a9"},
+    {file = "coverage-7.5.3-cp311-cp311-win_amd64.whl", hash = "sha256:8c836309931839cca658a78a888dab9676b5c988d0dd34ca247f5f3e679f4e7a"},
+    {file = "coverage-7.5.3-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:296a7d9bbc598e8744c00f7a6cecf1da9b30ae9ad51c566291ff1314e6cbbed8"},
+    {file = "coverage-7.5.3-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:34d6d21d8795a97b14d503dcaf74226ae51eb1f2bd41015d3ef332a24d0a17b3"},
+    {file = "coverage-7.5.3-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8e317953bb4c074c06c798a11dbdd2cf9979dbcaa8ccc0fa4701d80042d4ebf1"},
+    {file = "coverage-7.5.3-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:705f3d7c2b098c40f5b81790a5fedb274113373d4d1a69e65f8b68b0cc26f6db"},
+    {file = "coverage-7.5.3-cp312-cp312-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b1196e13c45e327d6cd0b6e471530a1882f1017eb83c6229fc613cd1a11b53cd"},
+    {file = "coverage-7.5.3-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:015eddc5ccd5364dcb902eaecf9515636806fa1e0d5bef5769d06d0f31b54523"},
+    {file = "coverage-7.5.3-cp312-cp312-musllinux_1_1_i686.whl", hash = "sha256:fd27d8b49e574e50caa65196d908f80e4dff64d7e592d0c59788b45aad7e8b35"},
+    {file = "coverage-7.5.3-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:33fc65740267222fc02975c061eb7167185fef4cc8f2770267ee8bf7d6a42f84"},
+    {file = "coverage-7.5.3-cp312-cp312-win32.whl", hash = "sha256:7b2a19e13dfb5c8e145c7a6ea959485ee8e2204699903c88c7d25283584bfc08"},
+    {file = "coverage-7.5.3-cp312-cp312-win_amd64.whl", hash = "sha256:0bbddc54bbacfc09b3edaec644d4ac90c08ee8ed4844b0f86227dcda2d428fcb"},
+    {file = "coverage-7.5.3-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:f78300789a708ac1f17e134593f577407d52d0417305435b134805c4fb135adb"},
+    {file = "coverage-7.5.3-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:b368e1aee1b9b75757942d44d7598dcd22a9dbb126affcbba82d15917f0cc155"},
+    {file = "coverage-7.5.3-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:f836c174c3a7f639bded48ec913f348c4761cbf49de4a20a956d3431a7c9cb24"},
+    {file = "coverage-7.5.3-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:244f509f126dc71369393ce5fea17c0592c40ee44e607b6d855e9c4ac57aac98"},
+    {file = "coverage-7.5.3-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c4c2872b3c91f9baa836147ca33650dc5c172e9273c808c3c3199c75490e709d"},
+    {file = "coverage-7.5.3-cp38-cp38-musllinux_1_1_aarch64.whl", hash = "sha256:dd4b3355b01273a56b20c219e74e7549e14370b31a4ffe42706a8cda91f19f6d"},
+    {file = "coverage-7.5.3-cp38-cp38-musllinux_1_1_i686.whl", hash = "sha256:f542287b1489c7a860d43a7d8883e27ca62ab84ca53c965d11dac1d3a1fab7ce"},
+    {file = "coverage-7.5.3-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:75e3f4e86804023e991096b29e147e635f5e2568f77883a1e6eed74512659ab0"},
+    {file = "coverage-7.5.3-cp38-cp38-win32.whl", hash = "sha256:c59d2ad092dc0551d9f79d9d44d005c945ba95832a6798f98f9216ede3d5f485"},
+    {file = "coverage-7.5.3-cp38-cp38-win_amd64.whl", hash = "sha256:fa21a04112c59ad54f69d80e376f7f9d0f5f9123ab87ecd18fbb9ec3a2beed56"},
+    {file = "coverage-7.5.3-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:f5102a92855d518b0996eb197772f5ac2a527c0ec617124ad5242a3af5e25f85"},
+    {file = "coverage-7.5.3-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:d1da0a2e3b37b745a2b2a678a4c796462cf753aebf94edcc87dcc6b8641eae31"},
+    {file = "coverage-7.5.3-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8383a6c8cefba1b7cecc0149415046b6fc38836295bc4c84e820872eb5478b3d"},
+    {file = "coverage-7.5.3-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:9aad68c3f2566dfae84bf46295a79e79d904e1c21ccfc66de88cd446f8686341"},
+    {file = "coverage-7.5.3-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2e079c9ec772fedbade9d7ebc36202a1d9ef7291bc9b3a024ca395c4d52853d7"},
+    {file = "coverage-7.5.3-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:bde997cac85fcac227b27d4fb2c7608a2c5f6558469b0eb704c5726ae49e1c52"},
+    {file = "coverage-7.5.3-cp39-cp39-musllinux_1_1_i686.whl", hash = "sha256:990fb20b32990b2ce2c5f974c3e738c9358b2735bc05075d50a6f36721b8f303"},
+    {file = "coverage-7.5.3-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:3d5a67f0da401e105753d474369ab034c7bae51a4c31c77d94030d59e41df5bd"},
+    {file = "coverage-7.5.3-cp39-cp39-win32.whl", hash = "sha256:e08c470c2eb01977d221fd87495b44867a56d4d594f43739a8028f8646a51e0d"},
+    {file = "coverage-7.5.3-cp39-cp39-win_amd64.whl", hash = "sha256:1d2a830ade66d3563bb61d1e3c77c8def97b30ed91e166c67d0632c018f380f0"},
+    {file = "coverage-7.5.3-pp38.pp39.pp310-none-any.whl", hash = "sha256:3538d8fb1ee9bdd2e2692b3b18c22bb1c19ffbefd06880f5ac496e42d7bb3884"},
+    {file = "coverage-7.5.3.tar.gz", hash = "sha256:04aefca5190d1dc7a53a4c1a5a7f8568811306d7a8ee231c42fb69215571944f"},
 ]
 
 [package.dependencies]
@@ -595,6 +595,24 @@ files = [
     {file = "decorator-4.4.2.tar.gz", hash = "sha256:e3a62f0520172440ca0dcc823749319382e377f37f140a0b99ef45fecb84bfe7"},
 ]
 
+[[package]]
+name = "deepdiff"
+version = "7.0.1"
+description = "Deep Difference and Search of any Python object/data. Recreate objects by adding adding deltas to each other."
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "deepdiff-7.0.1-py3-none-any.whl", hash = "sha256:447760081918216aa4fd4ca78a4b6a848b81307b2ea94c810255334b759e1dc3"},
+    {file = "deepdiff-7.0.1.tar.gz", hash = "sha256:260c16f052d4badbf60351b4f77e8390bee03a0b516246f6839bc813fb429ddf"},
+]
+
+[package.dependencies]
+ordered-set = ">=4.1.0,<4.2.0"
+
+[package.extras]
+cli = ["click (==8.1.7)", "pyyaml (==6.0.1)"]
+optimize = ["orjson"]
+
 [[package]]
 name = "diffusers"
 version = "0.27.2"
@@ -1075,7 +1093,7 @@ description = ""
 optional = true
 python-versions = "^3.10"
 files = []
-develop = true
+develop = false
 
 [package.dependencies]
 dora-rs = ">=0.3.4"
@@ -1083,8 +1101,11 @@ gymnasium = ">=0.29.1"
 pyarrow = ">=12.0.0"
 
 [package.source]
-type = "directory"
-url = "gym_dora"
+type = "git"
+url = "https://github.com/dora-rs/dora-lerobot.git"
+reference = "HEAD"
+resolved_reference = "ed0c00a4fdc6ec856c9842551acd7dc7ee776f79"
+subdirectory = "gym_dora"
 
 [[package]]
 name = "gym-pusht"
@@ -1289,13 +1310,13 @@ files = [
 
 [[package]]
 name = "huggingface-hub"
-version = "0.23.0"
+version = "0.23.2"
 description = "Client library to download and publish models, datasets and other repos on the huggingface.co hub"
 optional = false
 python-versions = ">=3.8.0"
 files = [
-    {file = "huggingface_hub-0.23.0-py3-none-any.whl", hash = "sha256:075c30d48ee7db2bba779190dc526d2c11d422aed6f9044c5e2fdc2c432fdb91"},
-    {file = "huggingface_hub-0.23.0.tar.gz", hash = "sha256:7126dedd10a4c6fac796ced4d87a8cf004efc722a5125c2c09299017fa366fa9"},
+    {file = "huggingface_hub-0.23.2-py3-none-any.whl", hash = "sha256:48727a16e704d409c4bb5913613308499664f22a99743435dc3a13b23c485827"},
+    {file = "huggingface_hub-0.23.2.tar.gz", hash = "sha256:f6829b62d5fdecb452a76fdbec620cba4c1573655a8d710c1df71735fd9edbd2"},
 ]
 
 [package.dependencies]
@@ -2081,18 +2102,15 @@ test = ["pytest (>=7.2)", "pytest-cov (>=4.0)"]
 
 [[package]]
 name = "nodeenv"
-version = "1.8.0"
+version = "1.9.0"
 description = "Node.js virtual environment builder"
 optional = true
-python-versions = ">=2.7,!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*"
+python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
 files = [
-    {file = "nodeenv-1.8.0-py2.py3-none-any.whl", hash = "sha256:df865724bb3c3adc86b3876fa209771517b0cfe596beff01a92700e0e8be4cec"},
-    {file = "nodeenv-1.8.0.tar.gz", hash = "sha256:d51e0c37e64fbf47d017feac3145cdbb58836d7eee8c6f6d3b6880c5456227d2"},
+    {file = "nodeenv-1.9.0-py2.py3-none-any.whl", hash = "sha256:508ecec98f9f3330b636d4448c0f1a56fc68017c68f1e7857ebc52acf0eb879a"},
+    {file = "nodeenv-1.9.0.tar.gz", hash = "sha256:07f144e90dae547bf0d4ee8da0ee42664a42a04e02ed68e06324348dafe4bdb1"},
 ]
 
-[package.dependencies]
-setuptools = "*"
-
 [[package]]
 name = "numba"
 version = "0.59.1"
@@ -2334,13 +2352,13 @@ files = [
 
 [[package]]
 name = "nvidia-nvjitlink-cu12"
-version = "12.4.127"
+version = "12.5.40"
 description = "Nvidia JIT LTO Library"
 optional = false
 python-versions = ">=3"
 files = [
-    {file = "nvidia_nvjitlink_cu12-12.4.127-py3-none-manylinux2014_x86_64.whl", hash = "sha256:06b3b9b25bf3f8af351d664978ca26a16d2c5127dbd53c0497e28d1fb9611d57"},
-    {file = "nvidia_nvjitlink_cu12-12.4.127-py3-none-win_amd64.whl", hash = "sha256:fd9020c501d27d135f983c6d3e244b197a7ccad769e34df53a42e276b0e25fa1"},
+    {file = "nvidia_nvjitlink_cu12-12.5.40-py3-none-manylinux2014_x86_64.whl", hash = "sha256:d9714f27c1d0f0895cd8915c07a87a1d0029a0aa36acaf9156952ec2a8a12189"},
+    {file = "nvidia_nvjitlink_cu12-12.5.40-py3-none-win_amd64.whl", hash = "sha256:c3401dc8543b52d3a8158007a0c1ab4e9c768fcbd24153a48c86972102197ddd"},
 ]
 
 [[package]]
@@ -2393,6 +2411,20 @@ numpy = [
     {version = ">=1.21.2", markers = "platform_system != \"Darwin\" and python_version >= \"3.10\" and python_version < \"3.11\""},
 ]
 
+[[package]]
+name = "ordered-set"
+version = "4.1.0"
+description = "An OrderedSet is a custom MutableSet that remembers its order, so that every"
+optional = false
+python-versions = ">=3.7"
+files = [
+    {file = "ordered-set-4.1.0.tar.gz", hash = "sha256:694a8e44c87657c59292ede72891eb91d34131f6531463aab3009191c77364a8"},
+    {file = "ordered_set-4.1.0-py3-none-any.whl", hash = "sha256:046e1132c71fcf3330438a539928932caf51ddbc582496833e23de611de14562"},
+]
+
+[package.extras]
+dev = ["black", "mypy", "pytest"]
+
 [[package]]
 name = "packaging"
 version = "24.0"
@@ -3196,13 +3228,13 @@ files = [
 
 [[package]]
 name = "requests"
-version = "2.32.1"
+version = "2.32.3"
 description = "Python HTTP for Humans."
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "requests-2.32.1-py3-none-any.whl", hash = "sha256:21ac9465cdf8c1650fe1ecde8a71669a93d4e6f147550483a2967d08396a56a5"},
-    {file = "requests-2.32.1.tar.gz", hash = "sha256:eb97e87e64c79e64e5b8ac75cee9dd1f97f49e289b083ee6be96268930725685"},
+    {file = "requests-2.32.3-py3-none-any.whl", hash = "sha256:70761cfe03c773ceb22aa2f671b4757976145175cdfca038c02654d061d6dcc6"},
+    {file = "requests-2.32.3.tar.gz", hash = "sha256:55365417734eb18255590a9ff9eb97e9e1da868d4ccd6402399eaf68af20a760"},
 ]
 
 [package.dependencies]
@@ -3218,16 +3250,16 @@ use-chardet-on-py3 = ["chardet (>=3.0.2,<6)"]
 
 [[package]]
 name = "rerun-sdk"
-version = "0.16.0"
+version = "0.16.1"
 description = "The Rerun Logging SDK"
 optional = false
 python-versions = "<3.13,>=3.8"
 files = [
-    {file = "rerun_sdk-0.16.0-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:1cc6dc66d089e296f945dc238301889efb61dd6d338b5d00f76981cf7aed0a74"},
-    {file = "rerun_sdk-0.16.0-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:faf231897655e46eb975695df2b0ace07db362d697e697f9a3dff52f81c0dc5d"},
-    {file = "rerun_sdk-0.16.0-cp38-abi3-manylinux_2_31_aarch64.whl", hash = "sha256:860a6394380d3e9b9e48bf34423bd56dda54d5b0158d2ae0e433698659b34198"},
-    {file = "rerun_sdk-0.16.0-cp38-abi3-manylinux_2_31_x86_64.whl", hash = "sha256:5b8d1476f73a3ad1a5d3f21b61c633f3ab62aa80fa0b049f5ad10bf1227681ab"},
-    {file = "rerun_sdk-0.16.0-cp38-abi3-win_amd64.whl", hash = "sha256:aff0051a263b8c3067243c0126d319845baf4fe640899f04aeef7daf151f35e4"},
+    {file = "rerun_sdk-0.16.1-cp38-abi3-macosx_10_12_x86_64.whl", hash = "sha256:170c6976634008611753e10dfef8cdc395ce8180e634c169e7c61cef2f89a277"},
+    {file = "rerun_sdk-0.16.1-cp38-abi3-macosx_11_0_arm64.whl", hash = "sha256:c9a76eab7eb5559276737dad655200e9350df0837158dbc5a896970ab4201454"},
+    {file = "rerun_sdk-0.16.1-cp38-abi3-manylinux_2_31_aarch64.whl", hash = "sha256:4d6436752d57e8b8038489a0e7e37f0c760b088e96db5fb81667d3a376d63fea"},
+    {file = "rerun_sdk-0.16.1-cp38-abi3-manylinux_2_31_x86_64.whl", hash = "sha256:37b7b47948471873e84f224b16f417a94a91c7cbd6c72c68281eeff1ba414b8f"},
+    {file = "rerun_sdk-0.16.1-cp38-abi3-win_amd64.whl", hash = "sha256:be88799c8afdf68eafa99e64e2e4f0a484e187e017a180219abbe6bb988acd4e"},
 ]
 
 [package.dependencies]
@@ -3407,36 +3439,36 @@ test = ["asv", "numpydoc (>=1.7)", "pooch (>=1.6.0)", "pytest (>=7.0)", "pytest-
 
 [[package]]
 name = "scipy"
-version = "1.13.0"
+version = "1.13.1"
 description = "Fundamental algorithms for scientific computing in Python"
 optional = true
 python-versions = ">=3.9"
 files = [
-    {file = "scipy-1.13.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:ba419578ab343a4e0a77c0ef82f088238a93eef141b2b8017e46149776dfad4d"},
-    {file = "scipy-1.13.0-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:22789b56a999265431c417d462e5b7f2b487e831ca7bef5edeb56efe4c93f86e"},
-    {file = "scipy-1.13.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:05f1432ba070e90d42d7fd836462c50bf98bd08bed0aa616c359eed8a04e3922"},
-    {file = "scipy-1.13.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b8434f6f3fa49f631fae84afee424e2483289dfc30a47755b4b4e6b07b2633a4"},
-    {file = "scipy-1.13.0-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:dcbb9ea49b0167de4167c40eeee6e167caeef11effb0670b554d10b1e693a8b9"},
-    {file = "scipy-1.13.0-cp310-cp310-win_amd64.whl", hash = "sha256:1d2f7bb14c178f8b13ebae93f67e42b0a6b0fc50eba1cd8021c9b6e08e8fb1cd"},
-    {file = "scipy-1.13.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:0fbcf8abaf5aa2dc8d6400566c1a727aed338b5fe880cde64907596a89d576fa"},
-    {file = "scipy-1.13.0-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:5e4a756355522eb60fcd61f8372ac2549073c8788f6114449b37e9e8104f15a5"},
-    {file = "scipy-1.13.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b5acd8e1dbd8dbe38d0004b1497019b2dbbc3d70691e65d69615f8a7292865d7"},
-    {file = "scipy-1.13.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9ff7dad5d24a8045d836671e082a490848e8639cabb3dbdacb29f943a678683d"},
-    {file = "scipy-1.13.0-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:4dca18c3ffee287ddd3bc8f1dabaf45f5305c5afc9f8ab9cbfab855e70b2df5c"},
-    {file = "scipy-1.13.0-cp311-cp311-win_amd64.whl", hash = "sha256:a2f471de4d01200718b2b8927f7d76b5d9bde18047ea0fa8bd15c5ba3f26a1d6"},
-    {file = "scipy-1.13.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:d0de696f589681c2802f9090fff730c218f7c51ff49bf252b6a97ec4a5d19e8b"},
-    {file = "scipy-1.13.0-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:b2a3ff461ec4756b7e8e42e1c681077349a038f0686132d623fa404c0bee2551"},
-    {file = "scipy-1.13.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6bf9fe63e7a4bf01d3645b13ff2aa6dea023d38993f42aaac81a18b1bda7a82a"},
-    {file = "scipy-1.13.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:1e7626dfd91cdea5714f343ce1176b6c4745155d234f1033584154f60ef1ff42"},
-    {file = "scipy-1.13.0-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:109d391d720fcebf2fbe008621952b08e52907cf4c8c7efc7376822151820820"},
-    {file = "scipy-1.13.0-cp312-cp312-win_amd64.whl", hash = "sha256:8930ae3ea371d6b91c203b1032b9600d69c568e537b7988a3073dfe4d4774f21"},
-    {file = "scipy-1.13.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:5407708195cb38d70fd2d6bb04b1b9dd5c92297d86e9f9daae1576bd9e06f602"},
-    {file = "scipy-1.13.0-cp39-cp39-macosx_12_0_arm64.whl", hash = "sha256:ac38c4c92951ac0f729c4c48c9e13eb3675d9986cc0c83943784d7390d540c78"},
-    {file = "scipy-1.13.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:09c74543c4fbeb67af6ce457f6a6a28e5d3739a87f62412e4a16e46f164f0ae5"},
-    {file = "scipy-1.13.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:28e286bf9ac422d6beb559bc61312c348ca9b0f0dae0d7c5afde7f722d6ea13d"},
-    {file = "scipy-1.13.0-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:33fde20efc380bd23a78a4d26d59fc8704e9b5fd9b08841693eb46716ba13d86"},
-    {file = "scipy-1.13.0-cp39-cp39-win_amd64.whl", hash = "sha256:45c08bec71d3546d606989ba6e7daa6f0992918171e2a6f7fbedfa7361c2de1e"},
-    {file = "scipy-1.13.0.tar.gz", hash = "sha256:58569af537ea29d3f78e5abd18398459f195546bb3be23d16677fb26616cc11e"},
+    {file = "scipy-1.13.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:20335853b85e9a49ff7572ab453794298bcf0354d8068c5f6775a0eabf350aca"},
+    {file = "scipy-1.13.1-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:d605e9c23906d1994f55ace80e0125c587f96c020037ea6aa98d01b4bd2e222f"},
+    {file = "scipy-1.13.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cfa31f1def5c819b19ecc3a8b52d28ffdcc7ed52bb20c9a7589669dd3c250989"},
+    {file = "scipy-1.13.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f26264b282b9da0952a024ae34710c2aff7d27480ee91a2e82b7b7073c24722f"},
+    {file = "scipy-1.13.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:eccfa1906eacc02de42d70ef4aecea45415f5be17e72b61bafcfd329bdc52e94"},
+    {file = "scipy-1.13.1-cp310-cp310-win_amd64.whl", hash = "sha256:2831f0dc9c5ea9edd6e51e6e769b655f08ec6db6e2e10f86ef39bd32eb11da54"},
+    {file = "scipy-1.13.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:27e52b09c0d3a1d5b63e1105f24177e544a222b43611aaf5bc44d4a0979e32f9"},
+    {file = "scipy-1.13.1-cp311-cp311-macosx_12_0_arm64.whl", hash = "sha256:54f430b00f0133e2224c3ba42b805bfd0086fe488835effa33fa291561932326"},
+    {file = "scipy-1.13.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e89369d27f9e7b0884ae559a3a956e77c02114cc60a6058b4e5011572eea9299"},
+    {file = "scipy-1.13.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a78b4b3345f1b6f68a763c6e25c0c9a23a9fd0f39f5f3d200efe8feda560a5fa"},
+    {file = "scipy-1.13.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:45484bee6d65633752c490404513b9ef02475b4284c4cfab0ef946def50b3f59"},
+    {file = "scipy-1.13.1-cp311-cp311-win_amd64.whl", hash = "sha256:5713f62f781eebd8d597eb3f88b8bf9274e79eeabf63afb4a737abc6c84ad37b"},
+    {file = "scipy-1.13.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:5d72782f39716b2b3509cd7c33cdc08c96f2f4d2b06d51e52fb45a19ca0c86a1"},
+    {file = "scipy-1.13.1-cp312-cp312-macosx_12_0_arm64.whl", hash = "sha256:017367484ce5498445aade74b1d5ab377acdc65e27095155e448c88497755a5d"},
+    {file = "scipy-1.13.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:949ae67db5fa78a86e8fa644b9a6b07252f449dcf74247108c50e1d20d2b4627"},
+    {file = "scipy-1.13.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:de3ade0e53bc1f21358aa74ff4830235d716211d7d077e340c7349bc3542e884"},
+    {file = "scipy-1.13.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:2ac65fb503dad64218c228e2dc2d0a0193f7904747db43014645ae139c8fad16"},
+    {file = "scipy-1.13.1-cp312-cp312-win_amd64.whl", hash = "sha256:cdd7dacfb95fea358916410ec61bbc20440f7860333aee6d882bb8046264e949"},
+    {file = "scipy-1.13.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:436bbb42a94a8aeef855d755ce5a465479c721e9d684de76bf61a62e7c2b81d5"},
+    {file = "scipy-1.13.1-cp39-cp39-macosx_12_0_arm64.whl", hash = "sha256:8335549ebbca860c52bf3d02f80784e91a004b71b059e3eea9678ba994796a24"},
+    {file = "scipy-1.13.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d533654b7d221a6a97304ab63c41c96473ff04459e404b83275b60aa8f4b7004"},
+    {file = "scipy-1.13.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:637e98dcf185ba7f8e663e122ebf908c4702420477ae52a04f9908707456ba4d"},
+    {file = "scipy-1.13.1-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:a014c2b3697bde71724244f63de2476925596c24285c7a637364761f8710891c"},
+    {file = "scipy-1.13.1-cp39-cp39-win_amd64.whl", hash = "sha256:392e4ec766654852c25ebad4f64e4e584cf19820b980bc04960bca0b0cd6eaa2"},
+    {file = "scipy-1.13.1.tar.gz", hash = "sha256:095a87a0312b08dfd6a6155cbbd310a8c51800fc931b8c0b84003014b874ed3c"},
 ]
 
 [package.dependencies]
@@ -3449,13 +3481,13 @@ test = ["array-api-strict", "asv", "gmpy2", "hypothesis (>=6.30)", "mpmath", "po
 
 [[package]]
 name = "sentry-sdk"
-version = "2.2.1"
+version = "2.3.1"
 description = "Python client for Sentry (https://sentry.io)"
 optional = false
 python-versions = ">=3.6"
 files = [
-    {file = "sentry_sdk-2.2.1-py2.py3-none-any.whl", hash = "sha256:7d617a1b30e80c41f3b542347651fcf90bb0a36f3a398be58b4f06b79c8d85bc"},
-    {file = "sentry_sdk-2.2.1.tar.gz", hash = "sha256:8aa2ec825724d8d9d645cab68e6034928b1a6a148503af3e361db3fa6401183f"},
+    {file = "sentry_sdk-2.3.1-py2.py3-none-any.whl", hash = "sha256:c5aeb095ba226391d337dd42a6f9470d86c9fc236ecc71cfc7cd1942b45010c6"},
+    {file = "sentry_sdk-2.3.1.tar.gz", hash = "sha256:139a71a19f5e9eb5d3623942491ce03cf8ebc14ea2e39ba3e6fe79560d8a5b1f"},
 ]
 
 [package.dependencies]
@@ -3704,17 +3736,17 @@ files = [
 
 [[package]]
 name = "sympy"
-version = "1.12"
+version = "1.12.1"
 description = "Computer algebra system (CAS) in Python"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "sympy-1.12-py3-none-any.whl", hash = "sha256:c3588cd4295d0c0f603d0f2ae780587e64e2efeedb3521e46b9bb1d08d184fa5"},
-    {file = "sympy-1.12.tar.gz", hash = "sha256:ebf595c8dac3e0fdc4152c51878b498396ec7f30e7a914d6071e674d49420fb8"},
+    {file = "sympy-1.12.1-py3-none-any.whl", hash = "sha256:9b2cbc7f1a640289430e13d2a56f02f867a1da0190f2f99d8968c2f74da0e515"},
+    {file = "sympy-1.12.1.tar.gz", hash = "sha256:2877b03f998cd8c08f07cd0de5b767119cd3ef40d09f41c30d722f6686b0fb88"},
 ]
 
 [package.dependencies]
-mpmath = ">=0.19"
+mpmath = ">=1.1.0,<1.4.0"
 
 [[package]]
 name = "tbb"
@@ -3745,13 +3777,13 @@ tests = ["pytest", "pytest-cov"]
 
 [[package]]
 name = "tifffile"
-version = "2024.5.10"
+version = "2024.5.22"
 description = "Read and write TIFF files"
 optional = true
 python-versions = ">=3.9"
 files = [
-    {file = "tifffile-2024.5.10-py3-none-any.whl", hash = "sha256:4154f091aa24d4e75bfad9ab2d5424a68c70e67b8220188066dc61946d4551bd"},
-    {file = "tifffile-2024.5.10.tar.gz", hash = "sha256:aa1e1b12be952ab20717d6848bd6d4a5ee88d2aa319f1152bff4354ad728ec86"},
+    {file = "tifffile-2024.5.22-py3-none-any.whl", hash = "sha256:e281781c15d7d197d7e12749849c965651413aa905f97a48b0f84bd90a3b4c6f"},
+    {file = "tifffile-2024.5.22.tar.gz", hash = "sha256:3a105801d1b86d55692a98812a170c39d3f0447aeacb1d94635d38077cb328c4"},
 ]
 
 [package.dependencies]
@@ -3907,13 +3939,13 @@ tutorials = ["matplotlib", "pandas", "tabulate", "torch"]
 
 [[package]]
 name = "typing-extensions"
-version = "4.11.0"
+version = "4.12.0"
 description = "Backported and Experimental Type Hints for Python 3.8+"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "typing_extensions-4.11.0-py3-none-any.whl", hash = "sha256:c1f94d72897edaf4ce775bb7558d5b79d8126906a14ea5ed1635921406c0387a"},
-    {file = "typing_extensions-4.11.0.tar.gz", hash = "sha256:83f085bd5ca59c80295fc2a82ab5dac679cbe02b9f33f7d83af68e241bea51b0"},
+    {file = "typing_extensions-4.12.0-py3-none-any.whl", hash = "sha256:b349c66bea9016ac22978d800cfff206d5f9816951f12a7d0ec5578b0a819594"},
+    {file = "typing_extensions-4.12.0.tar.gz", hash = "sha256:8cbcdc8606ebcb0d95453ad7dc5065e6237b6aa230a31e81d0f440c30fed5fd8"},
 ]
 
 [[package]]
@@ -4228,13 +4260,13 @@ multidict = ">=4.0"
 
 [[package]]
 name = "zarr"
-version = "2.18.1"
+version = "2.18.2"
 description = "An implementation of chunked, compressed, N-dimensional arrays for Python"
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "zarr-2.18.1-py3-none-any.whl", hash = "sha256:a1770d194eec4ec0a41a01295a6f724e1c3471d704d3aca906d3b3a7f8830245"},
-    {file = "zarr-2.18.1.tar.gz", hash = "sha256:28c360ed123e606c425a694a83300227a907cb86a995fc9eef620ecafbe5f92d"},
+    {file = "zarr-2.18.2-py3-none-any.whl", hash = "sha256:a638754902f97efa99b406083fdc807a0e2ccf12a949117389d2a4ba9b05df38"},
+    {file = "zarr-2.18.2.tar.gz", hash = "sha256:9bb393b8a0a38fb121dbb913b047d75db28de9890f6d644a217a73cf4ae74f47"},
 ]
 
 [package.dependencies]
@@ -4249,13 +4281,13 @@ jupyter = ["ipytree (>=0.2.2)", "ipywidgets (>=8.0.0)", "notebook"]
 
 [[package]]
 name = "zipp"
-version = "3.18.2"
+version = "3.19.0"
 description = "Backport of pathlib-compatible object wrapper for zip files"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "zipp-3.18.2-py3-none-any.whl", hash = "sha256:dce197b859eb796242b0622af1b8beb0a722d52aa2f57133ead08edd5bf5374e"},
-    {file = "zipp-3.18.2.tar.gz", hash = "sha256:6278d9ddbcfb1f1089a88fde84481528b07b0e10474e09dcfe53dad4069fa059"},
+    {file = "zipp-3.19.0-py3-none-any.whl", hash = "sha256:96dc6ad62f1441bcaccef23b274ec471518daf4fbbc580341204936a5a3dddec"},
+    {file = "zipp-3.19.0.tar.gz", hash = "sha256:952df858fb3164426c976d9338d3961e8e8b3758e2e059e0f754b8c4262625ee"},
 ]
 
 [package.extras]
@@ -4274,4 +4306,4 @@ xarm = ["gym-xarm"]
 [metadata]
 lock-version = "2.0"
 python-versions = ">=3.10,<3.13"
-content-hash = "ea4e8207316a8ec8a4b95d6a89cf488c8733a8e7ab43e5f669c889ee87f3bef3"
+content-hash = "23ddb8dd774a4faf85d08a07dfdf19badb7c370120834b71df4afca254520771"

pyproject.toml

@@ -41,12 +41,12 @@ numba = ">=0.59.0"
 torch = "^2.2.1"
 opencv-python = ">=4.9.0"
 diffusers = "^0.27.2"
-torchvision = ">=0.18.0"
+torchvision = ">=0.17.1"
 h5py = ">=3.10.0"
 huggingface-hub = {extras = ["hf-transfer"], version = "^0.23.0"}
 gymnasium = ">=0.29.1"
 cmake = ">=3.29.0.1"
-gym-dora = { path = "gym_dora", optional = true, develop = true}
+gym-dora = { git = "https://github.com/dora-rs/dora-lerobot.git", subdirectory = "gym_dora", optional = true }
 gym-pusht = { version = ">=0.1.3", optional = true}
 gym-xarm = { version = ">=0.1.1", optional = true}
 gym-aloha = { version = ">=0.1.1", optional = true}
@@ -59,6 +59,7 @@ imagecodecs = { version = ">=2024.1.1", optional = true }
 pyav = ">=12.0.5"
 moviepy = ">=1.0.3"
 rerun-sdk = ">=0.15.1"
+deepdiff = ">=7.0.1"
 
 
 [tool.poetry.extras]

tests/data/save_policy_to_safetensors/dora_aloha_real_act_real/actions.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fff6294b94cf42d4dd1249dcc5c3b0269d6d9c697f894e61b867d7ab81a94e4
+size 5104

tests/data/save_policy_to_safetensors/dora_aloha_real_act_real/grad_stats.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4aa23e51607604a18b70fa42edbbe1af34f119d985628fc27cc1bbb0efbc8901
+size 31688

tests/data/save_policy_to_safetensors/dora_aloha_real_act_real/output_dict.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fd368406c93cb562a69ff11cf7adf34a4b223507dcb2b9e9b8f44ee1036988a
+size 68

tests/data/save_policy_to_safetensors/dora_aloha_real_act_real/param_stats.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5663ee79a13bb70a1604b887dd21bf89d18482287442419c6cc6c5bf0e753e99
+size 34928

tests/data/save_policy_to_safetensors/dora_aloha_real_act_real_no_state/actions.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb1a45463efd860af2ca22c16c77d55a18bd96fef080ae77978845a2f22ef716
+size 5104

tests/data/save_policy_to_safetensors/dora_aloha_real_act_real_no_state/grad_stats.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa5a43e22f01d8e2f8d19f31753608794f1edbd74aaf71660091ab80ea58dc9b
+size 30808

tests/data/save_policy_to_safetensors/dora_aloha_real_act_real_no_state/output_dict.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97455b4360748c99905cd103473c1a52da6901d0a73ffbc51b5ea3eb250d1386
+size 68

tests/data/save_policy_to_safetensors/dora_aloha_real_act_real_no_state/param_stats.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54d1f75cf67a7b1d7a7c6865ecb9b1cc86a2f032d1890245f8996789ab6e0df6
+size 33608

tests/scripts/save_policy_to_safetensors.py

@@ -75,15 +75,16 @@ def get_policy_stats(env_name, policy_name, extra_overrides):
     # HACK: We reload a batch with no delta_timestamps as `select_action` won't expect a timestamps dimension
     dataset.delta_timestamps = None
     batch = next(iter(dataloader))
-    obs = {
-        k: batch[k]
-        for k in batch
-        if k in ["observation.image", "observation.images.top", "observation.state"]
-    }
+    obs = {}
+    for k in batch:
+        if k.startswith("observation"):
+            obs[k] = batch[k]
+
+    if "n_action_steps" in cfg.policy:
+        actions_queue = cfg.policy.n_action_steps
+    else:
+        actions_queue = cfg.policy.n_action_repeats
 
-    actions_queue = (
-        cfg.policy.n_action_steps if "n_action_steps" in cfg.policy else cfg.policy.n_action_repeats
-    )
     actions = {str(i): policy.select_action(obs).contiguous() for i in range(actions_queue)}
     return output_dict, grad_stats, param_stats, actions
 
@@ -114,6 +115,8 @@ if __name__ == "__main__":
             ["policy.n_action_steps=8", "policy.num_inference_steps=10", "policy.down_dims=[128, 256, 512]"],
         ),
         ("aloha", "act", ["policy.n_action_steps=10"]),
+        ("dora_aloha_real", "act_real", ["policy.n_action_steps=10"]),
+        ("dora_aloha_real", "act_real_no_state", ["policy.n_action_steps=10"]),
     ]
     for env, policy, extra_overrides in env_policies:
         save_policy_to_safetensors("tests/data/save_policy_to_safetensors", env, policy, extra_overrides)

tests/test_datasets.py

@@ -16,6 +16,7 @@
 import json
 import logging
 from copy import deepcopy
+from itertools import chain
 from pathlib import Path
 
 import einops
@@ -25,26 +26,34 @@ from datasets import Dataset
 from safetensors.torch import load_file
 
 import lerobot
-from lerobot.common.datasets.factory import make_dataset
-from lerobot.common.datasets.lerobot_dataset import (
-    LeRobotDataset,
-)
-from lerobot.common.datasets.push_dataset_to_hub.compute_stats import (
+from lerobot.common.datasets.compute_stats import (
+    aggregate_stats,
     compute_stats,
     get_stats_einops_patterns,
 )
+from lerobot.common.datasets.factory import make_dataset
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset, MultiLeRobotDataset
 from lerobot.common.datasets.utils import (
     flatten_dict,
     hf_transform_to_torch,
     load_previous_and_future_frames,
     unflatten_dict,
 )
-from lerobot.common.utils.utils import init_hydra_config
+from lerobot.common.utils.utils import init_hydra_config, seeded_context
 from tests.utils import DEFAULT_CONFIG_PATH, DEVICE
 
 
-@pytest.mark.parametrize("env_name, repo_id, policy_name", lerobot.env_dataset_policy_triplets)
+@pytest.mark.parametrize(
+    "env_name, repo_id, policy_name",
+    lerobot.env_dataset_policy_triplets
+    + [("aloha", ["lerobot/aloha_sim_insertion_human", "lerobot/aloha_sim_transfer_cube_human"], "act")],
+)
 def test_factory(env_name, repo_id, policy_name):
+    """
+    Tests that:
+        - we can create a dataset with the factory.
+        - for a commonly used set of data keys, the data dimensions are correct.
+    """
     cfg = init_hydra_config(
         DEFAULT_CONFIG_PATH,
         overrides=[
@@ -105,6 +114,39 @@ def test_factory(env_name, repo_id, policy_name):
             assert key in item, f"{key}"
 
 
+# TODO(alexander-soare): If you're hunting for savings on testing time, this takes about 5 seconds.
+def test_multilerobotdataset_frames():
+    """Check that all dataset frames are incorporated."""
+    # Note: use the image variants of the dataset to make the test approx 3x faster.
+    # Note: We really do need three repo_ids here as at some point this caught an issue with the chaining
+    # logic that wouldn't be caught with two repo IDs.
+    repo_ids = [
+        "lerobot/aloha_sim_insertion_human_image",
+        "lerobot/aloha_sim_transfer_cube_human_image",
+        "lerobot/aloha_sim_insertion_scripted_image",
+    ]
+    sub_datasets = [LeRobotDataset(repo_id) for repo_id in repo_ids]
+    dataset = MultiLeRobotDataset(repo_ids)
+    assert len(dataset) == sum(len(d) for d in sub_datasets)
+    assert dataset.num_samples == sum(d.num_samples for d in sub_datasets)
+    assert dataset.num_episodes == sum(d.num_episodes for d in sub_datasets)
+
+    # Run through all items of the LeRobotDatasets in parallel with the items of the MultiLerobotDataset and
+    # check they match.
+    expected_dataset_indices = []
+    for i, sub_dataset in enumerate(sub_datasets):
+        expected_dataset_indices.extend([i] * len(sub_dataset))
+
+    for expected_dataset_index, sub_dataset_item, dataset_item in zip(
+        expected_dataset_indices, chain(*sub_datasets), dataset, strict=True
+    ):
+        dataset_index = dataset_item.pop("dataset_index")
+        assert dataset_index == expected_dataset_index
+        assert sub_dataset_item.keys() == dataset_item.keys()
+        for k in sub_dataset_item:
+            assert torch.equal(sub_dataset_item[k], dataset_item[k])
+
+
 def test_compute_stats_on_xarm():
     """Check that the statistics are computed correctly according to the stats_patterns property.
 
@@ -315,3 +357,31 @@ def test_backward_compatibility(repo_id):
     # i = dataset.episode_data_index["to"][-1].item()
     # load_and_compare(i - 2)
     # load_and_compare(i - 1)
+
+
+def test_aggregate_stats():
+    """Makes 3 basic datasets and checks that aggregate stats are computed correctly."""
+    with seeded_context(0):
+        data_a = torch.rand(30, dtype=torch.float32)
+        data_b = torch.rand(20, dtype=torch.float32)
+        data_c = torch.rand(20, dtype=torch.float32)
+
+    hf_dataset_1 = Dataset.from_dict(
+        {"a": data_a[:10], "b": data_b[:10], "c": data_c[:10], "index": torch.arange(10)}
+    )
+    hf_dataset_1.set_transform(hf_transform_to_torch)
+    hf_dataset_2 = Dataset.from_dict({"a": data_a[10:20], "b": data_b[10:], "index": torch.arange(10)})
+    hf_dataset_2.set_transform(hf_transform_to_torch)
+    hf_dataset_3 = Dataset.from_dict({"a": data_a[20:], "c": data_c[10:], "index": torch.arange(10)})
+    hf_dataset_3.set_transform(hf_transform_to_torch)
+    dataset_1 = LeRobotDataset.from_preloaded("d1", hf_dataset=hf_dataset_1)
+    dataset_1.stats = compute_stats(dataset_1, batch_size=len(hf_dataset_1), num_workers=0)
+    dataset_2 = LeRobotDataset.from_preloaded("d2", hf_dataset=hf_dataset_2)
+    dataset_2.stats = compute_stats(dataset_2, batch_size=len(hf_dataset_2), num_workers=0)
+    dataset_3 = LeRobotDataset.from_preloaded("d3", hf_dataset=hf_dataset_3)
+    dataset_3.stats = compute_stats(dataset_3, batch_size=len(hf_dataset_3), num_workers=0)
+    stats = aggregate_stats([dataset_1, dataset_2, dataset_3])
+    for data_key, data in zip(["a", "b", "c"], [data_a, data_b, data_c], strict=True):
+        for agg_fn in ["mean", "min", "max"]:
+            assert torch.allclose(stats[data_key][agg_fn], einops.reduce(data, "n -> 1", agg_fn))
+        assert torch.allclose(stats[data_key]["std"], torch.std(data, correction=0))

tests/test_examples.py

@@ -45,11 +45,11 @@ def test_example_1():
 
 
 @require_package("gym_pusht")
-def test_examples_2_through_4():
+def test_examples_basic2_basic3_advanced1():
     """
     Train a model with example 3, check the outputs.
     Evaluate the trained model with example 2, check the outputs.
-    Calculate the validation loss with example 4, check the outputs.
+    Calculate the validation loss with advanced example 1, check the outputs.
     """
 
     ### Test example 3
@@ -97,7 +97,7 @@ def test_examples_2_through_4():
     assert Path("outputs/eval/example_pusht_diffusion/rollout.mp4").exists()
 
     ## Test example 4
-    file_contents = _read_file("examples/4_calculate_validation_loss.py")
+    file_contents = _read_file("examples/advanced/2_calculate_validation_loss.py")
 
     # Run on a single example from the last episode, use CPU, and use the local model.
     file_contents = _find_and_replace(

tests/test_policies.py

@@ -30,7 +30,7 @@ from lerobot.common.policies.factory import get_policy_and_config_classes, make_
 from lerobot.common.policies.normalize import Normalize, Unnormalize
 from lerobot.common.policies.policy_protocol import Policy
 from lerobot.common.utils.utils import init_hydra_config
-from tests.scripts.save_policy_to_safetensor import get_policy_stats
+from tests.scripts.save_policy_to_safetensors import get_policy_stats
 from tests.utils import DEFAULT_CONFIG_PATH, DEVICE, require_cpu, require_env, require_x86_64_kernel
 
 
@@ -72,6 +72,8 @@ def test_get_policy_and_config_classes(policy_name: str):
         ),
         # Note: these parameters also need custom logic in the test function for overriding the Hydra config.
         ("pusht", "act", ["env.task=PushT-v0", "dataset_repo_id=lerobot/pusht"]),
+        ("dora_aloha_real", "act_real", []),
+        ("dora_aloha_real", "act_real_no_state", []),
     ],
 )
 @require_env
@@ -84,6 +86,9 @@ def test_policy(env_name, policy_name, extra_overrides):
         - Updating the policy.
         - Using the policy to select actions at inference time.
         - Test the action can be applied to the policy
+
+    Note: We test various combinations of policy and dataset. The combinations are by no means exhaustive,
+          and for now we add tests as we see fit.
     """
     cfg = init_hydra_config(
         DEFAULT_CONFIG_PATH,
@@ -135,7 +140,7 @@ def test_policy(env_name, policy_name, extra_overrides):
 
     dataloader = torch.utils.data.DataLoader(
         dataset,
-        num_workers=4,
+        num_workers=0,
         batch_size=2,
         shuffle=True,
         pin_memory=DEVICE != "cpu",
@@ -291,6 +296,8 @@ def test_normalize(insert_temporal_dim):
             ["policy.n_action_steps=8", "policy.num_inference_steps=10", "policy.down_dims=[128, 256, 512]"],
         ),
         ("aloha", "act", ["policy.n_action_steps=10"]),
+        ("dora_aloha_real", "act_real", ["policy.n_action_steps=10"]),
+        ("dora_aloha_real", "act_real_no_state", ["policy.n_action_steps=10"]),
     ],
 )
 # As artifacts have been generated on an x86_64 kernel, this test won't

tests/test_sampler.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from datasets import Dataset
+
+from lerobot.common.datasets.sampler import EpisodeAwareSampler
+from lerobot.common.datasets.utils import (
+    calculate_episode_data_index,
+    hf_transform_to_torch,
+)
+
+
+def test_drop_n_first_frames():
+    dataset = Dataset.from_dict(
+        {
+            "timestamp": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6],
+            "index": [0, 1, 2, 3, 4, 5],
+            "episode_index": [0, 0, 1, 2, 2, 2],
+        },
+    )
+    dataset.set_transform(hf_transform_to_torch)
+    episode_data_index = calculate_episode_data_index(dataset)
+    sampler = EpisodeAwareSampler(episode_data_index, drop_n_first_frames=1)
+    assert sampler.indices == [1, 4, 5]
+    assert len(sampler) == 3
+    assert list(sampler) == [1, 4, 5]
+
+
+def test_drop_n_last_frames():
+    dataset = Dataset.from_dict(
+        {
+            "timestamp": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6],
+            "index": [0, 1, 2, 3, 4, 5],
+            "episode_index": [0, 0, 1, 2, 2, 2],
+        },
+    )
+    dataset.set_transform(hf_transform_to_torch)
+    episode_data_index = calculate_episode_data_index(dataset)
+    sampler = EpisodeAwareSampler(episode_data_index, drop_n_last_frames=1)
+    assert sampler.indices == [0, 3, 4]
+    assert len(sampler) == 3
+    assert list(sampler) == [0, 3, 4]
+
+
+def test_episode_indices_to_use():
+    dataset = Dataset.from_dict(
+        {
+            "timestamp": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6],
+            "index": [0, 1, 2, 3, 4, 5],
+            "episode_index": [0, 0, 1, 2, 2, 2],
+        },
+    )
+    dataset.set_transform(hf_transform_to_torch)
+    episode_data_index = calculate_episode_data_index(dataset)
+    sampler = EpisodeAwareSampler(episode_data_index, episode_indices_to_use=[0, 2])
+    assert sampler.indices == [0, 1, 3, 4, 5]
+    assert len(sampler) == 5
+    assert list(sampler) == [0, 1, 3, 4, 5]
+
+
+def test_shuffle():
+    dataset = Dataset.from_dict(
+        {
+            "timestamp": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6],
+            "index": [0, 1, 2, 3, 4, 5],
+            "episode_index": [0, 0, 1, 2, 2, 2],
+        },
+    )
+    dataset.set_transform(hf_transform_to_torch)
+    episode_data_index = calculate_episode_data_index(dataset)
+    sampler = EpisodeAwareSampler(episode_data_index, shuffle=False)
+    assert sampler.indices == [0, 1, 2, 3, 4, 5]
+    assert len(sampler) == 6
+    assert list(sampler) == [0, 1, 2, 3, 4, 5]
+    sampler = EpisodeAwareSampler(episode_data_index, shuffle=True)
+    assert sampler.indices == [0, 1, 2, 3, 4, 5]
+    assert len(sampler) == 6
+    assert set(sampler) == {0, 1, 2, 3, 4, 5}

tests/test_utils.py

@@ -11,22 +11,24 @@ from lerobot.common.datasets.utils import (
     hf_transform_to_torch,
     reset_episode_index,
 )
-from lerobot.common.utils.utils import seeded_context, set_global_seed
-
-
-@pytest.mark.parametrize(
-    "rand_fn",
-    (
-        [
-            random.random,
-            np.random.random,
-            lambda: torch.rand(1).item(),
-        ]
-        + [lambda: torch.rand(1, device="cuda")]
-        if torch.cuda.is_available()
-        else []
-    ),
+from lerobot.common.utils.utils import (
+    get_global_random_state,
+    seeded_context,
+    set_global_random_state,
+    set_global_seed,
 )
+
+# Random generation functions for testing the seeding and random state get/set.
+rand_fns = [
+    random.random,
+    np.random.random,
+    lambda: torch.rand(1).item(),
+]
+if torch.cuda.is_available():
+    rand_fns.append(lambda: torch.rand(1, device="cuda"))
+
+
+@pytest.mark.parametrize("rand_fn", rand_fns)
 def test_seeding(rand_fn: Callable[[], int]):
     set_global_seed(0)
     a = rand_fn()
@@ -46,6 +48,15 @@ def test_seeding(rand_fn: Callable[[], int]):
     assert c_ == c
 
 
+def test_get_set_random_state():
+    """Check that getting the random state, then setting it results in the same random number generation."""
+    random_state_dict = get_global_random_state()
+    rand_numbers = [rand_fn() for rand_fn in rand_fns]
+    set_global_random_state(random_state_dict)
+    rand_numbers_ = [rand_fn() for rand_fn in rand_fns]
+    assert rand_numbers_ == rand_numbers
+
+
 def test_calculate_episode_data_index():
     dataset = Dataset.from_dict(
         {

tests/test_visualize_dataset.py

@@ -25,9 +25,8 @@ from lerobot.scripts.visualize_dataset import visualize_dataset
 def test_visualize_dataset(tmpdir, repo_id):
     rrd_path = visualize_dataset(
         repo_id,
-        episode_index=0,
-        batch_size=32,
-        save=True,
+        episode_indices=[0],
         output_dir=tmpdir,
+        serve=False,
     )
     assert rrd_path.exists()