fix aloha conversion changes

Co-authored-by: Simon Alibert <75076266+aliberts@users.noreply.github.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 .
@@ -18,15 +19,16 @@ 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-act-ete-train-amp
-	${MAKE} test-act-ete-eval-amp
-	${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 \
@@ -38,8 +40,8 @@ test-act-ete-train:
 		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 \
@@ -48,11 +50,11 @@ 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 \
@@ -64,22 +66,22 @@ test-act-ete-train-amp:
 		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 \
 		training.batch_size=2 \
-		hydra.run.dir=tests/outputs/act/ \
+		hydra.run.dir=tests/outputs/act_amp/ \
 		use_amp=true
 
 test-act-ete-eval-amp:
 	python lerobot/scripts/eval.py \
-		-p tests/outputs/act/checkpoints/000002 \
+		-p tests/outputs/act_amp/checkpoints/000002/pretrained_model \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=8 \
-		device=cpu \
+		device=$(DEVICE) \
 		use_amp=true
 
 test-diffusion-ete-train:
@@ -94,19 +96,19 @@ test-diffusion-ete-train:
 		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:
@@ -121,19 +123,19 @@ test-tdmpc-ete-train:
 		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 \
@@ -141,4 +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/pusht.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 / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    observation.image: [3, 96, 96]
+    observation.state: ["${env.state_dim}"]
+  output_shapes:
+    action: ["${env.action_dim}"]
+
+  # Normalization / Unnormalization
+  input_normalization_modes:
+    observation.image: mean_std
+    # Use min_max normalization just because it's more standard.
+    observation.state: min_max
+  output_normalization_modes:
+    # Use min_max normalization just because it's more standard.
+    action: min_max
+
+  # Architecture.
+  # Vision backbone.
+  vision_backbone: resnet18
+  pretrained_backbone_weights: ResNet18_Weights.IMAGENET1K_V1
+  replace_final_stride_with_dilation: false
+  # Transformer layers.
+  pre_norm: false
+  dim_model: 512
+  n_heads: 8
+  dim_feedforward: 3200
+  feedforward_activation: relu
+  n_encoder_layers: 4
+    # Note: Although the original ACT implementation has 7 for `n_decoder_layers`, there is a bug in the code
+  # that means only the first layer is used. Here we match the original implementation by setting this to 1.
+  # See this issue https://github.com/tonyzhaozh/act/issues/25#issue-2258740521.
+  n_decoder_layers: 1
+  # VAE.
+  use_vae: true
+  latent_dim: 32
+  n_vae_encoder_layers: 4
+
+  # Inference.
+  temporal_ensemble_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! 🤗

lerobot/__init__.py

@@ -52,6 +52,7 @@ available_tasks_per_env = {
     ],
     "pusht": ["PushT-v0"],
     "xarm": ["XarmLift-v0"],
+    "dora": ["DoraAloha-v0", "DoraKoch-v0", "DoraReachy2-v0"],
 }
 available_envs = list(available_tasks_per_env.keys())
 
@@ -77,6 +78,23 @@ available_datasets_per_env = {
         "lerobot/xarm_push_medium_image",
         "lerobot/xarm_push_medium_replay_image",
     ],
+    "dora": [
+        "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 = [
@@ -116,6 +134,7 @@ available_policies = [
 
 available_policies_per_env = {
     "aloha": ["act"],
+    "dora": ["act"],
     "pusht": ["diffusion"],
     "xarm": ["tdmpc"],
 }

lerobot/common/datasets/factory.py

@@ -21,6 +21,20 @@ from omegaconf import OmegaConf
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
 
+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):
+                # TODO(rcadene, alexander-soare): remove `eval` to avoid exploit
+                cfg.training.delta_timestamps[key] = eval(delta_timestamps[key])
+
+
 def make_dataset(
     cfg,
     split="train",
@@ -31,18 +45,14 @@ def make_dataset(
             f"environment ({cfg.env.name=})."
         )
 
-    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])
+    resolve_delta_timestamps(cfg)
 
     # TODO(rcadene): add data augmentations
 
     dataset = LeRobotDataset(
         cfg.dataset_repo_id,
         split=split,
-        delta_timestamps=delta_timestamps,
+        delta_timestamps=cfg.training.get("delta_timestamps"),
     )
 
     if cfg.get("override_dataset_stats"):

lerobot/common/datasets/push_dataset_to_hub/aloha_dora_format.py

@@ -0,0 +1,255 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Contains utilities to process raw data format from dora-record
+"""
+
+import logging
+import re
+from pathlib import Path
+
+import pandas as pd
+import torch
+from datasets import Dataset, Features, Image, Sequence, Value
+
+from lerobot.common.datasets.utils import (
+    hf_transform_to_torch,
+)
+from lerobot.common.datasets.video_utils import VideoFrame
+from lerobot.common.utils.utils import init_logging
+
+
+def check_format(raw_dir) -> bool:
+    assert raw_dir.exists()
+
+    leader_file = list(raw_dir.glob("*.parquet"))
+    if len(leader_file) == 0:
+        raise ValueError(f"Missing parquet files in '{raw_dir}'")
+    return True
+
+
+def load_from_raw(raw_dir: Path, 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:
+        raise ValueError(f"Missing reference files for camera, starting with  in '{raw_dir}'")
+    # select first camera in alphanumeric order
+    reference_key = sorted(reference_files)[0].stem
+    reference_df = pd.read_parquet(raw_dir / f"{reference_key}.parquet")
+    reference_df = reference_df[["timestamp_utc", reference_key]]
+
+    # Merge all data stream using nearest backward strategy
+    df = reference_df
+    for path in raw_dir.glob("*.parquet"):
+        key = path.stem  # action or observation.state or ...
+        if key == reference_key:
+            continue
+        if "failed_episode_index" in key:
+            # TODO(rcadene): add support for removing episodes that are tagged as "failed"
+            continue
+        modality_df = pd.read_parquet(path)
+        modality_df = modality_df[["timestamp_utc", key]]
+        df = pd.merge_asof(
+            df,
+            modality_df,
+            on="timestamp_utc",
+            # "nearest" is the best option over "backward", since the latter can desynchronizes camera timestamps by
+            # matching timestamps that are too far appart, in order to fit the backward constraints. It's not the case for "nearest".
+            # However, note that "nearest" might synchronize the reference camera with other cameras on slightly future timestamps.
+            # are too far appart.
+            direction="nearest",
+            tolerance=pd.Timedelta(f"{1/fps} seconds"),
+        )
+    # Remove rows with episode_index -1 which indicates data that correspond to in-between episodes
+    df = df[df["episode_index"] != -1]
+
+    image_keys = [key for key in df if "observation.images." in key]
+
+    num_unaligned_images = 0
+    max_episode = 0
+
+    def get_episode_index(row):
+        nonlocal num_unaligned_images
+        nonlocal max_episode
+        episode_index_per_cam = {}
+        for key in image_keys:
+            if isinstance(row[key], float):
+                num_unaligned_images += 1
+                return float("nan")
+            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 episode_index > max_episode:
+                assert episode_index - max_episode == 1
+                max_episode = episode_index
+            else:
+                assert episode_index == max_episode
+        if len(set(episode_index_per_cam.values())) != 1:
+            raise ValueError(
+                f"All cameras are expected to belong to the same episode, but getting {episode_index_per_cam}"
+            )
+        return episode_index
+
+    df["episode_index"] = df.apply(get_episode_index, axis=1)
+
+    # dora only use arrays, so single values are encapsulated into a list
+    df["frame_index"] = df.groupby("episode_index").cumcount()
+    df = df.reset_index()
+    df["index"] = df.index
+
+    # set 'next.done' to True for the last frame of each episode
+    df["next.done"] = False
+    df.loc[df.groupby("episode_index").tail(1).index, "next.done"] = True
+
+    df["timestamp"] = df["timestamp_utc"].map(lambda x: x.timestamp())
+    # each episode starts with timestamp 0 to match the ones from the video
+    df["timestamp"] = df.groupby("episode_index")["timestamp"].transform(lambda x: x - x.iloc[0])
+
+    del df["timestamp_utc"]
+
+    # sanity check
+    num_rows_with_nan = df.isna().any(axis=1).sum()
+    assert (
+        num_rows_with_nan == num_unaligned_images
+    ), f"Found {num_rows_with_nan} rows with NaN values but {num_unaligned_images} unaligned images."
+    if num_unaligned_images > max_episode * 2:
+        # We allow a few unaligned images, typically at the beginning and end of the episodes for instance
+        # but if there are too many, we raise an error to avoid large chunks of missing data
+        raise ValueError(
+            f"Found {num_unaligned_images} unaligned images out of {max_episode} episodes. "
+            f"Check the timestamps of the cameras."
+        )
+
+    # Drop rows with NaN values now that we double checked and convert episode_index to int
+    df = df.dropna()
+    df["episode_index"] = df["episode_index"].astype(int)
+
+    # sanity check episode indices go from 0 to n-1
+    assert df["episode_index"].max() == max_episode
+    ep_ids = [ep_idx for ep_idx, _ in df.groupby("episode_index")]
+    expected_ep_ids = list(range(df["episode_index"].max() + 1))
+    if ep_ids != expected_ep_ids:
+        raise ValueError(f"Episodes indices go from {ep_ids} instead of {expected_ep_ids}")
+
+    # Create symlink to raw videos directory (that needs to be absolute not relative)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    videos_dir = out_dir / "videos"
+    videos_dir.symlink_to((raw_dir / "videos").absolute())
+
+    # sanity check the video paths are well formated
+    for key in df:
+        if "observation.images." not in key:
+            continue
+        for ep_idx in ep_ids:
+            video_path = videos_dir / f"{key}_episode_{ep_idx:06d}.mp4"
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
+
+    data_dict = {}
+    for key in df:
+        # is video frame
+        if "observation.images." in key:
+            # we need `[0] because dora only use arrays, so single values are encapsulated into a list.
+            # it is the case for video_frame dictionary = [{"path": ..., "timestamp": ...}]
+            data_dict[key] = [video_frame[0] for video_frame in df[key].values]
+
+            # sanity check the video path is well formated
+            video_path = videos_dir.parent / data_dict[key][0]["path"]
+            if not video_path.exists():
+                raise ValueError(f"Video file not found in {video_path}")
+        # is number
+        elif df[key].iloc[0].ndim == 0 or df[key].iloc[0].shape[0] == 1:
+            data_dict[key] = torch.from_numpy(df[key].values)
+        # is vector
+        elif df[key].iloc[0].shape[0] > 1:
+            data_dict[key] = torch.stack([torch.from_numpy(x.copy()) for x in df[key].values])
+        else:
+            raise ValueError(key)
+
+    # Get the episode index containing for each unique episode index
+    first_ep_index_df = df.groupby("episode_index").agg(start_index=("index", "first")).reset_index()
+    from_ = first_ep_index_df["start_index"].tolist()
+    to_ = from_[1:] + [len(df)]
+    episode_data_index = {
+        "from": from_,
+        "to": to_,
+    }
+
+    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):
+    init_logging()
+
+    if debug:
+        logging.warning("debug=True not implemented. Falling back to debug=False.")
+
+    # sanity check
+    check_format(raw_dir)
+
+    if fps is None:
+        fps = 30
+
+    if not video:
+        raise NotImplementedError()
+
+    data_df, episode_data_index = load_from_raw(raw_dir, out_dir, fps)
+    hf_dataset = to_hf_dataset(data_df, video)
+
+    info = {
+        "fps": fps,
+        "video": video,
+    }
+    return hf_dataset, episode_data_index, info

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)
@@ -69,65 +109,127 @@ 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)):

lerobot/common/policies/act/configuration_act.py

@@ -25,6 +25,14 @@ 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:
+        - "observation.state" is required as an input key.
+        - At least one key starting with "observation.image is required as an input.
+        - If there are multiple keys beginning with "observation.image" they are treated as multiple camera
+          views.
+          Right now we only support all images having the same shape.
+        - "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 +41,15 @@ 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

@@ -200,25 +200,28 @@ class ACT(nn.Module):
         self.config = config
         # BERT style VAE encoder with input [cls, *joint_space_configuration, *action_sequence].
         # The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
+        self.has_state = "observation.state" in config.input_shapes
+        self.latent_dim = config.latent_dim
         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.has_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
             # dimension.
+            num_input_token_encoder = 1 + 1 + config.chunk_size if self.has_state else 1 + config.chunk_size
             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 +241,17 @@ 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
-        )
+        if self.has_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(self.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.has_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.
@@ -285,7 +290,7 @@ 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 +298,16 @@ 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.has_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.has_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.
@@ -317,6 +327,7 @@ class ACT(nn.Module):
         else:
             # When not using the VAE encoder, we set the latent to be all zeros.
             mu = log_sigma_x2 = None
+            # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
             latent_sample = torch.zeros([batch_size, self.latent_dim], dtype=torch.float32).to(
                 batch["observation.state"].device
             )
@@ -326,8 +337,10 @@ class ACT(nn.Module):
         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"]
+            # TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
             cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
             cam_features = self.encoder_img_feat_input_proj(cam_features)  # (B, C, h, w)
             all_cam_features.append(cam_features)
@@ -337,13 +350,15 @@ 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.has_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.has_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"),
             ]
         )
@@ -357,6 +372,7 @@ class ACT(nn.Module):
 
         # Forward pass through the transformer modules.
         encoder_out = self.encoder(encoder_in, pos_embed=pos_embed)
+        # TODO(rcadene, alexander-soare): remove call to `device` ; precompute and use buffer
         decoder_in = torch.zeros(
             (self.config.chunk_size, batch_size, self.config.dim_model),
             dtype=pos_embed.dtype,

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -26,21 +26,29 @@ 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.
+        - At least one key starting with "observation.image is required as an input.
+        - If there are multiple keys beginning with "observation.image" they are treated as multiple camera
+          views.
+          Right now we only support all images having the same shape.
+        - "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 +163,7 @@ 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

@@ -304,7 +304,11 @@ 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 +427,15 @@ 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():

lerobot/configs/default.yaml

@@ -5,10 +5,17 @@ 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.
@@ -29,7 +36,9 @@ training:
   eval_freq: ???
   save_freq: ???
   log_freq: 250
-  save_model: true
+  save_checkpoint: true
+  num_workers: 4
+  batch_size: ???
 
 eval:
   n_episodes: 1
@@ -40,7 +49,7 @@ 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_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/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

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=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 / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    # 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

@@ -0,0 +1,111 @@
+# @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=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 / ${fps} for i in range(${policy.chunk_size})]"
+
+eval:
+  n_episodes: 50
+  batch_size: 50
+
+# See `configuration_act.py` for more details.
+policy:
+  name: act
+
+  # Input / output structure.
+  n_obs_steps: 1
+  chunk_size: 100 # chunk_size
+  n_action_steps: 100
+
+  input_shapes:
+    # 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]
+  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
+  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/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

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
 ```
 

lerobot/scripts/push_dataset_to_hub.py

@@ -84,10 +84,14 @@ def get_from_raw_to_lerobot_format_fn(raw_format):
         from lerobot.common.datasets.push_dataset_to_hub.umi_zarr_format import from_raw_to_lerobot_format
     elif raw_format == "aloha_hdf5":
         from lerobot.common.datasets.push_dataset_to_hub.aloha_hdf5_format import from_raw_to_lerobot_format
+    elif raw_format == "aloha_dora":
+        from lerobot.common.datasets.push_dataset_to_hub.aloha_dora_format import from_raw_to_lerobot_format
     elif raw_format == "xarm_pkl":
         from lerobot.common.datasets.push_dataset_to_hub.xarm_pkl_format import from_raw_to_lerobot_format
     else:
-        raise ValueError(raw_format)
+        raise ValueError(
+            f"The selected {raw_format} can't be found. Did you add it to `lerobot/scripts/push_dataset_to_hub.py::get_from_raw_to_lerobot_format_fn`?"
+        )
 
     return from_raw_to_lerobot_format
 

lerobot/scripts/train.py

@@ -18,13 +18,16 @@ import time
 from contextlib import nullcontext
 from copy import deepcopy
 from pathlib import Path
+from pprint import pformat
 
 import hydra
 import torch
-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.utils import cycle
 from lerobot.common.envs.factory import make_env
 from lerobot.common.logger import Logger, log_output_dir
@@ -34,6 +37,7 @@ 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,
 )
@@ -140,24 +144,6 @@ def update_policy(
     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"]
@@ -237,36 +223,91 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
     init_logging()
 
+    # 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
     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)
 
-    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)})")
@@ -278,7 +319,7 @@ 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}")
             with torch.no_grad(), torch.autocast(device_type=device.type) if cfg.use_amp else nullcontext():
                 eval_info = eval_policy(
@@ -294,12 +335,15 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
                 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)))
                 ),
@@ -309,7 +353,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     # create dataloader for offline training
     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=device.type != "cpu",
@@ -319,7 +363,7 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
     policy.train()
     is_offline = True
-    for step in range(cfg.training.offline_steps):
+    for _ in range(step, cfg.training.offline_steps):
         if step == 0:
             logging.info("Start offline training on a fixed dataset")
         batch = next(dl_iter)
@@ -337,7 +381,6 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
             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)
 
@@ -345,6 +388,18 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
         # so we pass in step + 1.
         evaluate_and_checkpoint_if_needed(step + 1)
 
+        step += 1
+
+    logging.info("End of offline training")
+
+    if cfg.training.online_steps == 0:
+        if cfg.training.eval_freq > 0:
+            eval_env.close()
+        return
+
+    # 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 = {}
@@ -365,8 +420,29 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
         drop_last=False,
     )
 
-    eval_env.close()
-    logging.info("End of training")
+    logging.info("End of online training")
+
+    if cfg.training.eval_freq > 0:
+        eval_env.close()
+    online_training_env.close()
+
+
+@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__":

pyproject.toml

@@ -41,11 +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 = { 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}
@@ -58,9 +59,11 @@ 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]
+dora = ["gym-dora"]
 pusht = ["gym-pusht"]
 xarm = ["gym-xarm"]
 aloha = ["gym-aloha"]

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:ebd21273f6048b66c806f92035352843a9069908b3296863fd55d34cf71cd0ef
+size 51248

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:b9bbf951891077320a5da27e77ddb580a6e833e8d3162b62a2f887a1989585cc
+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:d4070bd1f1cd8c72bc2daf628088e42b8ef113f6df0bfd9e91be052bc90038c3
+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 sha256:42f92239223bb4df32d5c3016bc67450159f1285a7ab046307b645f699ccc34e
+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 sha256:52f85d6262ad1dd0b66578b25829fed96aaaca3c7458cb73ac75111350d17fcf
+size 51248

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 sha256:5ba7c910618f0f3ca69f82f3d70c880d2b2e432456524a2a63dfd5c50efa45f0
+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 sha256:53ad410f43855254438790f54aa7c895a052776acdd922906ae430684f659b53
+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 "observation" in k:
+            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", []),
+        ("dora_aloha_real", "act_real_no_state", []),
     ]
     for env, policy, extra_overrides in env_policies:
         save_policy_to_safetensors("tests/data/save_policy_to_safetensors", env, policy, extra_overrides)

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
@@ -291,6 +293,8 @@ 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_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(
         {