Update lerobot/scripts/push_dataset_to_hub.py

Co-authored-by: Remi <re.cadene@gmail.com>

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

@@ -10,6 +10,7 @@ on:
 
 env:
   PYTHON_VERSION: "3.10"
+#   CI_SLACK_CHANNEL: ${{ secrets.CI_DOCKER_CHANNEL }}
 
 jobs:
   latest-cpu:
@@ -50,6 +51,30 @@ 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
@@ -88,40 +113,27 @@ jobs:
           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: |
-          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
-
-      - name: Login to DockerHub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ secrets.DOCKERHUB_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_PASSWORD }}
-
-      - name: Build and Push GPU dev
-        uses: docker/build-push-action@v5
-        with:
-          context: .
-          file: ./docker/lerobot-gpu-dev/Dockerfile
-          push: true
-          tags: huggingface/lerobot-gpu:dev
-          build-args: PYTHON_VERSION=${{ env.PYTHON_VERSION }}
+      # - 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 }}

.gitignore

@@ -2,16 +2,11 @@
 logs
 tmp
 wandb
-
-# Data
 data
 outputs
-
-# Apple
-.DS_Store
-
-# VS Code
 .vscode
+rl
+.DS_Store
 
 # HPC
 nautilus/*.yaml
@@ -95,7 +90,6 @@ instance/
 docs/_build/
 
 # PyBuilder
-.pybuilder/
 target/
 
 # Jupyter Notebook
@@ -108,6 +102,13 @@ ipython_config.py
 # pyenv
 .python-version
 
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
 # PEP 582; used by e.g. github.com/David-OConnor/pyflow
 __pypackages__/
 
@@ -118,15 +119,6 @@ celerybeat.pid
 # SageMath parsed files
 *.sage.py
 
-# Environments
-.env
-.venv
-env/
-venv/
-ENV/
-env.bak/
-venv.bak/
-
 # Spyder project settings
 .spyderproject
 .spyproject
@@ -144,9 +136,3 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
-
-# pytype static type analyzer
-.pytype/
-
-# Cython debug symbols
-cython_debug/

Makefile

@@ -20,19 +20,15 @@ build-gpu:
 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} test-act-pusht-tutorial
 
 test-act-ete-train:
 	python lerobot/scripts/train.py \
 		policy=act \
-		policy.dim_model=64 \
 		env=aloha \
 		wandb.enable=False \
 		training.offline_steps=2 \
@@ -55,40 +51,9 @@ test-act-ete-eval:
 		env.episode_length=8 \
 		device=cpu \
 
-test-act-ete-train-amp:
-	python lerobot/scripts/train.py \
-		policy=act \
-		policy.dim_model=64 \
-		env=aloha \
-		wandb.enable=False \
-		training.offline_steps=2 \
-		training.online_steps=0 \
-		eval.n_episodes=1 \
-		eval.batch_size=1 \
-		device=cpu \
-		training.save_model=true \
-		training.save_freq=2 \
-		policy.n_action_steps=20 \
-		policy.chunk_size=20 \
-		training.batch_size=2 \
-		hydra.run.dir=tests/outputs/act/ \
-		use_amp=true
-
-test-act-ete-eval-amp:
-	python lerobot/scripts/eval.py \
-		-p tests/outputs/act/checkpoints/000002 \
-		eval.n_episodes=1 \
-		eval.batch_size=1 \
-		env.episode_length=8 \
-		device=cpu \
-		use_amp=true
-
 test-diffusion-ete-train:
 	python lerobot/scripts/train.py \
 		policy=diffusion \
-		policy.down_dims=\[64,128,256\] \
-		policy.diffusion_step_embed_dim=32 \
-		policy.num_inference_steps=10 \
 		env=pusht \
 		wandb.enable=False \
 		training.offline_steps=2 \
@@ -109,7 +74,6 @@ test-diffusion-ete-eval:
 		env.episode_length=8 \
 		device=cpu \
 
-# TODO(alexander-soare): Restore online_steps to 2 when it is reinstated.
 test-tdmpc-ete-train:
 	python lerobot/scripts/train.py \
 		policy=tdmpc \
@@ -118,7 +82,7 @@ test-tdmpc-ete-train:
 		dataset_repo_id=lerobot/xarm_lift_medium \
 		wandb.enable=False \
 		training.offline_steps=2 \
-		training.online_steps=0 \
+		training.online_steps=2 \
 		eval.n_episodes=1 \
 		eval.batch_size=1 \
 		env.episode_length=2 \
@@ -136,6 +100,7 @@ test-tdmpc-ete-eval:
 		env.episode_length=8 \
 		device=cpu \
 
+
 test-default-ete-eval:
 	python lerobot/scripts/eval.py \
 		--config lerobot/configs/default.yaml \
@@ -143,21 +108,3 @@ test-default-ete-eval:
 		eval.batch_size=1 \
 		env.episode_length=8 \
 		device=cpu \
-
-
-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=cpu \
-		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

@@ -77,10 +77,6 @@ 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)
@@ -103,7 +99,6 @@ 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
@@ -163,10 +158,9 @@ See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ### Train your own policy
 
-Check out [example 3](./examples/3_train_policy.py) that illustrates how to train a model using our core library in python, and [example 4](./examples/4_train_policy_with_script.md) that shows how to use our training script from command line.
+Check out [example 3](./examples/3_train_policy.py) that illustrates how to start training a model.
 
 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 \
@@ -190,19 +184,7 @@ 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. 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.
+Note: For efficiency, during training every checkpoint is evaluated on a low number of episodes. After training, you may want to re-evaluate your best checkpoints on more episodes or change the evaluation settings. See `python lerobot/scripts/eval.py --help` for more instructions.
 
 ## Contribute
 

docker/lerobot-gpu-dev/Dockerfile

@@ -1,40 +0,0 @@
-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

@@ -4,15 +4,18 @@ 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${PYTHON_VERSION} /usr/bin/python
 RUN python -m venv /opt/venv
@@ -20,7 +23,8 @@ ENV PATH="/opt/venv/bin:$PATH"
 RUN echo "source /opt/venv/bin/activate" >> /root/.bashrc
 
 # Install LeRobot
-COPY . /lerobot
+RUN git lfs install
+RUN git clone https://github.com/huggingface/lerobot.git
 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

@@ -1,183 +0,0 @@
-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/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/advanced/1_train_act_pusht/train_act_pusht.md

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

gym_dora/README.md

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

gym_dora/example.py

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

gym_dora/gym_dora/__init__.py

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

gym_dora/gym_dora/env.py

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

gym_dora/poetry.lock

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

gym_dora/pyproject.toml

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

lerobot/common/datasets/factory.py

@@ -16,12 +16,15 @@
 import logging
 
 import torch
-from omegaconf import DictConfig, OmegaConf
+from omegaconf import OmegaConf
 
 from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
 
 
-def make_dataset(cfg: DictConfig, split="train") -> LeRobotDataset:
+def make_dataset(
+    cfg,
+    split="train",
+):
     if cfg.env.name not in cfg.dataset_repo_id:
         logging.warning(
             f"There might be a mismatch between your training dataset ({cfg.dataset_repo_id=}) and your "
@@ -40,7 +43,6 @@ def make_dataset(cfg: DictConfig, split="train") -> LeRobotDataset:
         cfg.dataset_repo_id,
         split=split,
         delta_timestamps=delta_timestamps,
-        n_end_keyframes_dropped=eval(cfg.training.get("n_end_keyframes_dropped", "0")),
     )
 
     if cfg.get("override_dataset_stats"):

lerobot/common/datasets/lerobot_dataset.py

@@ -44,26 +44,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         split: str = "train",
         transform: callable = None,
         delta_timestamps: dict[list[float]] | None = None,
-        n_end_keyframes_dropped: int = 0,
     ):
-        """
-        Args:
-            delta_timestamps: A dictionary mapping lists of relative times (Δt) to data keys. When a frame is
-                sampled from the underlying dataset, we treat it as a "keyframe" and load multiple frames
-                according to the list of Δt's. For example {"action": [-0.05, 0, 0.05]} indicates
-                that we want to load the current keyframe's action, as well as one from 50 ms ago, and one
-                50 ms into the future. The action key then contains a (3, action_dim) tensor (whereas without
-                `delta_timestamps` there would just be a (action_dim,) tensor. When the Δt's demand that
-                frames outside of an episode boundary are retrieved, a copy padding strategy is used. See
-                `load_previous_and_future_frames` for more details.
-            n_end_keyframes_dropped: Don't sample the last n items in each episode. This option is handy when
-                used in combination with `delta_timestamps` when, for example, the Δt's demand multiple future
-                frames, but we want to avoid introducing too much copy padding into the data distribution.
-                For example if `delta_timestamps = {"action": [0, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30]}`
-                and we sample the last frame in the episode, we would end up padding with 6 frames worth of
-                copies. Instead, we might want no padding (in which case we need n=6), or we might be okay
-                with up to 2 frames of padding (in which case we need n=4).
-        """
         super().__init__()
         self.repo_id = repo_id
         self.version = version
@@ -84,12 +65,6 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.info = load_info(repo_id, version, root)
         if self.video:
             self.videos_dir = load_videos(repo_id, version, root)
-        # If `n_end_keyframes_dropped == 0`, `self.index` contains exactly the indices of the hf_dataset. If
-        # `n_end_keyframes_dropped > 0`, `self.index` contains a subset of the indices of the hf_dataset where
-        # we drop those indices pertaining to the last n frames of each episode.
-        self.index = []
-        for from_ix, to_ix in zip(*self.episode_data_index.values(), strict=True):
-            self.index.extend(list(range(from_ix, to_ix - n_end_keyframes_dropped)))
 
     @property
     def fps(self) -> int:
@@ -132,11 +107,8 @@ class LeRobotDataset(torch.utils.data.Dataset):
 
     @property
     def num_samples(self) -> int:
-        """Number of possible samples in the dataset.
-
-        This is equivalent to the number of frames in the dataset minus n_end_keyframes_dropped.
-        """
-        return len(self.index)
+        """Number of samples/frames."""
+        return len(self.hf_dataset)
 
     @property
     def num_episodes(self) -> int:
@@ -156,7 +128,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return self.num_samples
 
     def __getitem__(self, idx):
-        item = self.hf_dataset[self.index[idx]]
+        item = self.hf_dataset[idx]
 
         if self.delta_timestamps is not None:
             item = load_previous_and_future_frames(

lerobot/common/datasets/push_dataset_to_hub/aloha_dora_format.py

@@ -0,0 +1,200 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Contains utilities to process raw data format from dora-record
+"""
+
+import logging
+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):
+    # 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
+        modality_df = pd.read_parquet(path)
+        modality_df = modality_df[["timestamp_utc", key]]
+        df = pd.merge_asof(
+            df,
+            modality_df,
+            on="timestamp_utc",
+            direction="backward",
+        )
+
+    # Remove rows with a NaN in any column. It can happened during the first frames of an episode,
+    # because some cameras didnt start recording yet.
+    df = df.dropna(axis=0)
+
+    # Remove rows with episode_index -1 which indicates a failed episode
+    df = df[df["episode_index"] != -1]
+
+    # dora only use arrays, so single values are encapsulated into a list
+    df["episode_index"] = df["episode_index"].map(lambda x: x[0])
+    df["frame_index"] = df.groupby("episode_index").cumcount()
+    df = df.reset_index()
+    df["index"] = df.index
+
+    # 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 episode indices go from 0 to n-1
+    ep_ids = [ep_idx for ep_idx, _ in df.groupby("episode_index")]
+    expected_ep_ids = list(range(df["episode_index"].max() + 1))
+    assert ep_ids == expected_ep_ids, f"Episodes indices go from {ep_ids} instead of {expected_ep_ids}"
+
+    # 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"
+            assert video_path.exists(), 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"]
+            assert video_path.exists(), 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
+    else:
+        raise NotImplementedError()
+
+    if not video:
+        raise NotImplementedError()
+
+    data_df, episode_data_index = load_from_raw(raw_dir, out_dir)
+    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

@@ -28,11 +28,11 @@ def make_env(cfg: DictConfig, n_envs: int | None = None) -> gym.vector.VectorEnv
         raise ValueError("`n_envs must be at least 1")
 
     kwargs = {
-        "obs_type": "pixels_agent_pos",
-        "render_mode": "rgb_array",
+        # "obs_type": "pixels_agent_pos",
+        # "render_mode": "rgb_array",
         "max_episode_steps": cfg.env.episode_length,
-        "visualization_width": 384,
-        "visualization_height": 384,
+        # "visualization_width": 384,
+        # "visualization_height": 384,
     }
 
     package_name = f"gym_{cfg.env.name}"

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -304,11 +304,7 @@ 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:
-            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=}."
-                )
+        if self.config.do_mask_loss_for_padding and "action_is_pad" in batch:
             in_episode_bound = ~batch["action_is_pad"]
             loss = loss * in_episode_bound.unsqueeze(-1)
 

lerobot/configs/default.yaml

@@ -10,9 +10,6 @@ hydra:
     name: default
 
 device: cuda  # cpu
-# `use_amp` determines whether to use Automatic Mixed Precision (AMP) for training and evaluation. With AMP,
-# automatic gradient scaling is used.
-use_amp: false
 # `seed` is used for training (eg: model initialization, dataset shuffling)
 # AND for the evaluation environments.
 seed: ???
@@ -20,7 +17,6 @@ dataset_repo_id: lerobot/pusht
 
 training:
   offline_steps: ???
-  # NOTE: `online_steps` is not implemented yet. It's here as a placeholder.
   online_steps: ???
   online_steps_between_rollouts: ???
   online_sampling_ratio: 0.5

lerobot/configs/env/dora.yaml

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

lerobot/configs/policy/act_real_world.yaml

@@ -1,13 +1,22 @@
 # @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
+seed: 1000
+dataset_repo_id: cadene/aloha_v2_static_dora_test
 
 override_dataset_stats:
-  observation.image:
+  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)
@@ -15,9 +24,9 @@ override_dataset_stats:
 training:
   offline_steps: 80000
   online_steps: 0
-  eval_freq: 10000
-  save_freq: 100000
-  log_freq: 250
+  eval_freq: 99999999999999
+  save_freq: 1000
+  log_freq: 100
   save_model: true
 
   batch_size: 8
@@ -44,19 +53,24 @@ policy:
   n_action_steps: 100
 
   input_shapes:
-    observation.image: [3, 96, 96]
+    # 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.image: mean_std
-    # Use min_max normalization just because it's more standard.
-    observation.state: min_max
+    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:
-    # Use min_max normalization just because it's more standard.
-    action: min_max
+    action: mean_std
 
   # Architecture.
   # Vision backbone.
@@ -70,7 +84,7 @@ policy:
   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
+  # 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

lerobot/configs/policy/diffusion.yaml

@@ -39,21 +39,11 @@ training:
   adam_weight_decay: 1.0e-6
   online_steps_between_rollouts: 1
 
-  # For each training batch we want (consider n_obs_steps=2, horizon=16):
-  # t           | -1,  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14
-  # action      |  a,  a,  a,  a,  a,  a,  a,  a,  a,  a,  a,   a,  a,  a,  a,  a
-  # observation |  o,  o,   ,   ,   ,   ,   ,   ,   ,   ,   ,    ,   ,   ,   ,
-  # Note that at rollout we only use some of the actions (consider n_action_steps=8):
-  # action used |   ,  a,  a,  a,  a,  a,  a,  a,  a,   ,   ,    ,   ,   ,   ,
   delta_timestamps:
     observation.image: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
     observation.state: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1)]"
     action: "[i / ${fps} for i in range(1 - ${policy.n_obs_steps}, 1 - ${policy.n_obs_steps} + ${policy.horizon})]"
 
-  # The original implementation doesn't sample keyframes for the last 7 steps. This is because, as described
-  # above, the last 7 actions from the diffusion model are not used.
-  n_end_keyframes_dropped: ${policy.horizon} - ${policy.n_action_steps} - ${policy.n_obs_steps} + 1
-
 eval:
   n_episodes: 50
   batch_size: 50

lerobot/configs/policy/tdmpc.yaml

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

lerobot/scripts/eval.py

@@ -46,7 +46,6 @@ import json
 import logging
 import threading
 import time
-from contextlib import nullcontext
 from copy import deepcopy
 from datetime import datetime as dt
 from pathlib import Path
@@ -521,7 +520,7 @@ def eval(
         raise NotImplementedError()
 
     # Check device is available
-    device = get_safe_torch_device(hydra_cfg.device, log=True)
+    get_safe_torch_device(hydra_cfg.device, log=True)
 
     torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
@@ -540,17 +539,16 @@ def eval(
         policy = make_policy(hydra_cfg=hydra_cfg, dataset_stats=make_dataset(hydra_cfg).stats)
     policy.eval()
 
-    with torch.no_grad(), torch.autocast(device_type=device.type) if hydra_cfg.use_amp else nullcontext():
-        info = eval_policy(
-            env,
-            policy,
-            hydra_cfg.eval.n_episodes,
-            max_episodes_rendered=10,
-            video_dir=Path(out_dir) / "eval",
-            start_seed=hydra_cfg.seed,
-            enable_progbar=True,
-            enable_inner_progbar=True,
-        )
+    info = eval_policy(
+        env,
+        policy,
+        hydra_cfg.eval.n_episodes,
+        max_episodes_rendered=10,
+        video_dir=Path(out_dir) / "eval",
+        start_seed=hydra_cfg.seed,
+        enable_progbar=True,
+        enable_inner_progbar=True,
+    )
     print(info["aggregated"])
 
     # Save info

lerobot/scripts/push_dataset_to_hub.py

@@ -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
 
@@ -140,7 +144,8 @@ def push_videos_to_hub(repo_id, videos_dir, revision):
 
 
 def push_dataset_to_hub(
-    data_dir: Path,
+    input_data_dir: Path,
+    output_data_dir: Path,
     dataset_id: str,
     raw_format: str | None,
     community_id: str,
@@ -157,34 +162,33 @@ def push_dataset_to_hub(
 ):
     repo_id = f"{community_id}/{dataset_id}"
 
-    raw_dir = data_dir / f"{dataset_id}_raw"
-
-    out_dir = data_dir / repo_id
-    meta_data_dir = out_dir / "meta_data"
-    videos_dir = out_dir / "videos"
+    meta_data_dir = output_data_dir / "meta_data"
+    videos_dir = output_data_dir / "videos"
 
     tests_out_dir = tests_data_dir / repo_id
     tests_meta_data_dir = tests_out_dir / "meta_data"
     tests_videos_dir = tests_out_dir / "videos"
 
-    if out_dir.exists():
-        shutil.rmtree(out_dir)
+    if output_data_dir.exists():
+        shutil.rmtree(output_data_dir)
 
     if tests_out_dir.exists() and save_tests_to_disk:
         shutil.rmtree(tests_out_dir)
 
-    if not raw_dir.exists():
-        download_raw(raw_dir, dataset_id)
+    if not input_data_dir.exists():
+        download_raw(input_data_dir, dataset_id)
 
     if raw_format is None:
         # TODO(rcadene, adilzouitine): implement auto_find_raw_format
         raise NotImplementedError()
-        # raw_format = auto_find_raw_format(raw_dir)
+        # raw_format = auto_find_raw_format(input_data_dir)
 
     from_raw_to_lerobot_format = get_from_raw_to_lerobot_format_fn(raw_format)
 
     # convert dataset from original raw format to LeRobot format
-    hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(raw_dir, out_dir, fps, video, debug)
+    hf_dataset, episode_data_index, info = from_raw_to_lerobot_format(
+        input_data_dir, output_data_dir, fps, video, debug
+    )
 
     lerobot_dataset = LeRobotDataset.from_preloaded(
         repo_id=repo_id,
@@ -198,7 +202,7 @@ def push_dataset_to_hub(
 
     if save_to_disk:
         hf_dataset = hf_dataset.with_format(None)  # to remove transforms that cant be saved
-        hf_dataset.save_to_disk(str(out_dir / "train"))
+        hf_dataset.save_to_disk(str(output_data_dir / "train"))
 
     if not dry_run or save_to_disk:
         # mandatory for upload
@@ -232,19 +236,25 @@ def push_dataset_to_hub(
             fname = f"{key}_episode_{episode_index:06d}.mp4"
             shutil.copy(videos_dir / fname, tests_videos_dir / fname)
 
-    if not save_to_disk and out_dir.exists():
+    if not save_to_disk and output_data_dir.exists():
         # remove possible temporary files remaining in the output directory
-        shutil.rmtree(out_dir)
+        shutil.rmtree(output_data_dir)
 
 
 def main():
     parser = argparse.ArgumentParser()
 
     parser.add_argument(
-        "--data-dir",
+        "--input-data-dir",
         type=Path,
         required=True,
-        help="Root directory containing datasets (e.g. `data` or `tmp/data` or `/tmp/lerobot/data`).",
+        help="Directory containing input raw datasets (e.g. `data/aloha_mobile_chair_raw` or `data/pusht_raw`).",
+    )
+    parser.add_argument(
+        "--output-data-dir",
+        type=Path,
+        required=True,
+        help="Root directory containing output dataset (e.g. `data/lerobot/aloha_mobile_chair` or `data/lerobot/pusht`).",
     )
     parser.add_argument(
         "--dataset-id",

lerobot/scripts/train.py

@@ -15,14 +15,15 @@
 # limitations under the License.
 import logging
 import time
-from contextlib import nullcontext
 from copy import deepcopy
 from pathlib import Path
 
+import datasets
 import hydra
 import torch
+from datasets import concatenate_datasets
+from datasets.utils import disable_progress_bars, enable_progress_bars
 from omegaconf import DictConfig
-from torch.cuda.amp import GradScaler
 
 from lerobot.common.datasets.factory import make_dataset
 from lerobot.common.datasets.utils import cycle
@@ -30,7 +31,6 @@ from lerobot.common.envs.factory import make_env
 from lerobot.common.logger import Logger, log_output_dir
 from lerobot.common.policies.factory import make_policy
 from lerobot.common.policies.policy_protocol import PolicyWithUpdate
-from lerobot.common.policies.utils import get_device_from_parameters
 from lerobot.common.utils.utils import (
     format_big_number,
     get_safe_torch_device,
@@ -69,6 +69,7 @@ def make_optimizer_and_scheduler(cfg, policy):
             cfg.training.adam_eps,
             cfg.training.adam_weight_decay,
         )
+        assert cfg.training.online_steps == 0, "Diffusion Policy does not handle online training."
         from diffusers.optimization import get_scheduler
 
         lr_scheduler = get_scheduler(
@@ -86,40 +87,21 @@ def make_optimizer_and_scheduler(cfg, policy):
     return optimizer, lr_scheduler
 
 
-def update_policy(
-    policy,
-    batch,
-    optimizer,
-    grad_clip_norm,
-    grad_scaler: GradScaler,
-    lr_scheduler=None,
-    use_amp: bool = False,
-):
+def update_policy(policy, batch, optimizer, grad_clip_norm, lr_scheduler=None):
     """Returns a dictionary of items for logging."""
-    start_time = time.perf_counter()
-    device = get_device_from_parameters(policy)
+    start_time = time.time()
     policy.train()
-    with torch.autocast(device_type=device.type) if use_amp else nullcontext():
-        output_dict = policy.forward(batch)
-        # TODO(rcadene): policy.unnormalize_outputs(out_dict)
-        loss = output_dict["loss"]
-    grad_scaler.scale(loss).backward()
-
-    # Unscale the graident of the optimzer's assigned params in-place **prior to gradient clipping**.
-    grad_scaler.unscale_(optimizer)
-
+    output_dict = policy.forward(batch)
+    # TODO(rcadene): policy.unnormalize_outputs(out_dict)
+    loss = output_dict["loss"]
+    loss.backward()
     grad_norm = torch.nn.utils.clip_grad_norm_(
         policy.parameters(),
         grad_clip_norm,
         error_if_nonfinite=False,
     )
 
-    # Optimizer's gradients are already unscaled, so scaler.step does not unscale them,
-    # although it still skips optimizer.step() if the gradients contain infs or NaNs.
-    grad_scaler.step(optimizer)
-    # Updates the scale for next iteration.
-    grad_scaler.update()
-
+    optimizer.step()
     optimizer.zero_grad()
 
     if lr_scheduler is not None:
@@ -133,7 +115,7 @@ def update_policy(
         "loss": loss.item(),
         "grad_norm": float(grad_norm),
         "lr": optimizer.param_groups[0]["lr"],
-        "update_s": time.perf_counter() - start_time,
+        "update_s": time.time() - start_time,
         **{k: v for k, v in output_dict.items() if k != "loss"},
     }
 
@@ -229,6 +211,103 @@ def log_eval_info(logger, info, step, cfg, dataset, is_offline):
     logger.log_dict(info, step, mode="eval")
 
 
+def calculate_online_sample_weight(n_off: int, n_on: int, pc_on: float):
+    """
+    Calculate the sampling weight to be assigned to samples so that a specified percentage of the batch comes from online dataset (on average).
+
+    Parameters:
+    - n_off (int): Number of offline samples, each with a sampling weight of 1.
+    - n_on (int): Number of online samples.
+    - pc_on (float): Desired percentage of online samples in decimal form (e.g., 50% as 0.5).
+
+    The total weight of offline samples is n_off * 1.0.
+    The total weight of offline samples is n_on * w.
+    The total combined weight of all samples is n_off + n_on * w.
+    The fraction of the weight that is online is n_on * w / (n_off + n_on * w).
+    We want this fraction to equal pc_on, so we set up the equation n_on * w / (n_off + n_on * w) = pc_on.
+    The solution is w = - (n_off * pc_on) / (n_on * (pc_on - 1))
+    """
+    assert 0.0 <= pc_on <= 1.0
+    return -(n_off * pc_on) / (n_on * (pc_on - 1))
+
+
+def add_episodes_inplace(
+    online_dataset: torch.utils.data.Dataset,
+    concat_dataset: torch.utils.data.ConcatDataset,
+    sampler: torch.utils.data.WeightedRandomSampler,
+    hf_dataset: datasets.Dataset,
+    episode_data_index: dict[str, torch.Tensor],
+    pc_online_samples: float,
+):
+    """
+    Modifies the online_dataset, concat_dataset, and sampler in place by integrating
+    new episodes from hf_dataset into the online_dataset, updating the concatenated
+    dataset's structure and adjusting the sampling strategy based on the specified
+    percentage of online samples.
+
+    Parameters:
+    - online_dataset (torch.utils.data.Dataset): The existing online dataset to be updated.
+    - concat_dataset (torch.utils.data.ConcatDataset): The concatenated dataset that combines
+      offline and online datasets, used for sampling purposes.
+    - sampler (torch.utils.data.WeightedRandomSampler): A sampler that will be updated to
+      reflect changes in the dataset sizes and specified sampling weights.
+    - hf_dataset (datasets.Dataset): A Hugging Face dataset containing the new episodes to be added.
+    - episode_data_index (dict): A dictionary containing two keys ("from" and "to") associated to dataset indices.
+      They indicate the start index and end index of each episode in the dataset.
+    - pc_online_samples (float): The target percentage of samples that should come from
+      the online dataset during sampling operations.
+
+    Raises:
+    - AssertionError: If the first episode_id or index in hf_dataset is not 0
+    """
+    first_episode_idx = hf_dataset.select_columns("episode_index")[0]["episode_index"].item()
+    last_episode_idx = hf_dataset.select_columns("episode_index")[-1]["episode_index"].item()
+    first_index = hf_dataset.select_columns("index")[0]["index"].item()
+    last_index = hf_dataset.select_columns("index")[-1]["index"].item()
+    # sanity check
+    assert first_episode_idx == 0, f"{first_episode_idx=} is not 0"
+    assert first_index == 0, f"{first_index=} is not 0"
+    assert first_index == episode_data_index["from"][first_episode_idx].item()
+    assert last_index == episode_data_index["to"][last_episode_idx].item() - 1
+
+    if len(online_dataset) == 0:
+        # initialize online dataset
+        online_dataset.hf_dataset = hf_dataset
+        online_dataset.episode_data_index = episode_data_index
+    else:
+        # get the starting indices of the new episodes and frames to be added
+        start_episode_idx = last_episode_idx + 1
+        start_index = last_index + 1
+
+        def shift_indices(episode_index, index):
+            # note: we dont shift "frame_index" since it represents the index of the frame in the episode it belongs to
+            example = {"episode_index": episode_index + start_episode_idx, "index": index + start_index}
+            return example
+
+        disable_progress_bars()  # map has a tqdm progress bar
+        hf_dataset = hf_dataset.map(shift_indices, input_columns=["episode_index", "index"])
+        enable_progress_bars()
+
+        episode_data_index["from"] += start_index
+        episode_data_index["to"] += start_index
+
+        # extend online dataset
+        online_dataset.hf_dataset = concatenate_datasets([online_dataset.hf_dataset, hf_dataset])
+
+    # update the concatenated dataset length used during sampling
+    concat_dataset.cumulative_sizes = concat_dataset.cumsum(concat_dataset.datasets)
+
+    # update the sampling weights for each frame so that online frames get sampled a certain percentage of times
+    len_online = len(online_dataset)
+    len_offline = len(concat_dataset) - len_online
+    weight_offline = 1.0
+    weight_online = calculate_online_sample_weight(len_offline, len_online, pc_online_samples)
+    sampler.weights = torch.tensor([weight_offline] * len_offline + [weight_online] * len(online_dataset))
+
+    # update the total number of samples used during sampling
+    sampler.num_samples = len(concat_dataset)
+
+
 def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = None):
     if out_dir is None:
         raise NotImplementedError()
@@ -237,11 +316,11 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
 
     init_logging()
 
-    if cfg.training.online_steps > 0:
-        raise NotImplementedError("Online training is not implemented yet.")
+    if cfg.training.online_steps > 0 and cfg.eval.batch_size > 1:
+        logging.warning("eval.batch_size > 1 not supported for online training steps")
 
     # Check device is available
-    device = get_safe_torch_device(cfg.device, log=True)
+    get_safe_torch_device(cfg.device, log=True)
 
     torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
@@ -259,7 +338,6 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     # 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)
 
     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())
@@ -280,15 +358,14 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
     def evaluate_and_checkpoint_if_needed(step):
         if 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(
-                    eval_env,
-                    policy,
-                    cfg.eval.n_episodes,
-                    video_dir=Path(out_dir) / "eval",
-                    max_episodes_rendered=4,
-                    start_seed=cfg.seed,
-                )
+            eval_info = eval_policy(
+                eval_env,
+                policy,
+                cfg.eval.n_episodes,
+                video_dir=Path(out_dir) / "eval",
+                max_episodes_rendered=4,
+                start_seed=cfg.seed,
+            )
             log_eval_info(logger, eval_info["aggregated"], step, cfg, offline_dataset, is_offline)
             if cfg.wandb.enable:
                 logger.log_video(eval_info["video_paths"][0], step, mode="eval")
@@ -312,30 +389,23 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
         num_workers=4,
         batch_size=cfg.training.batch_size,
         shuffle=True,
-        pin_memory=device.type != "cpu",
+        pin_memory=cfg.device != "cpu",
         drop_last=False,
     )
     dl_iter = cycle(dataloader)
 
     policy.train()
+    step = 0  # number of policy update (forward + backward + optim)
     is_offline = True
-    for step in range(cfg.training.offline_steps):
-        if step == 0:
+    for offline_step in range(cfg.training.offline_steps):
+        if offline_step == 0:
             logging.info("Start offline training on a fixed dataset")
         batch = next(dl_iter)
 
         for key in batch:
-            batch[key] = batch[key].to(device, non_blocking=True)
+            batch[key] = batch[key].to(cfg.device, non_blocking=True)
 
-        train_info = update_policy(
-            policy,
-            batch,
-            optimizer,
-            cfg.training.grad_clip_norm,
-            grad_scaler=grad_scaler,
-            lr_scheduler=lr_scheduler,
-            use_amp=cfg.use_amp,
-        )
+        train_info = update_policy(policy, batch, optimizer, cfg.training.grad_clip_norm, lr_scheduler)
 
         # TODO(rcadene): is it ok if step_t=0 = 0 and not 1 as previously done?
         if step % cfg.training.log_freq == 0:
@@ -345,6 +415,11 @@ 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
+
+    # 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 = {}
@@ -361,11 +436,58 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
         num_workers=4,
         batch_size=cfg.training.batch_size,
         sampler=sampler,
-        pin_memory=device.type != "cpu",
+        pin_memory=cfg.device != "cpu",
         drop_last=False,
     )
+    dl_iter = cycle(dataloader)
+
+    online_step = 0
+    is_offline = False
+    for env_step in range(cfg.training.online_steps):
+        if env_step == 0:
+            logging.info("Start online training by interacting with environment")
+
+        policy.eval()
+        with torch.no_grad():
+            eval_info = eval_policy(
+                online_training_env,
+                policy,
+                n_episodes=1,
+                return_episode_data=True,
+                start_seed=cfg.training.online_env_seed,
+                enable_progbar=True,
+            )
+
+        add_episodes_inplace(
+            online_dataset,
+            concat_dataset,
+            sampler,
+            hf_dataset=eval_info["episodes"]["hf_dataset"],
+            episode_data_index=eval_info["episodes"]["episode_data_index"],
+            pc_online_samples=cfg.training.online_sampling_ratio,
+        )
+
+        policy.train()
+        for _ in range(cfg.training.online_steps_between_rollouts):
+            batch = next(dl_iter)
+
+            for key in batch:
+                batch[key] = batch[key].to(cfg.device, non_blocking=True)
+
+            train_info = update_policy(policy, batch, optimizer, cfg.training.grad_clip_norm, lr_scheduler)
+
+            if step % cfg.training.log_freq == 0:
+                log_train_info(logger, train_info, step, cfg, online_dataset, is_offline)
+
+            # Note: evaluate_and_checkpoint_if_needed happens **after** the `step`th training update has completed,
+            # so we pass in step + 1.
+            evaluate_and_checkpoint_if_needed(step + 1)
+
+            step += 1
+            online_step += 1
 
     eval_env.close()
+    online_training_env.close()
     logging.info("End of training")
 
 

poetry.lock

@@ -767,6 +767,26 @@ files = [
 [package.dependencies]
 six = ">=1.4.0"
 
+[[package]]
+name = "dora-rs"
+version = "0.3.4"
+description = "`dora` goal is to be a low latency, composable, and distributed data flow."
+optional = true
+python-versions = "*"
+files = [
+    {file = "dora_rs-0.3.4-cp37-abi3-macosx_10_12_x86_64.whl", hash = "sha256:d1b738eea5a4966d731c26c6b6a0a50a491a24f7e9e335475f983cfc6f0da19e"},
+    {file = "dora_rs-0.3.4-cp37-abi3-macosx_11_0_arm64.whl", hash = "sha256:80b724871618c78a4e5863938fa66724176cc40352771087aebe1e62a8141157"},
+    {file = "dora_rs-0.3.4-cp37-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:3a3919e157b47dc1dbc74c040a73087a4485f0d1bee99b6adcdbc36559400fe2"},
+    {file = "dora_rs-0.3.4-cp37-abi3-manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:f7c95f6e5858fd651d6cd220e4f052e99db2944b9c37fb0b5402d60ac4b41a63"},
+    {file = "dora_rs-0.3.4-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:37d915fbbca282446235c98a9ca08389aa3ef3155d4e88c6c136326e9a830042"},
+    {file = "dora_rs-0.3.4-cp37-abi3-win32.whl", hash = "sha256:c9f7f22f65c884ec9bee0245ce98d0c7fad25dec0f982e566f844b5e8e58818f"},
+    {file = "dora_rs-0.3.4-cp37-abi3-win_amd64.whl", hash = "sha256:0a6a37f96a9f6e13b58b02a6ea75af192af5fbe4f456f6a67b1f239c3cee3276"},
+    {file = "dora_rs-0.3.4.tar.gz", hash = "sha256:05c5d0db0d23d7c4669995ae34db11cd636dbf91f5705d832669bd04e7452903"},
+]
+
+[package.dependencies]
+pyarrow = "*"
+
 [[package]]
 name = "einops"
 version = "0.8.0"
@@ -1048,6 +1068,24 @@ mujoco = ">=2.3.7,<3.0.0"
 dev = ["debugpy (>=1.8.1)", "pre-commit (>=3.7.0)"]
 test = ["pytest (>=8.1.0)", "pytest-cov (>=5.0.0)"]
 
+[[package]]
+name = "gym-dora"
+version = "0.1.0"
+description = ""
+optional = true
+python-versions = "^3.10"
+files = []
+develop = true
+
+[package.dependencies]
+dora-rs = ">=0.3.4"
+gymnasium = ">=0.29.1"
+pyarrow = ">=12.0.0"
+
+[package.source]
+type = "directory"
+url = "gym_dora"
+
 [[package]]
 name = "gym-pusht"
 version = "0.1.4"
@@ -1655,9 +1693,13 @@ files = [
     {file = "lxml-5.2.2-cp36-cp36m-win_amd64.whl", hash = "sha256:edcfa83e03370032a489430215c1e7783128808fd3e2e0a3225deee278585196"},
     {file = "lxml-5.2.2-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:28bf95177400066596cdbcfc933312493799382879da504633d16cf60bba735b"},
     {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_12_i686.manylinux2010_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:3a745cc98d504d5bd2c19b10c79c61c7c3df9222629f1b6210c0368177589fb8"},
+    {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1b590b39ef90c6b22ec0be925b211298e810b4856909c8ca60d27ffbca6c12e6"},
     {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b336b0416828022bfd5a2e3083e7f5ba54b96242159f83c7e3eebaec752f1716"},
+    {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_28_aarch64.whl", hash = "sha256:c2faf60c583af0d135e853c86ac2735ce178f0e338a3c7f9ae8f622fd2eb788c"},
     {file = "lxml-5.2.2-cp37-cp37m-manylinux_2_28_x86_64.whl", hash = "sha256:4bc6cb140a7a0ad1f7bc37e018d0ed690b7b6520ade518285dc3171f7a117905"},
+    {file = "lxml-5.2.2-cp37-cp37m-musllinux_1_1_aarch64.whl", hash = "sha256:7ff762670cada8e05b32bf1e4dc50b140790909caa8303cfddc4d702b71ea184"},
     {file = "lxml-5.2.2-cp37-cp37m-musllinux_1_1_x86_64.whl", hash = "sha256:57f0a0bbc9868e10ebe874e9f129d2917750adf008fe7b9c1598c0fbbfdde6a6"},
+    {file = "lxml-5.2.2-cp37-cp37m-musllinux_1_2_aarch64.whl", hash = "sha256:a6d2092797b388342c1bc932077ad232f914351932353e2e8706851c870bca1f"},
     {file = "lxml-5.2.2-cp37-cp37m-musllinux_1_2_x86_64.whl", hash = "sha256:60499fe961b21264e17a471ec296dcbf4365fbea611bf9e303ab69db7159ce61"},
     {file = "lxml-5.2.2-cp37-cp37m-win32.whl", hash = "sha256:d9b342c76003c6b9336a80efcc766748a333573abf9350f4094ee46b006ec18f"},
     {file = "lxml-5.2.2-cp37-cp37m-win_amd64.whl", hash = "sha256:b16db2770517b8799c79aa80f4053cd6f8b716f21f8aca962725a9565ce3ee40"},
@@ -2941,13 +2983,13 @@ files = [
 
 [[package]]
 name = "pytest"
-version = "8.2.0"
+version = "8.2.1"
 description = "pytest: simple powerful testing with Python"
 optional = true
 python-versions = ">=3.8"
 files = [
-    {file = "pytest-8.2.0-py3-none-any.whl", hash = "sha256:1733f0620f6cda4095bbf0d9ff8022486e91892245bb9e7d5542c018f612f233"},
-    {file = "pytest-8.2.0.tar.gz", hash = "sha256:d507d4482197eac0ba2bae2e9babf0672eb333017bcedaa5fb1a3d42c1174b3f"},
+    {file = "pytest-8.2.1-py3-none-any.whl", hash = "sha256:faccc5d332b8c3719f40283d0d44aa5cf101cec36f88cde9ed8f2bc0538612b1"},
+    {file = "pytest-8.2.1.tar.gz", hash = "sha256:5046e5b46d8e4cac199c373041f26be56fdb81eb4e67dc11d4e10811fc3408fd"},
 ]
 
 [package.dependencies]
@@ -3029,6 +3071,7 @@ files = [
     {file = "PyYAML-6.0.1-cp311-cp311-win_amd64.whl", hash = "sha256:bf07ee2fef7014951eeb99f56f39c9bb4af143d8aa3c21b1677805985307da34"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:855fb52b0dc35af121542a76b9a84f8d1cd886ea97c84703eaa6d88e37a2ad28"},
     {file = "PyYAML-6.0.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:40df9b996c2b73138957fe23a16a4f0ba614f4c0efce1e9406a184b6d07fa3a9"},
+    {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a08c6f0fe150303c1c6b71ebcd7213c2858041a7e01975da3a99aed1e7a378ef"},
     {file = "PyYAML-6.0.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6c22bec3fbe2524cde73d7ada88f6566758a8f7227bfbf93a408a9d86bcc12a0"},
     {file = "PyYAML-6.0.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8d4e9c88387b0f5c7d5f281e55304de64cf7f9c0021a3525bd3b1c542da3b0e4"},
     {file = "PyYAML-6.0.1-cp312-cp312-win32.whl", hash = "sha256:d483d2cdf104e7c9fa60c544d92981f12ad66a457afae824d146093b8c294c54"},
@@ -3153,13 +3196,13 @@ files = [
 
 [[package]]
 name = "requests"
-version = "2.31.0"
+version = "2.32.1"
 description = "Python HTTP for Humans."
 optional = false
-python-versions = ">=3.7"
+python-versions = ">=3.8"
 files = [
-    {file = "requests-2.31.0-py3-none-any.whl", hash = "sha256:58cd2187c01e70e6e26505bca751777aa9f2ee0b7f4300988b709f44e013003f"},
-    {file = "requests-2.31.0.tar.gz", hash = "sha256:942c5a758f98d790eaed1a29cb6eefc7ffb0d1cf7af05c3d2791656dbd6ad1e1"},
+    {file = "requests-2.32.1-py3-none-any.whl", hash = "sha256:21ac9465cdf8c1650fe1ecde8a71669a93d4e6f147550483a2967d08396a56a5"},
+    {file = "requests-2.32.1.tar.gz", hash = "sha256:eb97e87e64c79e64e5b8ac75cee9dd1f97f49e289b083ee6be96268930725685"},
 ]
 
 [package.dependencies]
@@ -3406,13 +3449,13 @@ test = ["array-api-strict", "asv", "gmpy2", "hypothesis (>=6.30)", "mpmath", "po
 
 [[package]]
 name = "sentry-sdk"
-version = "2.2.0"
+version = "2.2.1"
 description = "Python client for Sentry (https://sentry.io)"
 optional = false
 python-versions = ">=3.6"
 files = [
-    {file = "sentry_sdk-2.2.0-py2.py3-none-any.whl", hash = "sha256:674f58da37835ea7447fe0e34c57b4a4277fad558b0a7cb4a6c83bcb263086be"},
-    {file = "sentry_sdk-2.2.0.tar.gz", hash = "sha256:70eca103cf4c6302365a9d7cf522e7ed7720828910eb23d43ada8e50d1ecda9d"},
+    {file = "sentry_sdk-2.2.1-py2.py3-none-any.whl", hash = "sha256:7d617a1b30e80c41f3b542347651fcf90bb0a36f3a398be58b4f06b79c8d85bc"},
+    {file = "sentry_sdk-2.2.1.tar.gz", hash = "sha256:8aa2ec825724d8d9d645cab68e6034928b1a6a148503af3e361db3fa6401183f"},
 ]
 
 [package.dependencies]
@@ -3434,7 +3477,7 @@ django = ["django (>=1.8)"]
 falcon = ["falcon (>=1.4)"]
 fastapi = ["fastapi (>=0.79.0)"]
 flask = ["blinker (>=1.1)", "flask (>=0.11)", "markupsafe"]
-grpcio = ["grpcio (>=1.21.1)"]
+grpcio = ["grpcio (>=1.21.1)", "protobuf (>=3.8.0)"]
 httpx = ["httpx (>=0.16.0)"]
 huey = ["huey (>=2)"]
 huggingface-hub = ["huggingface-hub (>=0.22)"]
@@ -3556,19 +3599,18 @@ test = ["pytest"]
 
 [[package]]
 name = "setuptools"
-version = "69.5.1"
+version = "70.0.0"
 description = "Easily download, build, install, upgrade, and uninstall Python packages"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "setuptools-69.5.1-py3-none-any.whl", hash = "sha256:c636ac361bc47580504644275c9ad802c50415c7522212252c033bd15f301f32"},
-    {file = "setuptools-69.5.1.tar.gz", hash = "sha256:6c1fccdac05a97e598fb0ae3bbed5904ccb317337a51139dcd51453611bbb987"},
+    {file = "setuptools-70.0.0-py3-none-any.whl", hash = "sha256:54faa7f2e8d2d11bcd2c07bed282eef1046b5c080d1c32add737d7b5817b1ad4"},
+    {file = "setuptools-70.0.0.tar.gz", hash = "sha256:f211a66637b8fa059bb28183da127d4e86396c991a942b028c6650d4319c3fd0"},
 ]
 
 [package.extras]
-docs = ["furo", "jaraco.packaging (>=9.3)", "jaraco.tidelift (>=1.4)", "pygments-github-lexers (==0.0.5)", "rst.linker (>=1.9)", "sphinx (>=3.5)", "sphinx-favicon", "sphinx-inline-tabs", "sphinx-lint", "sphinx-notfound-page (>=1,<2)", "sphinx-reredirects", "sphinxcontrib-towncrier"]
-testing = ["build[virtualenv]", "filelock (>=3.4.0)", "importlib-metadata", "ini2toml[lite] (>=0.9)", "jaraco.develop (>=7.21)", "jaraco.envs (>=2.2)", "jaraco.path (>=3.2.0)", "mypy (==1.9)", "packaging (>=23.2)", "pip (>=19.1)", "pytest (>=6,!=8.1.1)", "pytest-checkdocs (>=2.4)", "pytest-cov", "pytest-enabler (>=2.2)", "pytest-home (>=0.5)", "pytest-mypy", "pytest-perf", "pytest-ruff (>=0.2.1)", "pytest-timeout", "pytest-xdist (>=3)", "tomli", "tomli-w (>=1.0.0)", "virtualenv (>=13.0.0)", "wheel"]
-testing-integration = ["build[virtualenv] (>=1.0.3)", "filelock (>=3.4.0)", "jaraco.envs (>=2.2)", "jaraco.path (>=3.2.0)", "packaging (>=23.2)", "pytest", "pytest-enabler", "pytest-xdist", "tomli", "virtualenv (>=13.0.0)", "wheel"]
+docs = ["furo", "jaraco.packaging (>=9.3)", "jaraco.tidelift (>=1.4)", "pygments-github-lexers (==0.0.5)", "pyproject-hooks (!=1.1)", "rst.linker (>=1.9)", "sphinx (>=3.5)", "sphinx-favicon", "sphinx-inline-tabs", "sphinx-lint", "sphinx-notfound-page (>=1,<2)", "sphinx-reredirects", "sphinxcontrib-towncrier"]
+testing = ["build[virtualenv] (>=1.0.3)", "filelock (>=3.4.0)", "importlib-metadata", "ini2toml[lite] (>=0.14)", "jaraco.develop (>=7.21)", "jaraco.envs (>=2.2)", "jaraco.path (>=3.2.0)", "mypy (==1.9)", "packaging (>=23.2)", "pip (>=19.1)", "pyproject-hooks (!=1.1)", "pytest (>=6,!=8.1.1)", "pytest-checkdocs (>=2.4)", "pytest-cov", "pytest-enabler (>=2.2)", "pytest-home (>=0.5)", "pytest-mypy", "pytest-perf", "pytest-ruff (>=0.2.1)", "pytest-subprocess", "pytest-timeout", "pytest-xdist (>=3)", "tomli", "tomli-w (>=1.0.0)", "virtualenv (>=13.0.0)", "wheel"]
 
 [[package]]
 name = "shapely"
@@ -4223,6 +4265,7 @@ testing = ["big-O", "jaraco.functools", "jaraco.itertools", "jaraco.test", "more
 [extras]
 aloha = ["gym-aloha"]
 dev = ["debugpy", "pre-commit"]
+dora = ["gym-dora"]
 pusht = ["gym-pusht"]
 test = ["pytest", "pytest-cov"]
 umi = ["imagecodecs"]
@@ -4231,4 +4274,4 @@ xarm = ["gym-xarm"]
 [metadata]
 lock-version = "2.0"
 python-versions = ">=3.10,<3.13"
-content-hash = "c3044329cfad91ffd91b411e85f16d8dfdcdfd7b9186d38fff5e18f4ee647e7b"
+content-hash = "ea4e8207316a8ec8a4b95d6a89cf488c8733a8e7ab43e5f669c889ee87f3bef3"

pyproject.toml

@@ -41,11 +41,12 @@ numba = ">=0.59.0"
 torch = "^2.2.1"
 opencv-python = ">=4.9.0"
 diffusers = "^0.27.2"
-torchvision = ">=0.17.1"
+torchvision = ">=0.18.0"
 h5py = ">=3.10.0"
 huggingface-hub = {extras = ["hf-transfer"], version = "^0.23.0"}
 gymnasium = ">=0.29.1"
 cmake = ">=3.29.0.1"
+gym-dora = { path = "gym_dora", optional = true, develop = true}
 gym-pusht = { version = ">=0.1.3", optional = true}
 gym-xarm = { version = ">=0.1.1", optional = true}
 gym-aloha = { version = ">=0.1.1", optional = true}
@@ -61,6 +62,7 @@ rerun-sdk = ">=0.15.1"
 
 
 [tool.poetry.extras]
+dora = ["gym-dora"]
 pusht = ["gym-pusht"]
 xarm = ["gym-xarm"]
 aloha = ["gym-aloha"]

tests/test_datasets.py

@@ -115,7 +115,6 @@ def test_compute_stats_on_xarm():
 
     # reduce size of dataset sample on which stats compute is tested to 10 frames
     dataset.hf_dataset = dataset.hf_dataset.select(range(10))
-    dataset.index = [i for i in dataset.index if i < 10]
 
     # Note: we set the batch size to be smaller than the whole dataset to make sure we are testing batched
     # computation of the statistics. While doing this, we also make sure it works when we don't divide the

tests/test_examples.py

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