LeRoboCasa is a forked version of RoboCasa, maintained with a devcontainer-first workflow. The container provisions the right system dependencies, runs uv sync, and configures macros automatically.

This repository now targets Python 3.12. If you are refreshing an existing local environment, run:

uv python install 3.12
uv venv --python 3.12 --clear
uv lock
uv sync

You have to install docker to be able to use the container setup.

1. Open this repository in VS Code.
2. Run the command Dev Containers: Reopen in Container.
3. Wait for container setup to complete (.devcontainer/post-create.sh runs automatically).
4. Open the integrated terminal to enter the container shell.

Build and start the container from the repository root:

devcontainer up --remove-existing-container --workspace-folder .

Run:

devcontainer exec --workspace-folder . -- bash -l

to enter the container shell.

The devcontainer exposes a browser desktop at port 6080.

1. Open http://localhost:6080 in your browser.
2. If you are using VS Code, you can also open the forwarded 6080 port from the Ports panel.

Use Chrome, Edge, or Firefox. Safari is not supported for this display workflow.

After the container is up, you can run demos with:

uv run python -m lerobocasa.demos.demo_kitchen_scenes

After selecting options in the terminal, a window with a scene should pop up on the desktop in the web browser.

If you need to force setup steps manually inside the container:

uv run python -m lerobocasa.scripts.setup_macros
uv run python -m lerobocasa.scripts.download_kitchen_assets

You can start a simulation node with:

uv run python -m lerobocasa.launch.simulation_node \
  --policy-port 8000

The policy port flag allows you to connect to the node with some policy server later on.

Assuming that you have a simulation node running, you can replay a lerobot v3.0 dataset by starting the example replay policy server in another shell:

uv run --with lerobot examples/replay_policy_server.py \
  --dataset_repo_id robotgeneralist/PickPlaceCounterToCabinet_pretrain \
  --port 8000

If the replay crashes without any explicit error, try rebuilding the container. Sometimes OOM errors are thrown. Not sure why.

First try at training and evaluating models can be found on this branch in our internal openpi repo. If the policy server is running inside a devcontainer on a remote machine, remember to properly forward appropriate ports, e.g., for jerryrig:

ssh -N -L 0.0.0.0:18000:127.0.0.1:8000 jerryrig

Then the simulation node can be run with:

uv run python -m lerobocasa.launch.simulation_node --policy-host host.docker.internal --policy-port 18000

The policy does random movements on a 5k checkpoint, so I suspect bugs are present. But at least on a general communication level the setup is working.

Inside the simulation node, you can teleoperate the robot and record new trajectories. Press:

• t to toggle teleoperation,
• Enter to start / stop recording,
• p to connect or disconnect from a policy server.

Note: If using SpaceMouse elsewhere, you may need to modify SPACEMOUSE_PRODUCT_ID in lerobocasa/macros_private.py.

If you want to verify the recordings, you can replay raw recordings using the replay_recording_server.py:

uv run --with lerobot examples/replay_recording_server.py \
  --recordings_dir recordings \
  --port 8000

To convert the recordings into a lerobot v3.0 dataset and upload to HF:

uv run python -m lerobocasa.converters.convert_recordings_lerobot_v3 \
    --recordings-dir recordings \
    --output-dir /tmp/pickplace_target_v3 \
    --upload-repo-id robotgeneralist/lerobocasa_custom_recordings \
    --overwrite

Explore 2500+ kitchen scenes:

uv run python -m lerobocasa.demos.demo_kitchen_scenes

View and interact with both human-designed and AI-generated objects:

uv run python -m lerobocasa.demos.demo_objects

Note: By default, this demo shows objaverse objects. To view AI-generated objects, add the flag --obj_types aigen.

THIS IS PART OF THE ORIGINAL ROBOCASA REPO. HAVEN'T YET VERIFIED IF THE THING BELOW WORKS

You can create environments using gym wrappers and run rollouts:

import gymnasium as gym
import lerobocasa
from lerobocasa.utils.env_utils import run_random_rollouts

env = gym.make(
   "lerobocasa/PickPlaceCounterToCabinet",
    split="pretrain", # use 'pretrain' or 'target' kitchen scenes and objects
    seed=0 # seed environment as needed. set seed=None to run unseeded
)

# run rollouts with random actions and save video
run_random_rollouts(
    env, num_rollouts=3, num_steps=100, video_path="/tmp/test.mp4"
)

Code: MIT License

Assets and Datasets: CC BY 4.0

RoboCasa365:

@inproceedings{robocasa365,
  title={RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots},
  author={Soroush Nasiriany and Sepehr Nasiriany and Abhiram Maddukuri and Yuke Zhu},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}

RoboCasa (Original Release):

@inproceedings{robocasa2024,
  title={RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots},
  author={Soroush Nasiriany and Abhiram Maddukuri and Lance Zhang and Adeet Parikh and Aaron Lo and Abhishek Joshi and Ajay Mandlekar and Yuke Zhu},
  booktitle={Robotics: Science and Systems (RSS)},
  year={2024}
}