lerobot-genesis

The bridge between the Genesis physics simulator and Hugging Face LeRobot.

Genesis is a fast, open robotics simulator; LeRobot is the open standard for robot datasets and policies. There is no official link between them — this package is that link, the same role NVIDIA's IsaacLab-Arena plays for Isaac Lab. It lets you expose any Genesis scene as a Gymnasium environment, record rollouts as a LeRobotDataset, and (from v0.2) evaluate a trained LeRobot policy in Genesis through the standard lerobot-eval workflow.

   Genesis scene ──▶ GenesisEnv (gym.Env) ──▶ LeRobotDataset ──▶ train a LeRobot policy
                          ▲                                              │
                          └────────────────── lerobot-eval ◀────────────┘   (v0.2)

Install

pip install "lerobot-genesis[all]"   # bridge + genesis-world + lerobot

genesis-world needs a CUDA GPU and a few system libraries — see the Genesis install guide. The bridge itself is pure Python; pip install lerobot-genesis (no extras) imports without either heavy dependency present.

Quickstart — record a dataset

Drive a robot in a Genesis scene and save the rollout in LeRobot's format:

from lerobot_genesis import GenesisEnv, GenesisRobotDriver, LeRobotDatasetSink, record_episodes

driver = GenesisRobotDriver(robot="path/to/robot.urdf")          # any URDF / MJCF / USD
env = GenesisEnv(driver, task="reach the target")
sink = LeRobotDatasetSink(
    "me/my-genesis-dataset",
    fps=30,
    state_dim=driver.state_dim,
    action_dim=driver.action_dim,
    image_shape=driver.image_shape,
    task="reach the target",
)

record_episodes(env, policy=lambda obs: env.action_space.sample(), sink=sink, n_episodes=10)

A runnable end-to-end example (a Franka arm, no asset download) lives in examples/record_franka.py.

Quickstart — evaluate a policy in Genesis

lerobot-genesis-eval is lerobot-eval with the Genesis env registered, so a LeRobot policy (ACT, SmolVLA, GR00T, …) can be rolled out in a Genesis scene. Smoke-test the wiring with a from-scratch policy (no checkpoint needed):

lerobot-genesis-eval --env.type=genesis --policy.type=act \
    --eval.batch_size=1 --eval.n_episodes=1 --eval.use_async_envs=false --policy.device=cpu

Point --policy.path at a trained checkpoint to score it, and --env.task at your own registered Genesis task. Training GR00T? lerobot_genesis.groot.write_modality_json emits the meta/modality.json it needs.

How it's designed

The simulator-specific work sits behind one small Protocol, so the Gymnasium contract and the recorder are testable on a CPU with no GPU and no upstream installed.

• GenesisEnv — a gymnasium.Env over a Genesis scene. It owns the gym contract (the 5-tuple step, info["is_success"], the pixels/agent_pos observation, a continuous action Box) and delegates the physics to an injected SceneDriver.
• GenesisRobotDriver — the reference SceneDriver: loads a robot, maps a normalised action onto a chosen subset of joints, steps physics, and reads a camera and joint state. Bring your own driver for richer scenes (objects, tasks, rewards).
• record_episodes / LeRobotDatasetSink — roll a policy and write a LeRobotDataset (the verified v3 create → add_frame → save_episode → finalize loop).
• GenesisEnvConfig — registers Genesis with LeRobot as --env.type=genesis.

Relationship to gym-genesis

Hugging Face's gym-genesis provides raw Gymnasium task scenes for Genesis. This package is complementary: it adds the LeRobot integration layer — dataset recording, env registration, and policy evaluation — and can wrap a gym-genesis env just as well as your own.

Roadmap

• v0.1 — GenesisEnv, reference driver, LeRobotDataset recorder, --env.type=genesis.
• v0.2 — EnvHub make_env + lerobot-eval policy evaluation; GR00T modality.json adapter; native vectorised (multi-env) rollouts.

License

MIT. An independent, community project — not affiliated with the Genesis or LeRobot teams.