feat(envs): add RoboCerebra long-horizon manipulation benchmark

[pkooij]

Summary

• Adds new robocerebra env type wrapping LIBERO's libero_10 suite with RoboCerebra-specific defaults (256×256, 20 FPS, camera keys image/wrist_image matching the HF dataset)
• Fixes a gap in the LIBERO factory: camera_name_mapping was accepted by LiberoEnv.__init__ but never forwarded through create_libero_envs → _make_env_fns → _make_env; now correctly propagated end-to-end (also benefits existing LiberoEnv users who pass --env.camera_name_mapping via CLI)
• Adds robocerebra optional dependency group in pyproject.toml (aliases lerobot[libero])

Dataset

The dataset CollisionCode/RoboCerebra_lerobot_v3.0 is already in LeRobot v3.0 format — no conversion needed.

Camera keys in the dataset match our defaults:

• observation.images.image (agent-view)
• observation.images.wrist_image (wrist)

Files changed

File	Change
src/lerobot/envs/configs.py	Add RoboCerebraEnv; fix LiberoEnv.create_envs to forward camera_name_mapping
src/lerobot/envs/libero.py	Thread camera_name_mapping through create_libero_envs → _make_env_fns → _make_env
pyproject.toml	Add robocerebra extra (= lerobot[libero], Linux only)
tests/envs/test_robocerebra_env.py	10 unit tests (no LIBERO install needed — factory fully mocked)
docs/source/robocerebra.md	Install, dataset, task table, eval commands, config reference, citation

Test plan

• 10 unit tests pass (pytest tests/envs/test_robocerebra_env.py) — no LIBERO needed, all env creation mocked
• pre-commit run -a passes on all changed files
• Eval smoke test (requires Linux + LIBERO + GPU):

  # Test existing libero path (user-provided command)
  lerobot-eval \
      --policy.path=pepijn223/smolvla_libero \
      --env.type=libero \
      --env.task=libero_spatial \
      --eval.batch_size=1 \
      --eval.n_episodes=1 \
      --eval.use_async_envs=false \
      --policy.device=cuda \
      '--env.camera_name_mapping={"agentview_image": "camera1", "robot0_eye_in_hand_image": "camera2"}' \
      --policy.empty_cameras=1
  
  # Test new robocerebra env type
  lerobot-eval \
      --policy.path=pepijn223/smolvla_libero \
      --env.type=robocerebra \
      --env.task=libero_10 \
      --eval.batch_size=1 \
      --eval.n_episodes=1 \
      --eval.use_async_envs=false \
      --policy.device=cuda

Publishing proposal

The dataset is already on HuggingFace at CollisionCode/RoboCerebra_lerobot_v3.0. To mirror it under the lerobot/ org:

huggingface-cli repo create robocerebra --type dataset --organization lerobot
huggingface-cli upload lerobot/robocerebra CollisionCode/RoboCerebra_lerobot_v3.0 --repo-type dataset

Requires lerobot/ org write access — can be done separately after merge.

🤖 Generated with Claude Code