ember-lab-berkeley/robocasa365-pretrain-composite

Pretraining (Human) — composite

Human teleoperation data for 235 composite (multi-step) tasks (~100 demos/task). The long-horizon half of RoboCasa365's 482-hour pretraining-human dataset.

Part of the RoboCasa365 collection. Flat LeRobot v3.0 mirror of RoboCasa365 — standard layout, drop-in loadable.

Stats

• Episodes: 24,687
• Frames: 27,610,913 (20 fps → 383 h)
• Tasks: 4,428 (natural-language phrasings; underlying RoboCasa task classes: 235)
• Cameras: 3 × 256×256 h264 video (robot0_agentview_left / _right / _eye_in_hand)
• Robot: PandaOmron (Panda 7-DoF arm + Omron mobile base + torso lift)

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("ember-lab-berkeley/robocasa365-pretrain-composite")

Feature schema (shared across all 6 robocasa365 repos)

• observation.state: float64[16] — base_pos(3) + base_quat(4) + eef_pos_rel(3) + eef_quat_rel(4) + gripper_qpos(2)
• action: float64[12] — base_motion(4) + control_mode(1) + eef_delta_pos(3) + eef_delta_aa(3) + gripper(1)
  • control_mode convention: -1 = arm active (eef delta + gripper drive), +1 = base active (base_motion drives). Source of truth: robosuite.controllers.composite.HybridMobileBase.
• observation.images.robot0_agentview_left / _right / _eye_in_hand: video, 256×256×3

License

MIT, inherited from upstream RoboCasa365.

Citation

RoboCasa365 paper (ICLR 2026)