HIW 500

HIW-500 (Humanoids In-the-Wild, 500 hours) is a robot learning dataset from the BitRobot Foundation, with Unitree and Hugging Face credited in the citation. It contains human whole-body teleoperation of a Unitree G1 humanoid performing household tasks in real homes in Southeast Asia, rather than in lab-only scenes. Layout, lighting, clutter and object state vary between episodes.

Each episode combines head and wrist camera streams (RGB and stereo, 480P, 30 FPS) with robot state and action traces (29-DoF joints, end-effector, IMU, odometry), language annotations, episode info, and camera intrinsics and extrinsics. Robot subtasks carry fine-grained action labels. Tasks listed include setting the table, restocking the fridge, kitchen organization, sweeping the floor, clothes washing, picking up trash, hanging keys and hangers, and cleaning up a room. Creator coverage describes the dataset as over 10 TB across 10+ tasks.

The data is hosted on Hugging Face as raw ROS bag/MCAP recordings, with a LeRobot-format version for robot learning workflows. Creator coverage suggests the LeRobot version was still pending at the time. The public dataset is under CC BY 4.0. BitRobot also offers commercial licensing for additional humanoid data and custom data collection on request. A roadmap item mentions releasing more tasks and environments.