HIW-500: Humanoids In-the-Wild Dataset

https://bitrobot-foundation.github.io/humanoids-in-the-wild-500-hours/

HIW-500: Humanoids In-the-Wild Dataset is a large-scale dataset for whole-body humanoid robot learning in natural home environments. It captures human teleoperation demonstrations on Unitree G1 across real homes in Southeast Asia, where layouts, object states, lighting, clutter, and operator styles vary from episode to episode.

The dataset is designed for research on mobile manipulation, bimanual interaction, long-horizon household skills, imitation learning, and general-purpose robot learning from in-the-wild demonstrations.

Dataset Overview

• 500+ hours of humanoid robot demonstrations
• 23K+ episodes
• Around 10 TB of data
• 10+ household tasks
• 12 real homes
• 161 subtask labels
• 148K+ subtask annotations

Data Modalities

Each episode records human whole-body teleoperation of Unitree G1 in real homes. The dataset combines synchronized visual observations, robot states, actions, and metadata.

Camera Streams

• Head camera: RGB stereo, 480p, 30 FPS
• Wrist camera: RGB, stereo IR, 480p, 30 FPS

Robot State and Actions

• 29-DoF joint states
• End-effector state
• IMU
• Odometry
• Action traces from human whole-body teleoperation

Metadata

• Language annotations
• Episode information
• Camera intrinsics and extrinsics

Dataset Statistics

👁 Task duration

👁 Task episodes

👁 Average duration

Dataset Access

The dataset is hosted on Hugging Face in two formats:

• Raw ROS bag / MCAP recordings: https://huggingface.co/datasets/BitRobot/HIW-500
• LeRobot format: https://huggingface.co/datasets/BitRobot/HIW-500-LeRobot

For raw data from the Building Children Table task, see our challenge dataset: https://huggingface.co/datasets/BitRobot/2026-humanoid-ikea-assembly-challenge

License

The public HIW-500 dataset is released under the CC BY 4.0 license. If you're interested in additional datasets similar to HIW-500, be it for commercial or academic purposes, pls contact us.

Citation

If you use this dataset, please cite:

@misc{hiw500_2026,
 title={HIW-500: Humanoids In-the-Wild Dataset for Robot Learning},
 author={BitRobot and Unitree and Hugging Face},
 year={2026},
 howpublished={\url{https://bitrobot-foundation.github.io/humanoids-in-the-wild-500-hours/}}
}