HIW-500: Humanoids In-the-Wild Dataset

Your browser does not support the video tag.

HIW-500 Humanoids In-the-Wild Dataset

An in-the-wild dataset

for whole-body humanoid robot learning.

II

Contributors

Overview

Built for learning from real homes, not lab-only scenes.

HIW-500: Humanoids In-the-Wild Dataset focuses on task execution in natural environments where layout,<br>object state, lighting, clutter, and human operating style vary from episode to episode.

The dataset is collected in real homes in Southeast Asia through human whole-body teleoperation of Unitree G1 ,<br>providing demonstrations for mobile manipulation, bimanual interaction, and long-horizon household skills.

0+<br>Hours

0+<br>Episodes

Data

0+<br>Tasks

Homes

Hardware

Acquisition setup

The hardware stack uses grippers for household manipulation, with a head camera<br>and wrist-mounted cameras for visual observations.

01<br>Gripper manipulation

End-effector grippers are used for household grasping, placing, and object interaction.

02<br>Head camera

Head-mounted camera provides a wider scene view for task context and navigation.

03<br>Hand stereo IR cameras

Wrist-mounted stereo IR cameras capture close-range observations around the robot hands.

Dataset

What data is collected?

Each episode records human whole-body teleoperation of Unitree G1 in real homes,<br>combining camera streams, robot states, action traces, and language annotations.

Head camera

RGB

Stereo

480P

30 FPS

Wrist camera

RGB

IR

Stereo

480P

30 FPS

State and actions

29-DoF Joints

End-Effector

IMU

Odometry

Metadata

Language annotation

Episode info

Camera intrinsics & extrinsics

Live demo

Open the Rerun preview in a new tab

Dataset Statistics

Task Coverage

Building children table

Hang hanger

Clean up the room

Setting the table

Restocking fridge

Kitchen organization

Hang keys on a hook

Move pillow to sofa

Sweep floor

Picking trash

Clothes washing

Task Episodes

Average Duration

Subtask Labels

161<br>Total subtask labels

148k+<br>Subtask annotation

Top 40<br>Shown in treemap

Every robot subtask is annotated with a fine-grained action label. Hover tiles for details.

Roadmap

Open Source timeline

June 2026

Release V1: 500+ hours of data

V2

Release more tasks and environments

LeRobot Integration

We host the dataset on Hugging Face in two formats: raw ROS bag / MCAP recordings and<br>a version in LeRobot format for robot learning workflows.

View raw dataset

View dataset in LeRobot format (Coming Soon)

Citation

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

License Humanoid Data for Training & Evaluation

Access our growing in-the-wild humanoid dataset or request custom data collection

Request Data Access