BitRobot Network partnered with Unitree Robotics for hardware and Hugging Face for hosting to release HIW-500, a 10-terabyte collection of teleoperated humanoid movements gathered across 12 real homes rather than labs, capturing long-horizon household tasks with stereo cameras, full robot state, and sub-task labels on Unitree G1 platforms.
1/ Introducing HIW-500 (Humanoids-in-the-Wild 500): the largest open-source humanoid teleop dataset collected in real homes
Built w/ @UnitreeRobotics @huggingface across 12 homes in Southeast Asia, it covers: > 500+ hrs > 23K+ episodes > 10+ TB > 10+ household tasks
The full ROSbag files and a ready-to-use LeRobot version sit publicly on Hugging Face, letting teams pull 29-DoF actions, RGB and IR video, and episode annotations without custom conversion steps.
While the 23,000 episodes target messy domestic settings, the collection stays limited to Southeast Asian homes and ten task families, leaving open how well models will transfer to other regions or entirely new activities.
Users are excited about the HIW-500 humanoid teleop dataset release on Hugging Face because it supplies valuable real-home data for robot training and advances open robotics.
No Digg Deeper questions have been answered for this story yet.
more and more large-scale robotics datasets on the @huggingface hub!!
1/ Introducing HIW-500 (Humanoids-in-the-Wild 500): the largest open-source humanoid teleop dataset collected in real homes
Built w/ @UnitreeRobotics @huggingface across 12 homes in Southeast Asia, it covers: > 500+ hrs > 23K+ episodes > 10+ TB > 10+ household tasks
The largest open-source humanoid teleop dataset EVER just dropped on @LeRobotHF @huggingface
https://huggingface.co/datasets/BitRobot/HIW-500-LeRobot
We re-encoded the full dataset into LeRobot format: ~10TB → ~2TB, no loss of fidelity. Same trajectories, a fraction of the footprint, far easier to stream and train on.
Explore the dataset in LeRobot Visualizer: live 3D render of the robot, synced camera feeds, subtask annotations and language instructions, all in your browser.
Most importantly: Unitree G1 is supported in LeRobot, meaning you can load HIW-500 and start training imitation / VLA policies on 20,000 real-home humanoid episodes on your G1.
Huge thanks to @BitRobotNetwork for sharing the data. Go build. 🤖
1/ Introducing HIW-500 (Humanoids-in-the-Wild 500): the largest open-source humanoid teleop dataset collected in real homes
Built w/ @UnitreeRobotics @huggingface across 12 homes in Southeast Asia, it covers: > 500+ hrs > 23K+ episodes > 10+ TB > 10+ household tasks
8/ Full 10TB dataset in ROSbag + @LeRobotHF ↓
Dataset: https://huggingface.co/datasets/BitRobot/In-The-Wild-Humanoid-Dataset LeRobot: https://huggingface.co/datasets/BitRobot/In-The-Wild-Humanoid-Dataset-LeRobot
For research / commercial use of the expanded HIW-500 dataset → https://bitrobot-foundation.github.io/humanoids-in-the-wild-500-hours/ or reach out to us directly!
2/ Humanoid robots have made tremendous progress in locomotion.
But whole-body mobile manipulation in real homes remains one of the biggest unsolved challenges in robotics.
Most datasets are collected in controlled environments. Real homes are messy, unpredictable, and full of edge cases.
To build "robot butlers" that work in the real world, we need data from the real world. HIW-500 is a step in that direction.
5/ The dataset contains 10+ long-horizon household tasks, with some episodes lasting more than 8 minutes.
And 10+ long-horizon tasks spanning navigation, object manipulation, and multi-step activities that require coordination across the entire robot body.
3/ To build this dataset, we used the following hardware setup:
> Unitree G1 humanoids > Stereo head camera > IR stereo wrist cameras > Grippers
This was designed to capture both whole-body motion and manipulation in everyday household environments.
7/ This project represents months of work alongside incredible partners.
Anyone who has tried teleoperating a humanoid robot knows how challenging it is. Collecting thousands of high-quality whole-body demonstrations inside real homes is an entirely different level of difficulty.
Huge thanks to @UnitreeRobotics for the hardware support.
And to @huggingface for partnering with us once again after our previous work on open-sourcing large-scale urban navigation datasets.
@LeRobotHF @huggingface The hardware is here. The data is open. Go build 🦾
4/ Each episode includes:
> Head camera RGB (480p, 30 FPS) > Wrist camera RGB + IR (480p, 30 FPS) > Full robot state & actions (29 DoF, IMU, odometry) > Episode and sub-task annotations
6/ With thousands of demonstrations per task and detailed sub-task annotations, researchers can train and evaluate models across multiple levels of abstraction.
@BitRobotNetwork @UnitreeRobotics @huggingface 🦾
@BitRobotNetwork @UnitreeRobotics @huggingface 🔥
@BitRobotNetwork @UnitreeRobotics @huggingface Great
@BitRobotNetwork @UnitreeRobotics @huggingface That's huge
@BitRobotNetwork @UnitreeRobotics @huggingface That's great! Putting web3 x robotics on the map.
@BitRobotNetwork @UnitreeRobotics @huggingface Massive step forward for real world humanoid training
@LeRobotHF @huggingface Open robotics in action 🦾🦾🦾
@BitRobotNetwork @UnitreeRobotics @huggingface A real training environment for robots produces great results. Kudos!
This data is very useful. Because no matter how much we try to obtain robot data by designing engines and pipelines from different places, the highest quality data is always the data collected directly from the robot. Especially the fact that this is a deployment environment rather than a lab environment is very valuable. Great work!
@BitRobotNetwork @UnitreeRobotics @huggingface 棒棒棒
