Humanoid robots need large amounts of data to learn how to balance, move, and interact with objects, but existing sources have clear limits. Internet video shows diverse behavior but cannot capture precise physical states, while laboratory motion capture records accurate movement but usually covers only a narrow set of actions. The HiPHI dataset is designed to fill that gap.
According to a white paper published by IEEE Spectrum and sponsored by Noitom Robotics, HiPHI was captured with optical motion capture at sub-millimeter accuracy. It includes synchronized object trajectories and meshes, making human-object interaction data useful for teaching robots tasks such as carrying, pushing, and pulling. The dataset also introduces a benchmark suite for measuring motion diversity and interaction grounding.
The paper reports that reinforcement learning policies trained on the dataset improve with scale and can be transferred to a physical Unitree G1 humanoid robot. Because the source is a sponsored white paper, the results are presented by the vendor rather than independently verified, but the dataset itself offers a concrete step toward more data-rich humanoid robot training.