IEEE Spectrum breed · Technology

HiPHI: A Large-Scale Benchmark for High-Precision Human Motion and Object Interaction

A new large-scale motion capture dataset, HiPHI, has been developed to address data limitations hindering humanoid robot learning. The dataset aims to improve robot task performance by providing comprehensive motion data.

Humanoid robots need extensive data for learning movement and interaction. Internet videos lack precise physical states, and lab systems cover limited actions, creating a data gap.

HiPHI is a 617.5-hour whole-body motion dataset captured with optical motion capture at sub-millimeter accuracy. It includes 245.7 hours of human-object interaction with synchronized object trajectories and meshes.

The dataset's coverage is organized using FrameNet, a linguistic framework for human action, ensuring broad whole-body motion is captured.

The white paper introduces a benchmark suite to measure motion diversity and interaction grounding for rigorous evaluation.

Reinforcement learning policies trained on HiPHI show improved performance with scale and have been transferred to a physical Unitree G1 humanoid robot.

The dataset is accessible via a registration process on a dedicated hub for researchers in embodied AI and Physical AI.

AI-samenvatting op basis van de bron.