Noitom Robotics released HiPHI, a large-scale motion capture dataset, to address the data gap in humanoid robot learning. The dataset includes 617.5 hours of whole-body human motion recorded with sub-millimeter accuracy. It is the company's first major update in the field of embodied AI since its earlier motion capture projects.
The HiPHI dataset contains 245.7 hours of human-object interaction data, featuring synchronized object trajectories and meshes. This enables the teaching of real-world tasks like carrying, pushing, and pulling to robots. The dataset is organized using FrameNet, a linguistic framework for human action.
"HiPHI is built to close the data gap limiting humanoid robot learning," said Noitom Robotics. "It provides the precision and diversity needed for training reinforcement learning policies." The dataset includes a benchmark suite for measuring motion diversity and interaction grounding.
The announcement follows Noitom's collaboration with IEEE Spectrum Magazine on exploring new technologies in embodied AI. The company emphasized that HiPHI will support sim-to-real transfer of policies to physical humanoid robots.
Noitom did not specify the exact deployment timeline for HiPHI and highlighted the challenge of scaling reinforcement learning policies. The company noted that the dataset will be made available to researchers and engineers in the near future.
Source: ieee
