A new white paper introduces HiPHI, a large‑scale motion‑capture dataset that fills a critical data gap for training humanoid robots, showing that policies learned from it transfer to real hardware.
Core Innovations in HiPHI
- Dataset spans thousands of whole‑body motions, systematically organized by the FrameNet linguistic framework.
- Provides synchronized object trajectories and high‑resolution meshes, enabling realistic human‑object interaction scenarios.
- Reinforcement‑learning agents trained on HiPHI demonstrate performance gains that scale with dataset size.
- Sim‑to‑real transfer validated on a physical humanoid robot, proving the dataset’s practical value.
Why Existing Data Falls Short
Internet video and legacy motion‑capture libraries lack the breadth and systematic coverage needed for embodied AI. HiPHI addresses this by combining linguistic guidance with precise motion capture, ensuring that a wide spectrum of whole‑body actions is represented. The inclusion of object meshes and trajectories further bridges the gap between simulated training environments and the messy, interactive real world.
Implications for Humanoid AI
HiPHI’s scale and realism open new avenues for training robust, transferable policies. Researchers can now explore complex tasks—carrying, pushing, pulling—directly from data that mirrors real‑world physics. The successful sim‑to‑real transfer suggests a viable path toward rapid deployment of humanoid robots in industrial and service settings, potentially accelerating the adoption of embodied AI across sectors.