HiPHI Dataset Delivers 617.5 Hours of Sub-Millimeter Human Motion Capture with 245.7 Hours of Object Interaction
HiPHI provides the largest synchronized human-object motion capture resource to date. Policies trained on it demonstrate scale-dependent gains and successful transfer to the Unitree G1. The dataset directly addresses data limitations that have constrained embodied humanoid learning.
The white paper details release of the HiPHI benchmark, a 617.5-hour whole-body dataset captured via optical systems. It adds synchronized object meshes and trajectories for 245.7 hours of carrying, pushing, and pulling actions. FrameNet linguistic categories guided collection to ensure systematic coverage of whole-body motion beyond prior narrow laboratory sets.
Scale experiments show reinforcement learning policies improve monotonically with added hours. Sim-to-real transfer succeeds on the Unitree G1, confirming that sub-millimeter precision and object grounding reduce distribution shift compared with internet video or unsynchronized mocap. Benchmark metrics quantify gains in motion diversity and interaction success.
Existing datasets either lack physical state precision or interaction breadth. HiPHI closes both gaps simultaneously. Operational impact appears in reduced sample complexity for contact-rich tasks and measurable policy robustness on physical hardware.
Next deployments will test whether performance continues to scale past current volumes and whether FrameNet coverage extends to multi-agent or tool-use sequences on additional platforms.
Noitom Robotics: HiPHI-trained policies reach 75% success on 15 unseen interaction tasks on Unitree G1 hardware by end of 2025.
Sources (2)
- [1]Primary Source(https://content.knowledgehub.wiley.com/hiphi-a-large-scale-benchmark-for-high-precision-human-motion-and-object-interaction/)
- [2]Supporting Source(https://unitree.com/g1)