Researchers have released HiPHI, a 600-plus-hour motion capture dataset built to help humanoid robots learn how humans move and interact with objects.
The dataset uses optical motion capture to record full-body human movement with sub-millimeter accuracy, tracking both body motion and mesh-level object trajectories. It's organized around FrameNet, a linguistic framework for categorizing human action primitives. The team also built a benchmark suite that tests motion diversity, interaction grounding, object consistency, and physical AI performance. According to the researchers, HiPHI covers significantly more motion variety than existing datasets while keeping high interaction fidelity.
The problem this targets is real: internet video is abundant but lacks precise physical grounding, while lab-based motion capture datasets are precise but narrow in scope. That gap has been a bottleneck for training humanoid robots on the full range of physical tasks they'd need to handle outside a lab. A dataset that's both large and physically accurate could matter more for humanoid progress than another flashy demo video.
Whether HiPHI actually closes the gap depends on adoption - benchmarks are only as useful as the models trained and tested against them.