A new dataset called HiPHI has been developed to address the data limitations in training humanoid robots for complex physical interactions. This large-scale dataset, comprising 617.5 hours of motion capture data, systematically covers a wide range of human actions and object interactions using FrameNet as a guiding framework. The data includes synchronized object trajectories and meshes, making it suitable for teaching robots real-world tasks. Policies trained on HiPHI have demonstrated improved performance with increased data scale and successful transfer to a physical Unitree G1 humanoid robot. AI
IMPACT This dataset could accelerate the development of more capable humanoid robots by providing richer training data for physical interaction and task execution.
RANK_REASON The cluster describes a new dataset and benchmark for AI research, presented in a white paper. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
