The HiPHI dataset uses the FrameNet linguistic framework to systematically capture high-precision human-object interactions. It provides synchronized object trajectories and meshes that internet videos lack. This allows reinforcement learning policies to transfer directly to physical humanoid robots.