HSImul3R: Reconstructing Simulatable Human-Scene Interactions Directly from Human Video

Daxiao Robotics, alongside NTU S-Lab and the Shanghai AI Laboratory, has unveiled HSImul3R, a framework designed to reconstruct physically simulatable human-scene interactions from raw human demonstration videos. By extracting kinematically valid poses, dynamic scene contacts, and interactive trajectories, the approach bridges unstructured monocular video with physics simulation engines, supplying structured data for robotic policy training.