Shanghai AI Laboratory

HSImul3R, a human–scene interaction reconstruction framework from Daxiao Robotics, Nanyang Technological University S-Lab, and Shanghai AI Laboratory, targets uncalibrated sparse-view input and produces simulation-ready scenes. The paper, accepted at ECCV 2026, describes a closed-loop optimization combining reinforcement learning for the scene with direct simulation-reward optimization, then transfers the refined motion to a Unitree G1 humanoid. On HSIBench, interaction stability reaches 53.68%, 30.56%, and 13.92% across three task tiers, while interpenetration drops from 69.51% to 22.90%. Code and paper are public; license was not stated.