CVPR 2026: RePHO reconstructs physically plausible human-object interactions

A team of researchers – Dingbang Huang, Etienne Vouga, Qixing Huang, and Georgios Pavlakos – has introduced RePHO, a novel framework for recovering physically plausible human-object interactions (HOI) from a single monocular video. While existing kinematic-based methods can produce visually plausible motion, they frequently suffer from physically implausible artifacts such as interpenetration (body parts clipping through objects) and object floating (objects not resting on surfaces). RePHO addresses these issues by starting from a kinematic estimate and then refining it using a reinforcement learning (RL) policy. This policy is trained to reproduce the interaction in a physics simulator, ensuring that the reconstructed motion obeys physical laws.

A key technical challenge is that kinematic estimates are inherently noisy, making naive RL training prone to failure. To overcome this, RePHO introduces an adaptive sampling strategy with a dual self-updating mechanism. This mechanism identifies the frames that contain the most informative and reliable kinematic reconstructions, progressively improving the quality of the physics-guided refinement. The result is a physically consistent HOI sequence that avoids common artifacts. The method was evaluated on two standard HOI benchmarks and demonstrated clear improvements in physical plausibility metrics over existing state-of-the-art approaches. This work, published at CVPR 2026, opens up new possibilities for more realistic human-object interaction modeling in applications like graphics, robotics, and AR/VR.