Learning Force-Reactive Humanoid Loco-Manipulation

Published 2026-10-28Automatically prepared brief

The paper introduces FLORE, a framework for learning force-reactive humanoid loco-manipulation using robot-free demonstrations and a force-reactive diffusion policy.

The research question addresses the challenge of controlling both motion and contact force in humanoid robots, where force coordination across hands and the body is essential and cannot be achieved through kinematic trajectories alone.

The proposed approach involves collecting demonstrations using force-sensing grippers and wearable trackers, followed by training a force-reactive diffusion policy that uses vision, proprioception, and measured wrenches to predict actions and force commands. A hybrid compliance whole-body controller then executes these commands while maintaining balance.

The authors report that in simulation, the controller effectively tracks force commands with minimal impact on motion accuracy. In real-world tests across three tasks, the force-reactive policy outperformed a position-only policy, succeeding in 80% of rollouts compared to 26.7%.

About this brief

This article summarizes a public preprint abstract. Its findings have not been independently reproduced for this publication. Consult the paper for methods, evaluation scope and limitations. Automated briefs may contain errors; the original source takes precedence.