From Policy Actions to Whole-Body Dynamics: A Unified Projection-Based Framework for Versatile Legged Loco-Manipulation

Abstract

Learning-based locomotion policies are widely adopted for legged robots but face integration challenges with model-based whole-body optimization for loco-manipulation, as they output joint-space commands instead of dynamics-level variables that enable explicit enforcement of rigid-body dynamics and contact objectives. To resolve this, we introduce Policy Action Dynamics Projection (PADP), an optimization-based dynamics projection that maps joint-space policy outputs into contact-consistent whole-body variables (i.e., generalized accelerations, contact wrenches, and joint torques) while satisfying rigid-body dynamics and contact conditions. Building on this dynamics-level interface, we further develop a unified loco-manipulation controller that directly leverages a generic pretrained locomotion policy and integrates model-based manipulation tasks in a subsequent unified whole-body QP. This formulation enables precise and coordinated whole-body loco-manipulation performance that goes far beyond the policy’s original command space without retraining. Simulation and hardware experiments on both point-foot and wheeled-foot quadruped manipulators demonstrate stable dynamic loco-manipulation with improved agility and precision over conventional whole-body control baselines.