Learning Humanoid Loco-Manipulation for Transporting Unwieldy Construction Objects
Abstract
Enabling robust loco-manipulation for transporting unwieldy objects is a critical prerequisite for deploying humanoid robots in construction. However, this task is complicated by severe physical constraints, including irregular object geometries, limited actuator torque, and high moments of inertia caused by eccentric centers of mass. These factors induce complex robot-object dynamics that threaten locomotion stability — a challenge further exacerbated by the scarcity of domain-specific interaction data. Unlike previous approaches that rely on either large-scale motion capture data or torso-supported strategies (“hugging motion”) to carry large objects, our method introduces (1) a scalable pipeline for synthesizing high-fidelity demonstrations in simulation and (2) a residual policy that refines locomotion priors for dexterous end-effector-based transportation. Experimental results demonstrate that our residual policy successfully transports unwieldy objects (up to 5 kg) while maintaining balance, outperforming both learning-from-scratch and fine-tuning baselines. This study serves as an initial exploration into equipping humanoids with the complex loco-manipulation skills required for construction sites.