Residual Whole-Body Control for Contact-Aware Humanoid Loco-Manipulation with Phase-Conditioned Motion Priors

Humanoid object transport couples high-dimensional locomotion with non-smooth contact interactions. Retargeted demonstration priors can fail under contact mismatches, payload changes, and dynamic perturbations. We present residual learning for humanoid object carrying (ResCarry), a residual whole-body control framework built on phase-conditioned motion priors. It first trains an object-centric command prior with vector-quantized codebooks and a causal transformer to capture multi-phase interactions. The prior is then fixed, and a bounded residual whole-body policy learns joint-reference corrections from proprioceptive, contact, and object-state feedback. To reduce catastrophic failures, an empirical upper-tail hazard metric from a multi-atom quantile critic adapts the expected-cost penalty in proximal policy optimization (PPO). A deployable risk estimator slows command progression in high-hazard states. Across six contact-specified furniture manipulation tasks, ResCarry achieves 59.7% success in Isaac Lab. Zero-shot transfer to Unitree MuJoCo reaches 54.8%.