A Review of Whole-Body Control for Humanoid Robots: From Optimization-Based Methods to Learning-Based Generalist Policies

Humanoid whole-body control (WBC) coordinates locomotion and manipulation under intermittent contact and friction constraints. This review provides a unified analytical framework that compares optimization-based controllers with learning-based generalist policies through dynamics, contact feasibility, task objectives, command interfaces and execution feedback. Optimization-based controllers can enforce model-derived constraints online but depend on model fidelity, state estimation and timely computation. Learning-based policies encode motion and task regularities from reference data, rewards, motion priors and policy architectures. Trained policies can generate actions with low online computational cost, but reliable execution depends on coverage of states and commands during training and on task-relevant observations during deployment. The review organizes evidence across model-based control, reinforcement learning, imitation, motion priors, teleoperation and loco-manipulation. Generalization is assessed separately across motions, tasks, environments and embodiments, with tracking distinguished from autonomous task competence. Open challenges in hybrid constraint-aware learning and reliable long-horizon operation on physical humanoids are also discussed.