Humanoid Loco-Manipulation

Humanoid loco-manipulation studies how a humanoid can move through its environment while reaching, grasping, pushing, or carrying objects without losing balance. Common approaches combine model-based whole-body control with reinforcement learning or imitation, often using motion retargeting, hierarchical skills, and vision-language or world-model planners. Human demonstrations, simulation-generated data, and robot experience train policies that are evaluated in simulation and on physical humanoids. A central challenge is making these systems reliable across unfamiliar scenes and object loads, especially during contact, while keeping actions physically feasible and enabling recovery from errors.