Co²Skill: Whole-Body Control via Skill Composition for Long-Horizon Human-Environment Interaction

This research presents Co²Skill, a framework that combines whole-body scene interaction and dexterous object manipulation through a unified policy. The method uses task and phase-dependent observation masks, along with a goal-conditioned curriculum, to enable sequential task execution and simultaneous scene interaction. It is evaluated on tasks such as sitting, standing, climbing, and goal-directed manipulation, demonstrating skill composition in indoor environments.