Long-Horizon Skill Learning

The question

Longer manipulation tasks require more than a sequence of isolated actions. A robot must interpret a scene, organize skills, and use feedback to revise its plan.

The approach

The framework combines scene understanding, a Fast Learner and a hierarchical Skill Library. A language model generates task sequences and control code, while interaction history supports the extraction and reuse of skills.

From method to robot

The work includes Meta-World experiments and physical demonstrations involving Franka manipulation and a mobile platform. Tasks include pick-and-place, rotation, unscrewing and disassembly. The mobile manipulation project describes the supporting system integration.

Publication