Reconfiguring Procedural Knowledge for Compositional Robot Skill Adaptation

Impact 7 arXiv cs.AI

Summary

  • Researchers introduce RecPo, a framework for compositional robot skill adaptation that reconfigures existing procedural knowledge for new goals.
  • RecPo represents skills task-agnostically through multiple procedure views and materializes executable skill sequences by composing stored transitions.
  • The framework maintains a knowledge base of skill-induced transitions, interpreting task specifications to bind skills for execution.
  • On an extended Franka Kitchen benchmark, RecPo achieves 91.9% ordered completion with 1.0 ms task inference per episode.
  • Real-robot studies demonstrate adaptation through local memory edits without requiring task-specific retraining.

Why it matters

RecPo enables efficient robot skill adaptation and long-horizon task execution through modular memory reconfiguration without retraining, advancing practical robotics deployment.

Who should care