Fuse Existing Knowledge into Model‑Free Reinforcement Learning for Peg‑in‑Hole Assembly

Develop a principled method to fuse existing knowledge into model‑free reinforcement learning for robotic peg‑in‑hole assembly in a natural and effective manner, so that prior information can be incorporated during policy learning without relying on explicit contact state recognition.

Background

The paper reviews learning-from-environments approaches, noting that model-free reinforcement learning has been applied to peg-in-hole assembly but typically suffers from data inefficiency. Several studies attempt to incorporate prior knowledge or demonstrations to improve practicality.

Despite these advances, the authors point out that for peg-in-hole assembly it is not clear how to integrate prior information into a model-free learning pipeline in a principled way. Addressing this gap would enable more efficient and robust policy learning by leveraging available knowledge about the task or environment.