Integrating Model-based and Model-free Reinforcement Learning for Peg-in-hole Assembly

Establish integrated reinforcement learning frameworks that combine model-based and model-free methods for peg-in-hole assembly, including accurate learning of transition dynamics and mechanisms to balance planning with the learned model against direct interaction with the real environment.

Background

The survey outlines complementary strengths of model-free RL (flexibility, ease of implementation) and model-based RL (sample efficiency, fast convergence), and reviews recent attempts to combine them.

The authors explicitly pose the open question of how to combine these approaches effectively, highlighting two key challenges: learning accurate dynamics models and balancing model-based planning with direct environment interaction during policy learning.