Synergizing Reinforcement Learning and Classic Control: Optimized Policy Search for Continuum Robots

Continuum robots pose significant challenges in accurate control due to their high flexibility and nonlinearity. To address these issues, system identification and linearization techniques have been explored extensively. However, selecting an optimal controller for the linearized model remains a critical challenge. This paper presents and analyzes three frameworks that make a fusion of the reinforcement learning (RL) and conventional controllers: RL-tuned LQR (RL-LQR), model-reference RL (MRRL), and RL from demonstration (RLfD). These frameworks enable optimal parameter selection for conventional controllers and enhance policy performance beyond conventional methods. Experimental validation on a continuum robot platform demonstrates the effectiveness of the proposed approaches in comparison with conventional controllers.