Towards Safe Control of Continuum Manipulator Using Shielded Multiagent Reinforcement Learning

Continuum robotic manipulators are increasingly adopted in minimal invasive surgery. However, their nonlinear behavior is challenging to model accurately, especially when subject external interaction, potentially leading poor control performance. In this letter, we investigate the feasibility of adopting a model-free multiagent reinforcement learning (RL), namely deep Q network (MADQN), 2-degree freedom (DoF) cable-driven continuum surgical manipulator. The robot formulated as one DoF, agent problem MADQN framework improve efficiency. Combined with shielding scheme that enables dynamic variation action set boundary, leads efficient and importantly safer robot. Shielded enabled perform point trajectory tracking submillimeter root mean square errors under loads, soft obstacles, rigid collision, which common interaction scenarios encountered by manipulators. controller was further proven be effective miniature high structural nonlinearitiy, achieving accuracy payload.