RL-based Shape Control of Rod-driven Soft Arms Using Strain Models

The continuous shapes of soft continuum robots with infinite degrees of freedom have potential applications in unstructured environments. However, their nonlinear behavior and compliance make them difficult to model and control. In this study, we propose a shape control scheme for rod-driven soft continuum robots and train the policy by using the reinforcement learning (RL) algorithmic implementation in a simulation environment with an theoretical strain model. The trained policy performs continuous, entire shape control (with maximum distance less than 8 mm) as well as the simulation model. By substituting the tracked positions in real robot from the strain model, the proposed control scheme bridges the large sim2real gap and reduces the distance of controlled shape to the target by 50%, from 15.93 to 7.92 mm. Our method demonstrates an effective and feasible way to control the entire 3D shape and configuration of a soft arm.