Pushing with Soft Robotic Arms via Deep Reinforcement Learning

Soft robots can adaptively interact with unstructured environments. However, nonlinear soft material properties challenge modeling and control. Learning-based controllers that leverage efficient mechanical models are promising for solving complex interaction tasks. This article develops a closed-loop pose/force controller for a dexterous soft manipulator enabling dynamic pushing tasks using deep reinforcement learning. Force tests investigate the mechanical properties of a soft robot module, resulting in orthogonal forces of $9-13$ N. Then, the policy is trained in simulation leveraging a dynamic Cosserat rod model of the soft robot. Domain randomization mitigate the sim-to-real gap while careful reward engineering induced pose and force control even without explicit force inputs. Despite the approximate simulation, the sim-to-real transfer achieved an average reaching distance of $34\pm 14$ mm ($8.1\%L\pm 3.4\%L$), an average orientation error of $0.40\pm 0.29$ rad ($23°\pm 17°$) and applied pushing forces up to $3$ N. Such performance is reasonable for the intended assistive tasks of the manipulator. The experiments uncovered that the soft robot interacting with the environment exhibited torsional and counter-balancing movements. Although not explicitly enforced, they emerged from the mechanical intelligence of the manipulator. The results demonstrate the potential of soft robotic manipulation via reinforcement learning.