Closed‐Loop Soft Robot Control Frameworks with Coordinated Policies Based on Reinforcement Learning and Proprioceptive Self‐Sensing

Abstract Recent advances in soft robots have been achieved by using compliant materials and exploiting the advantages of the soft structural designs of living organisms. Living organisms (which have theoretically infinite degrees of freedom) are not only mechanically soft but are also capable of smooth harmonic motions, thanks to global coordination and the individual sensing and control of local tissues. Despite improvements in structural designs, few soft robot control frameworks for global object‐oriented behaviors are reported. Such a framework will require the use of multiple segments, with local sensing and independent control using coordinated policies. Here, a class of reinforcement learning based control frameworks for soft robots (with high degrees of freedom) is presented, and their ability to conduct global tasks is demonstrated. Coordinated control policies are formulated to control multiple segments with independently controllable embedded actuators, based on localized proprioceptive self‐sensing capabilities. The control frameworks are employed to develop soft physical robots. Demonstrations and experiments include the forward and backward locomotion of multichannel soft robotic flatworms. This approach is applicable to multifunctional, high degrees of freedom soft robots, as demonstrated by experiments with light‐sensitive locomotion.