Learning to Control Soft Robots
Published:
Introduction
Classically controlled robots have revolutionised assembly lines where the environment is restrictedand predictable. However, this control scheme has proven less effective in uncertain, unstructured environments due to the unmodelled nonlinearities in the morphology and its interaction with the environment [Atkeson et al., 2015]. Embodied intelligence suggests that we are not controlled centrally but rather that the morphology and environment also contribute to behaviour. Therefore, the passive dynamics of the robot could be exploited to simplify the controller by making the passive dynamics closer to the desired behaviour [Pfeifer and Bongard, 2006]. Consequently, there has been a growing interest in soft robots which have many passive degrees of freedom and are often underactuated. However, classical control necessitates exact kinematic and dynamic models which are hard to derive analytically for soft robots due to the nonlinear dynamics of the body, external forces, and uncertain, unstructured environment. Model-free approaches are a potential solution to these problems as they can approximate inverse kinematic and dynamic models that account for the unknown nonlinearities.