Learnability of the Moving Surface Profiles of a Soft Robotic Sorting Table

This paper analyzes the application of machine learning techniques to the control of a soft, peristaltic, xy-sorting table. In particular, we address peristaltic tables made of a soft upper silicone layer and actuated by an array of integrated air-filled chambers. The chambers are pneumatically inflated in order to deform the table and move objects on the table. To control the robot precisely, it is necessary to model both the inverse mapping between the control signals of the actuators and the resulting surface deformation. There is currently no parametric model available. In this paper, we, therefore, study if nonparametric approaches are applicable. In these approaches, the mapping would be learned from a database of input signals and observed behaviors. From our analysis, we conclude that the most promising research direction consists in the nonparametric modeling of a limited set of peristaltic actuation patterns. However, the nonlinear hardware design that impedes a parametric model also affects the nonparametric optimization process. Our simulation suggests that the optimization is nonconvex, approximately monotonous, and feasible in terms of the number of observations of the physical robot.