A Control Basis for Learning Multifingered Grasps

摘要

In this article, we introduce a robust controller that uses contact position and normal feedback to generate contact configurations for statically stable grasps. The approach uses a context sensitive composition of two controllers that minimize force and moment residuals in the grasp configuration. Equilibria in the composite controller correspond to optimal contact configurations for two and three contacts on regular, convex polygons. The preimage is used to generalize the controller to arbitrary object geometries by learning a policy for compensation and to address object recognition and contact (de)allocation.