Acquiring state from control dynamics to learn grasping policies for robot hands

Abstract

A prominent emerging theory of sensorimotor development in biological systems proposes that control knowledge is encoded in the dynamics of physical interaction with the world. From this perspective, the musculoskeletal system is coupled through sensor feedback and neurological structure to a non-stationary world. Control is derived by reinforcing and learning to predict constructive patterns of interaction. We have adopted the traditions of dynamic pattern theory in which behavior is an artifact of coupled dynamical systems with a number of controllable degrees of freedom. For grasping and manipulation, we propose a closed-loop control process that is parametric in the number and identity of contact resources. We have shown previously that this controller generates a necessary condition for force closure grasps. In this paper, we will show how control decisions can be made by estimating patterns of membership in a family of prototypical dynamic models. A grasp controller can thus be tuned on-line to optimize performance over a variety of object geometries. This same process can be used to match the dynamics to several previously acquired haptic categories. We will illustrate how a grasping policy can be acquired that is incrementally optimal for several objects using our Salisbury hand with tactile sensor feedback.