Guiding Grasping with Proprioception and Markov Models

Abstract

This paper describes the application of a partially observed Markov decision process (POMDP) to guide the control decisions made during the task of grasping objects with a simple compliant grasper in unstructured environments. The decision process relies only on the sensing of angular deflection of the compliant gripper joints - proprioceptive information available on most robot hands and grippers. This information is used to infer the state of contact between the gripper and the object and guide a set of actions to be undertaken in order to lead to a successful grasp. We believe that the performance of the gripper under a POMDP model built from this limited sensory information will serve as a valuable baseline for comparison with more complex sensing modalities, allowing for quantitative analysis of the tradeoffs between commonly available sensory suites.