Grasp Approach Under Positional Uncertainty Using Compliant Tactile Sensing Modules and Reinforcement Learning

Abstract

Object grasping is a complex task that requires high environmental awareness. While vision generally provides highly detailed environmental information, light changes, object transparency, camera resolution, and other factors such as occlusion and clutter affect its perception of object pose. Due to these limitations, there may be some deviation between the estimated and actual object pose in unstructured environments. The use of compliant tactile sensors relaxes the requirement of strict finger position planning while providing essential information regarding contact with the target object. Therefore, under positional uncertainty, the robotic system may use compliant tactile sensors to perform multiple attempts before a successful grasp. In the present paper, we investigate using reinforcement learning and compliant tactile sensors to provide adaptive grasping under pose uncertainty. Here, we identify a policy that models an object position estimation error while minimizing the exploratory sensor contact before obtaining a grasp. Our method was able to perform a successful grasp while reducing the number of attempts from an average of five to an average of two per episode.