Learning to pick up objects through active exploration

Robots need to perceive and manipulate objects in their environment, yet robust object manipulation remains a challenging problem. Many aspects of a perception and manipulation system need to be customized for a particular object and environment, such as where to grasp an object, what algorithm to use for segmentation, and at which height to visually servo above an object on the table. To address these limitations, we propose an approach for enabling a robot to learn about objects through active exploration and adapt its grasping model accordingly. We frame the problem of model adaptation as a bandit problem, specifically the identification of the best of the arms of an N-armed bandit, [5] where the robot aims to minimize simple regret after a finite exploration period [1]. Our robot can obtain a high-quality reward signal (although sometimes at a higher cost in time and sensing) by actively collecting additional information from the environment, and use this reward signal to adaptively identify grasp points that are likely to succeed. This paper provides an overview of our previous work [3] using this approach to actively infer grasp points and adds a description of our efforts learning the height at which to servo to an object.