Active Sensing Strategies for Robotic Platforms, with an Application in Vision-Based Gripping

We present a vision-based robotic system which uses a combination of several active sensing strategies to grip a free-standing small target object with an initially unknown position and orientation. The object position is determined and maintained with a probabilistic visual tracking system. The cameras on the robot contain a motorized zoom lens, allowing the focal lengths of the cameras to be adjusted during the approach. Our system uses an entropy-based approach to find the optimal zoom levels for reducing the uncertainty in the position estimation in real-time. The object can only be gripped efficiently from a few distinct directions, requiring the robot to first determine the pose of the object in a classification step, and then decide on the correct angle of approach in a grip planning step. The optimal angle is trained and selected using reinforcement learning, requiring no user-supplied knowledge about the object. The system is evaluated by comparing the experimental results to ground-truth information.