Robotic pick-and-place of partially visible and novel objects

• Online Access: http://hdl.handle.net/2047/D20412868

If robots are to be capable of performing tasks in uncontrolled, natural environments, they must be able to handle objects they have never seen before, i.e., novel objects. We study the problem of grasping a partially visible, novel object and placing it in a desired way, e.g., placing a bottle upright onto a coaster. There are two main approaches to this problem: policy learning, where a direct mapping from observations to actions is learned, and modular systems, where a perceptual module predicts the objects' geometry and a planning module calculates a sequence of grasps and places valid for the perceived geometry. We have two contributions. The first relates to policy learning. We develop efficient mechanisms for sampling six degree-of-freedom gripper poses. Efficient sampling enables the use of established value-based reinforcement learning algorithms for pick-and-place of novel objects. Our second contribution relates to modular systems. We show that perceptual uncertainty is relevant to regrasping performance, and we compare different ways of incorporating perceptual uncertainty into the regrasp planning cost. Overall, we increase the range of objects robots can pick-and-place reliably without human intervention. This gets us a step closer to robots that work outside of factories and laboratories, i.e., in uncontrolled environments.--Author's abstract