Learning a Generative Transition Model for Uncertainty-Aware Robotic\n Manipulation

Abstract

Robot learning of real-world manipulation tasks remains challenging and time\nconsuming, even though actions are often simplified by single-step manipulation\nprimitives. In order to compensate the removed time dependency, we additionally\nlearn an image-to-image transition model that is able to predict a next state\nincluding its uncertainty. We apply this approach to bin picking, the task of\nemptying a bin using grasping as well as pre-grasping manipulation as fast as\npossible. The transition model is trained with up to 42000 pairs of real-world\nimages before and after a manipulation action. Our approach enables two\nimportant skills: First, for applications with flange-mounted cameras, picks\nper hours (PPH) can be increased by around 15% by skipping image measurements.\nSecond, we use the model to plan action sequences ahead of time and optimize\ntime-dependent rewards, e.g. to minimize the number of actions required to\nempty the bin. We evaluate both improvements with real-robot experiments and\nachieve over 700 PPH in the YCB Box and Blocks Test.\n