Online Learning of Part Deformation Models in Robotic Cleaning of Compliant Objects

Abstract

We present an approach to automatically learn a bimanual robotic cleaning task on compliant objects. One robot grasps the object, while the other robot cleans it. Given a part with unknown deformation characteristics, the system visually detects the regions to be cleaned, and generates plans for both the grasping and cleaning arms. As the system performs cleaning attempts and gains experience with multiple new parts, it learns models of the part deformation depending on the cleaning force and grasping parameters. A planner iteratively generates tool paths for both robots using the available knowledge to optimize the cleaning time, including (1) delays from regrasping a part to minimize deflection and (2) time taken for repeated cleaning attempts over regions that remained dirty. A nonparametric deflection model is learned separately for each part, with minimal assumptions of the material behavior. We demonstrate the approach on a system of two KUKA LWR iiwa robots and a set of thin planar parts. Results indicate that the system is effective at rapidly learning part deformation models to enable effective iterative cleaning performance.