Perception-based learning for motion in contact in task planning

This paper presents a new approach to error detection during motion in contact under uncertainty for robotic manufacturing tasks. In this approach, artificial neural networks are used for perception-based learning. The six force-and-torque signals from the wrist sensor of a robot arm are fed into the network. A self-organizing map is what learns the different contact states in an unsupervised way. The method is intended to work properly in complex real-world manufacturing environments, for which existent approaches based on geometric analytical models may not be feasible, or may be too difficult. It is used for different tasks involving motion in contact, particularly the peg-in-hole insertion task, and complex insertion or extraction operations in a flexible manufacturing system. Several real examples for these cases are presented.