Towards learning contact states during peg-in-hole assembly with a dual-arm robot

The use of dual arm robots for human like operations such as the peg-in-hole tasks can be facilitated by learning contact states during manipulation. In this paper, we propose a control scheme that includes the learning of contact states during operations. The approach includes the use of a Motoman SDA-20 dual-arm robot equipped with two three-fingered gripper and a force/torque sensing capability. A real operation using the assembly of an automotive starter motor is used as a case of study. Several patterns are generated and classified during the stages of the starter assembly. Three categories were considered during operations: moving both arms simultaneously and moving the right arm while keeping the left arm static and vice versa. Contact states were generated using the robot and learned by an Artificial Neural Network (ANN) Fuzzy ARTMAP architecture. The output of the ANN has been envisaged to be a motion command to diminish the constrained forces while moving the arm in the assembly direction. Results during contact state classification have shown that manipulative forces can be recognized by the Neural Network Controller (NNC) so that a valid motion command can be issued to the arm robot favoring the assembly task.