Learing Compliant Motions by Task-Demonstration in Virtual Environments

We are proposing a supervised learning approach in robot force control which enables robot programming by demonstration through an operator. The learning element of the controller is based on a neural network prestructured to represent knowledge in terms of rules. Tasks are demonstrated in a virtual environment. Visual feedback using sensor ball devices and graphic display of forces as well as visual and proprioceptive feedback using haptic interfaces are being considered. Experimental results are shown for the “Put Block in a Corner of a Box Problem”.