Dynamic control of a six degree-of-freedom robot manipulator using neural networks

Summary form only given. A dynamic controller for a full six-degree-of-freedom manipulator has been developed based on a backpropagation neural network. Unsupervised learning called feedback error learning is used to train the net. Although absolutely no dynamic model or its parameters were known (the robot is treated as a complete black box), it implicitly learns the robot´s dynamic properties through repetitive movement trials. Importantly, this black box model can automatically take care of some of the unmodeled effects such as friction and vibrations. Its control performance has been tested on a simulated PUMA 560, demonstrating fast learning and convergence. Furthermore, the neurocontroller exhibits adaptation to changing loads without load sensors, generalization over unlearned trajectories, and robustness against sensor noise