Using a neural network to learn the dynamics of the CMU Direct-Drive Arm II

Computing the inverse dynamics of a robot arm is an active area of research in the control literature. We apply a backpropagation network to this problem and measure its performance on the first 2 joints of the CMU Direct-Drive Arm II for a family of "pick-and-place" trajectories. Trained on a random sample of actual trajectories, the network is shown to generalize with a root mean square error/standard deviation (RMSS) of 0.10. The resulting weights can be interpreted in terms of the velocity and acceleration filters used in conventional control theory. We also report preliminary results on learning a larger subset of state space for a simulated arm. This research was sponsored in part by National Science Foundation grants DMC-8520475, EET-8716324, and by the Office of Naval Research under contract number N00014-86-K-0678. Barak Pearlmutter is a Fannie and John Hertz Foundation fellow.