{"title": "Using Backpropagation with Temporal Windows to Learn the Dynamics of the CMU Direct-Drive Arm II", "book": "Advances in Neural Information Processing Systems", "page_first": 356, "page_last": 363, "abstract": null, "full_text": "356 \n\nUSING BACKPROPAGATION \nWITH TEMPORAL WINDOWS \nTO LEARN THE DYNAMICS \n\nOF THE CMU DIRECT-DRIVE ARM II \n\nK. Y. Goldberg and B. A. Pearlmutter \n\nSchool of Computer Science \nCarnegie Mellon University \n\nPittsburgh, PA 15213 \n\nABSTRACT \n\nComputing the inverse dynamics of a robot ann is an active area of research \nin the control literature. We hope to learn the inverse dynamics by training \na neural network on the measured response of a physical ann. The input to \nthe network is a temporal window of measured positions; output is a vector \nof torques. We train the network on data measured from the first two joints \nof the CMU Direct-Drive Arm II as it moves through a randomly-generated \nsample of \"pick-and-place\" trajectories. We then test generalization with \na new trajectory and compare its output with the torque measured at the \nphysical arm. The network is shown to generalize with a root mean square \nerror/standard deviation (RMSS) of 0.10. We interpreted the weights of the \nnetwork in tenns of the velocity and acceleration filters used in conventional \ncontrol theory. \n\nINTRODUCTION \n\nDynamics is the study of forces. The dynamic response of a robot arm relates joint \ntorques to the position, velocity, and acceleration of its links. In order to control an ann \nat high spee