Online multi-target learning of inverse dynamics models for computed-torque control of compliant manipulators
Inverse dynamics models are applied to a plethoraof robot control tasks such as computed-torque control, whichare essential for trajectory execution. The analytical derivationof such dynamics models for robotic manipulators can bechallenging and depends on their physical characteristics. Thispaper proposes a machine learning approach for modelinginverse dynamics and provides information about its implemen-tation on a physical robotic system. The proposed algorithmcan perform online multi-target learning, thus allowing effi-cient implementations on real robots. Our approach has beentested both offline, on datasets captured from three differentrobotic systems and online, on a physical system. The proposedalgorithm exhibits state-of-the-art performance in terms ofgeneralization ability and convergence. Furthermore, it hasbeen implemented within ROS for controlling a Baxter robot.Evaluation results show that its performance is comparable tothe built-in inverse dynamics model of the robot.