A Practical Method for Acquiring Inverse Dynamics Model and its Application to Mechanical Impedance Control of Human-Cooperative Robot

Abstract

This paper proposes a practical method, which is appropriate to actual use, to improve the feasibility of an external force/torque estimator proposed by the authors based on an inverse dynamics model using a recurrent neural network (RNN) and reports on evaluation results of its feasibility through experiments. Specifically, a method for making RNN learn the inverse dynamics more practically is proposed and the effectiveness is analyzed. In order to acquire the inverse dynamics model, the RNN is trained to learn the relationship between force/torque and motion when a robot is manipulating an object according to random trajectories based on band-limited M-sequences. The validity is evaluated by experiments using trajectories unknown to the trained RNN. Namely, experiment is conducted in which a robot is manipulating a rigid body with 6 degrees of freedom. The result showed that the acquired model by learning is effective even for the unknown trajectories. Furthermore, it has been confirmed by experiments that a mechanical impedance control for human-cooperative robots can be realized with the proposed external force estimator even while the robot is manipulating an object in trajectories unknown to the trained RNN.