Learning Parametric Inverse Dynamics Models from multiple conditions for fast adaptive computed torque control

In this paper, we propose a novel approach for learning an inverse dynamics model of a serial-link robot only from data to be suitable for achieving computed torque control under unknown conditions, i.e., adaptive computed torque control. In our approach, we first collect a varied set of data from the robot under multiple conditions, each of which is constructed by putting loads on the body of the robot or giving the robot tools and bags to imitate real environmental situations. Then, a subspace representation that contains the various inverse dynamics models is extracted from the data as the Parametric Inverse Dynamics Model (PIDM) which is composed of the basis functions and weight coefficients. Using the PIDM, fast adaptive computed torque control under an unknown condition can be efficiently performed by automatically adjusting the weight coefficients of the basis functions unlike solving a high-dimensional learning problem as previous studies. To validate our approach, we applied the proposed method for the problem of adaptive computed torque control on trajectory-tracking tasks with an 7 DoF anthropomorphic manipulator and demonstrated the effectiveness. As a result, our approach achieved fast adaptation of computed torque control for the manipulator (within two seconds) even under unknown conditions such as holding and mounting unknown objects.