Robust Experimental Study of Data-driven Optimal Control for an Underactuated Rotary Flexible Joint

Abstract

As an important component of industrial robot, the motion control of rotary flexible joint (RFJ) system is of great significance, especially when the system has unmodeled dynamics or is seriously disturbed. This paper presents an experimental robustness study on a kind data-driven optimal control approach based on an underactuated rotary flexible joint system. The data-driven approach combines the off-policy optimal control algorithm and the popular integral reinforcement learning technique. Through literature review, we find that the key step of the control design lies in that it learns the optimal value function and control policy simultaneously from the input and output (I/O) data. However, the I/O data are often disturbed by the system uncertainty or environmental noise, and then it will indirectly affect the optimal control performance. To investigate the robustness of the data-driven optimal control approach, we artificially set different experimental scenarios and take numerous control experiments on a RFJ experimental setup. The experimental results show that the data-driven optimal control method is quite robust against the system uncertainties in terms of maintaining the stability and delivering satisfactory tracking performance, even when the uncertainty is not a small quantity. In addition, the disturbance originating from environmental noise has certain impact on the controlling of RFJ system, but as long as the noise power is not too large, the control algorithm can converge to a satisfactory result. Finally, we find that the probing signal up has strong influence to this control algorithm, which reminds us to be cautious when selecting the probing signal.