Improved Robustness of Reinforcement Learning Based on Uncertainty and Disturbance Estimator

摘要

This paper proposes a method to improve the robustness of RLs based on model-free uncertainty and disturbance estimator (RL-based UDE). In the real environment, instead of using optimal trajectory and control techniques to perform complex tasks, it learns through RL and supplements robustness by using uncertainty and disturbance estimator (UDE). From UDE, the robotics system can be improved the stability by appropriately canceling the uncertainty and disturbance without efforts to obtain model information; hence the UDE can compensate for the performance degradation of RL when system is non-stationary. In addition, the performance can be improved by reducing the sensor noise from low-pass filter of UDE. It is shown through an experiment that the proposed RL-based UDE provides robustness.