Online Gain Tuning Using Neural Networks: A Comparative Study

Abstract

This paper addresses the problem of adapting a control system to unseen conditions, specifically to the problem of trajectory tracking in off-road conditions. Three different approaches are considered and compared for this comparative study: The first approach is a classical reinforcement learning method to define the steering control of the system. The second strategy uses an end-to-end reinforcement learning method, allowing for the training of a policy for the steering of the robot. The third strategy uses a hybrid gain tuning method, allowing for the adaptation of the settling distance with respect to the robot’s capabilities according to the perception, in order to optimize the robot’s behavior with respect to an objective function. The three methods are described and compared to the results obtained using constant parameters in order to identify their respective strengths and weaknesses. They have been implemented and tested in real conditions on an off-road mobile robot with variable terrain and trajectories. The hybrid method allowing for an overall reduction of 53.2% when compared with a predictive control law. A thorough analysis of the methods are then performed, and further insights are obtained in the context of gain tuning for steering controllers in dynamic environments. The performance and transferability of these methods are demonstrated, as well as their robustness to changes in the terrain properties. As a result, tracking errors are reduced while preserving the stability and the explainability of the control architecture.