Improvement of a Learning-Based Tuning Approach for Trajectory Tracking Controller

Controller design and tuning are crucial to the driving performance of autonomous vehicles. Typically, engineers define a control structure in advance and then tune its parameters in a real-world driving scenario through trial-and-error method. This method is undoubtedly time and resource consuming, and dis-courages the exploration of different control structures. In this paper, we present an improved learning-based controller tuning approach and achieve the rapid tuning of a trajectory tracking controller for a real vehicle. The proposed approach employs the framework of model-based reinforcement learning, in which a probabilistic dynamics model is first learned from the real data and then the controller parameters are updated based on the learned model. To facilitate real-world applications, the aleatoric uncertainties of the real vehicle and the measurement system are distinguished and separated during the learning of probabilistic model. Additionally, the original state propagation framework is extended by incorporating steering response model and prepro-cessed Gaussian noise. Simulation and experimental results show that controllers can be well tuned within 3 trials by our approach and the comprehensive tracking performance (i.e., cumulative cost) can be improved by up to 76% compared with the initial controller parameters. More importantly, the tuned controllers maintain high tracking accuracy in untrained driving scenarios. We believe that this approach will help to accelerate the controller design and tuning of autonomous vehicles.