Architecture for testing learning-based autonomous vehicle control design

This paper presents the architecture for the testing of autonomous vehicle controllers designed using machine learning. A localization system feeds measurement data into a ground station. The ground station processes this data and provides the position of the vehicle to the controller, which has been tuned offline via machine learning. Control inputs are then generated and provided to the vehicle in order to accomplish the desired mission. The performance of the learned controller is compared to a nominal case. During the offline tuning process, a vehicle model is used to simulate the actual vehicle. Once tuning is complete, the parameters are used to control a real vehicle in order to accomplish the desired mission. The proposed architecture was tested by tuning a model predictive controller to guide a differential drive robot to a series of waypoints using a Reinforced Learning technique known as Learning Automata. The controller was then tested online with a similar but different problem on a real differential drive robot. The results showed that after tuning, the vehicle performed significantly better. This demonstrated that offline learning techniques can be used to optimally select controller parameters.