Cooperative autonomous vehicle control with deep reinforcement learning in lane-free roundabouts and its adaptability to various geometries

This study investigates deep reinforcement learning (DRL) mechanisms for achieving cooperative driving control of connected autonomous vehicles (CAVs) on lane-free roads. This study investigates the effectiveness of each of the two methods proposed in our conference paper to verify whether the best performance is achieved when both methods are incorporated. This study also evaluates the generalization performance of training models by conducting driving tests on several test courses to see if the training models can be adapted to courses other than the training course. In addition, this study proposed improvement in the use of course direction information to enhance the generalization performance of the training models. The result shows the proposed improvement can increase the generalizability of the trained model and the efficiency of vehicle flow.