Reinforcement learning-based autonomous driving control for efficient road utilization in lane-less environments

Abstract

In recent years, research on autonomous driving using reinforcement learning has been attracting attention. Much of the current research focuses on simply replacing human driving with autonomous driving. Compared to conventional human-driven vehicles, autonomous vehicles can utilize a wide variety of sensor measurements and share information with nearby vehicles through vehicle-to-vehicle communication for driving control. By actively utilizing these capabilities, we can consider overall optimal control through coordination of groups of autonomous vehicles, which is completely different from human driving control. One example is adaptive vehicle control in an environment that does not assume lane separation or directional separation (Single Carriageway Environment). In this study, we construct a simulation environment and focus on the efficient use of a Single Carriageway Environment, aiming to develop driving control strategies using reinforcement learning. In an environment with a road width equivalent to four lanes, without lane or directional separation, we acquire adaptive vehicle control through reinforcement learning using information obtained from sensors and vehicle-to-vehicle communication. To verify the effectiveness of the proposed method, we construct two types of environments: a Single Carriageway Environment and a conventional road environment with directional separation (Dual Carriageway Environment). We evaluate road utilization effectiveness by measuring the number of vehicles passing through and the average number of vehicles present on the road. The result of the evaluation shows that, in the Single Carriageway Environment, our method has adapted to road congestion and was seen to effectively utilize the available road space. Furthermore, both the number of vehicles passing through and the average number of vehicles present have also improved.