One of the key challenges in developing autonomous vehicles is planning safe and efficient trajectories in complex environments, such as intersections. This paper proposes an offline RL approach for planning trajectories for autonomous vehicles at crossroads with other actors. It enables the possibility of using pre-recorded expert trajectories for algorithm tuning. We study the influence of the quality of collected trajectories on various offline reinforcement learning methods. Our approach has the potential to overcome the limitations of online RL and provide an effective planning solution for autonomous vehicles in dynamic environments.
Melkumov, M., Panov, A. I. (2023). Planning Maneuvers for Autonomous Driving Based on Offline Reinforcement Learning: Comparative Study // In: Kovalev, S., Kotenko, I., Sukhanov, A. (eds) Proceedings of the Seventh International Scientific Conference “Intelligent Information Technologies for Industry” (IITI’23). IITI 2023. Lecture Notes in Networks and Systems, vol 776. Springer, Cham. Pp. 65–74.