Reinforcement Learning Based Route And Stop Planning For Autonomous Vehicle Shuttle Service

AV shuttles can be a complement to the existing bus services to fill the gap between the mobility demand of a city population and the supply of the existing bus services. In contrast to traditional approaches, this problem requires the consideration of new factors due to new objectives of the AV shuttle service such as mobility service for elderly and physically disabled persons, first/last mile transits, lessening of parking woes of the commuters, small local shop connectivity, and etc. In this work, we propose methods for route and stop planning for AV shuttle service in an urban city. We utilize a large scale human mobility dataset collected from users' cellphones' GPS sensors in Richmond city (as a city example) to identify the different potential transportation demands of AV shuttles. We propose an optimization problem-based solution and a deep reinforcement learning (RL) based solution. We conduct comprehensive evaluation based on our human mobility dataset along with real-world road network information. Our evaluation shows the proposed RL based system outperforms the comparison methods in terms of addressing the transportation needs as mentioned above by a significant margin while also minimizing the travel time on roads.