Modeling and Planning for Connected, Cooperative, and Automated Mobility Systems

Networks of connected, cooperative, and automated vehicles will operate future mobility systems. Such mobility systems will reshape our current transportation systems to be safer, more efficient, and more affordable. They will help us eliminate fatal collisions, improve traffic throughput and energy efficiency, and make traveling cheaper with a shared-trip service. Connectivity and automation technologies are not yet reaching their full potential. Nowadays, many vehicles are connected, but mainly for infotainment, navigation, and maintenance purposes. In the meantime, they are also increasingly automated by using measurements from onboard sensors. However, these connected-automated vehicles (CAVs) are not yet cooperative because their actions are not coordinated based on shared information among each other. Apart from cooperation among each other, CAVs of a cooperative, connected, and automated mobility system must also interact with human traffic participants efficiently and safely. The main challenge is how to anticipate human factors, such as human routing and driving behaviors. Most existing approaches to vehicle coordination overlook or over-simplify the behaviors of other human traffic participants. This dissertation studies the modeling and planning of cooperative, connected, and automated mobility systems. Modeling refers to the generation of models with predictive abilities for intelligent agents to understand the world where they are situated. At the same time, planning refers to optimal policies for intelligent agents through coordination and communication to improve traffic efficiency, safety, and mobility. Three applications are studied throughout the dissertation, covering the essential tasks of connected, cooperative, and automated mobility systems. Firstly, a ride-sharing problem was solved among a shared CAV fleet to determine an optimal dispatching strategy. We proposed a decentralized ride-sharing algorithm using multi-agent reinforcement learning (RL). To better represent vehicles and requests with changing quantities and attributes, we formulated them as graphs and implemented graph neural networks to learn their relational information. We adopted a hierarchical framework to unify pick-up, drop-off, and rebalance tasks and distributed decision-making and computation among vehicles and servers. This dissertation shows that the RL-based decentralized method can achieve dispatching performance on par with the centralized optimization-based method and better computation efficiency. Next, the CAVs inside a bounded urban area determine their routes given origins and destinations. We studied the energy-efficient routing of plug-in hybrid vehicles. We developed an energy consumption model that could capture plug-in hybrid vehicles' distinct energy consumption patterns. An energy-efficient routing algorithm was proposed to minimize energy consumption. The proposed algorithm included travel time constraints to consider the balance between saving travel time and energy. It can simultaneously find the optimal path and powertrain switching strategy. We also studied the cooperation among vehicles, proposed a cooperative routing algorithm based on decentralized Monte Carlo tree search, and showed that the proposed cooperative routing algorithm could save travel time and energy. Finally, the CAVs predict the motion of other human-driven vehicles and plan their future trajectories accordingly in a roundabout scenario. We proposed a multi-modal trajectory prediction algorithm that could separately consider the uncertainty from routing intentions and maneuvers using graph transformer networks. With prediction in Frenet space, the proposed algorithm can predict distinct future trajectories. We implemented a deterministic sampling-based trajectory planning algorithm for planning and integrated it with the proposed prediction algorithm. This dissertation shows that the proposed prediction and planning integration could achieve a safer and smoother interaction with other CAVs and human-driven vehicles.PhDMechanical EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/176499/1/boqili_1.pd