Rethinking Multi-Modal Tokenization for Reinforcement Learning with Transformers in Mobility-on-Demand Tasks

Deep Reinforcement Learning (Deep RL) has been applied to Autonomous Mobility on Demand (AMoD) systems to optimize vehicle assignment, fleet distribution, and passenger wait times. Recently, transformer-based models, such as the Decision Transformer and Trajectory Transformer, have emerged as promising alternatives by framing policy learning as sequence modeling, improving sample efficiency and capturing long-range dependencies. These models, however, face challenges when handling the high-dimensional, multi-modal state spaces typical of AMoD environments. In this work, we propose researching new tokenization and embedding techniques that extend transformer architectures for these complex tasks by expanding input modalities and improving token representations. Our approach will initially be evaluated on the vehicle assignment task, with the goal of future use in implementing micro-transit solutions to address the last-mile problem in urban transportation.