Real-time Information Dissemination for Efficiency in a Robo-taxi System (RIDERS)

Real-time traffic information is critical for efficient robo-taxi fleet management, yet such data is often costly to obtain and incomplete in mixed traffic with low autonomous vehicle (AV) penetration. This project addresses this challenge by proposing a hierarchical reinforcement learning (HRL) framework that enables fleets to act not only as mobility providers but also as mobile sensors. Using New York City taxi demand data on the Manhattan road network, we simulate fleets that begin with partial knowledge of link-level speeds. Two actor–critic agents are employed: a zone-level agent reallocates idle vehicles across demand zones, while a route-level agent selects paths that balance service efficiency with information gain. Numerical experiments compare an information-focused strategy (AVR-IF) against a baseline (AVR-BA) under weekday and weekend demand with varying fleet sizes. Results show that AVR-IF uncovers vastly more link-level traffic information (~75,000 observations vs. ~3,000 for the baseline) while maintaining similar passenger wait times and vehicle miles traveled. These findings demonstrate that fleets can generate valuable real-time network intelligence without increasing operational costs or user burden. Beyond improving adaptability and responsiveness, such dual-purpose operation reduces reliance on third-party data and creates opportunities for collaboration between operators and public agencies. Ultimately, integrating information collection into everyday operations can make robo-taxi fleets more effective and impactful than traditional taxi or ride-sourcing services, advancing both mobility efficiency and urban traffic management.