Simulation Study Using NYC Taxi Data Found Robo-Taxi Fleet-as-Sensor Generated Approximately 75,000 Link-Level Traffic Observations Versus 2,000–4,000 for the Baseline, While Maintaining Similar Wait Times and VMT.
Study Developed Framework that Allows for Fleets to Double as Mobile Sensors to Obtain Less Costly and More Complete Real-time Traffic Information.
New York City, New York, United States
RIDERS: Real-Time Information Dissemination for Efficiency in a Robo-Taxi System
Summary Information
Real-time traffic data are critical for efficient robo-taxi fleet management, but such data are often costly and incomplete with low automated vehicle (AV) penetration. This study examined the feasibility of a solution in which idle robo-taxis were themselves used to generate and update the traffic data used in fleet decision making.
METHODOLOGY
The study modeled a centralized robo-taxi fleet management system where operations focused on passenger vehicle assignment and vehicle repositioning, using data from New York City (NYC)’s taxi demand data in Manhattan from April of 2016. The study adopted an artificial intelligent (AI) method using hierarchical reinforcement learning (HRL) which captured both fleet balancing decisions, which reallocated idle vehicles to better match demand, and routing choices, determine the specific routes taken by the vehicles. The study compared a baseline model (AVR-BA) and the information-focused model (AVR-IF) with fleets of 60, 120, and 240 vehicles and evaluated under both weekday and weekend demand patterns from NYC taxi data. The simulations used a simplified road network in which traffic conditions were represented by hourly link speeds. Vehicles began each simulation with incomplete traffic information, which was gradually updated as they traveled through the network and observed actual link speeds.
FINDINGS
- The AVR-IF uncovered 75,000 link level observations, compared to the baseline with 2,000-4,000 links uncovered, about 24 times more when using 3,000 as the average baseline value.
- The AVR-IF did not harm service quality. On weekdays, average daily passenger wait times dropped to 23 minutes with 120 or more vehicles, compared to the baseline of 27 minutes with 60 vehicles.
- Using the fleet as sensors did not inflate their overall distance travelled.
- Fleets could act as both mobility providers and sensing systems, which extends the reach of real-time traffic intelligence at no extra cost to operators, and reduces dependence on expensive third-party data sources and enables more efficient fleet deployment. In addition, passengers can benefit from more reliable service and consistent wait times.
- At the system level, fleets functioning as mobile sensors can expand traffic monitoring coverage without generating additional mileage, supporting smoother traffic flow and more resilient transportation networks.
