Chalmers study: new method plans driverless electric taxi fleets

Every morning, everyone wants to travel into the city center. Every afternoon, the flow reverses. For a self-driving taxi fleet, that means empty cars driving to wherever the next passengers wait, and electric cars must also stop to recharge. Researchers at Chalmers University of Technology in Sweden have now developed an optimization method built for this problem at the scale of a whole city, one that keeps working when traffic does not behave as forecast.

The stakes are real. According to a report cited by the researchers, large-scale shared, self-driving electric passenger transport can reduce an individual's cost per trip by up to 60% compared with traditional taxis, while also cutting emissions per trip significantly. The gain depends on the fleet being managed well. Without efficient control, the researchers note, the mismatch between available taxis and demand can increase waiting times and queue lengths, and worsen urban congestion. As lead author Sten Elling Tingstad Jacobsen puts it, "where and how long to charge is a difficult question."

The method combines two tools. Bayesian neural networks, a machine learning technique, forecast travel demand and charger availability in each part of the city, using historical taxi trip data from San Francisco and Chicago. Unlike many machine learning models, they also estimate how uncertain their own predictions are, which helps when real travel patterns differ from forecasts, for example during unusual weather. A transparent optimization model, with clearly defined goals and constraints, then decides where vehicles go and when they charge. The AI supports those decisions rather than making them on its own. To keep the calculation tractable, the problem is split into smaller parts that are solved separately and then combined. Jacobsen says the result is "very close to optimum while remaining computationally practical."

What distinguishes the work, according to Jacobsen, is that it explicitly accounts for the uncertainties of a large city and optimizes fleet operations accordingly, and that the framework was developed and tested for different types of vehicles. The study compares three electric vehicle types with different battery sizes and energy efficiency, and examines how those differences affect fleet performance and service availability. The team also investigates deep reinforcement learning, a form of AI that learns through interaction, as a way to make decisions in real time. The announcement does not give a figure for the method's gains over other planning approaches, and it does not say when it could be used outside the study.

For fleet operators and city planners, the practical gain would be fewer empty trips and fewer cars stuck waiting, if the method holds up beyond historical data. Autonomous taxi services already run in several cities in the United States and China, where managing large fleets in complex traffic remains difficult. Balázs Kulcsár, the professor who supervised the study, says the aim is more reliable and trustworthy autonomous mobility. Whether that reaches everyday service depends on tests the research does not yet report.