Driverless electric taxi fleets learn to adapt to the unexpected

A new optimisation approach developed at Chalmers University of Technology seeks to tackle one of the less visible problems facing the robotaxi industry: how thousands of electric autonomous vehicles can be kept in the right place, sufficiently charged and ready for passengers when demand refuses to follow the forecast.

The technological challenge facing the robotaxi industry extends considerably further than teaching an individual car how to drive itself. For an autonomous taxi business to operate efficiently, an entire fleet needs to function as a coordinated system. Cars must be positioned where customers are likely to request them while avoiding unnecessary empty journeys. With electric vehicles, operators also need to decide when and where cars should recharge without taking too much passenger-carrying capacity off the road.

Researchers at Chalmers University of Technology in Sweden have developed an optimisation method intended to address these competing demands while recognising something that transport models cannot eliminate: uncertainty.

The research is particularly timely. Robotaxis are moving beyond experiments into commercial and pre-commercial fleets. In the US and China, private driverless taxi fleets more than doubled during 2025 to around 8,000 vehicles operating across more than two dozen major cities, according to International Energy Agency figures cited by Digital Journal. Europe is also seeing trials involving companies including Waymo, Wayve, Uber, Apollo Go, WeRide and Pony.ai.

The robotaxi positioning problem

The difficulty is that passenger demand is inherently uneven. A city might experience strong flows towards its centre in the morning before the direction reverses later in the day. If an autonomous taxi finishes a journey somewhere with little subsequent demand, there is no human driver making an intuitive decision about where to wait next. The fleet management system needs to determine where that vehicle would be more useful.

This process, generally termed rebalancing, can involve running vehicles without passengers. Get the decisions wrong and the purported efficiency of autonomous transport begins to deteriorate.

Sten Elling Tingstad Jacobsen, who holds a PhD in Control Engineering from Chalmers University of Technology, identifies the additional complication posed by electrification: vehicles periodically need to stop working in order to charge. Determining where they should recharge, and for how long, consequently becomes part of the fleet-wide optimisation problem.

His research combines machine-learning-based forecasts of passenger demand with optimisation techniques for planning fleet movements. It also investigates deep reinforcement learning for real-time decision-making in complex transport systems.  A peer reviewed paper considered where unoccupied vehicles should be positioned so they are better prepared for subsequent passenger requests.

Making room for uncertainty

The newer research introduces an important extra dimension: electric cars require access to charging infrastructure. Historical taxi journey data from San Francisco and Chicago were used to predict travel demand and charger availability across different parts of the cities. The researchers applied Bayesian neural networks, or BNNs, to the forecasting problem.

A significant feature of this approach is uncertainty estimation. Instead of treating a forecast as if it were necessarily correct, a BNN can also produce information about uncertainty surrounding its prediction. This matters because cities do not behave identically from one day to the next. Weather and other departures from normal travel patterns can make historical demand an imperfect guide to what happens next.  The distinction also provides an interesting model for practical AI deployment.

Machine learning supplies the forecasts, but it does not independently decide where the taxis will travel or recharge. Those operational choices are handled through an optimisation model with stated objectives and constraints. In other words, predictive AI provides an input into the decision-making process rather than becoming an opaque autonomous fleet controller. The researchers also evaluated three types of electric vehicle with differing battery capacities and energy efficiencies to examine how vehicle characteristics affect fleet performance and service availability.

Scale presents another difficulty. Attempting to calculate every possible movement, charging event and passenger requirement across a large fleet can rapidly make the optimisation problem computationally unwieldy. The researchers address this by dividing the overall problem into smaller components that can be solved efficiently and subsequently combined. Jacobsen describes the resulting answer as very close to the optimum while remaining computationally practical.

The methodology is detailed in the open-access paper Combined stochastic and robust optimization for electric autonomous mobility-on-demand with nested Benders decomposition, published in Transportation Research Part C: Emerging Technologies.  The title points to the central idea. Stochastic optimisation provides a way of dealing with outcomes characterised by uncertainty, while robust optimisation seeks solutions that continue to perform acceptably when conditions differ from expectations. The research applies these approaches to electric autonomous mobility-on-demand and uses nested Benders decomposition to make the large optimisation task manageable.

Why fleet intelligence matters

Such work highlights an important distinction in debates about autonomous transport. A technically capable self-driving car is not, by itself, an efficient autonomous transport network. A commercially useful robotaxi fleet also depends upon orchestration. Too many vehicles arriving in the same place can waste road capacity. Too few vehicles create queues and longer waits. Charging adds a further constraint because an electric taxi connected to a charger is temporarily unavailable to passengers.

Balázs Kulcsár, Professor in the Department of Electrical Engineering at Chalmers University of Technology and supervisor of the research, notes that self-driving fleets continue to face uncertain, complex traffic environments and interactions with human drivers. He sees robust optimisation as a way of supporting better traffic coordination and more dependable autonomous mobility. 

This becomes more significant as fleets grow. Waymo, for example, claimed around 3,000 driverless taxis across a dozen US cities by June 2026, while Apollo Go had a similar number deployed across 27 Chinese cities, according to figures reported by Digital Journal.

There is also a wider environmental calculation. Electrifying a vehicle removes tailpipe emissions, but autonomous driving introduces additional computing requirements. Previous research has raised questions over how computing energy consumption could affect the environmental balance as autonomous vehicles reach very large scales. The energy consumed by autonomous-vehicle computing therefore needs to form part of the sustainability discussion.