How thousands of self-driving electric taxis could be coordinated

Imagine ordering a ride on your phone and a self-driving electric car arrives within minutes. It picks you up, perhaps shares your ride with other passengers, and takes you to your destination – without a driver and without emissions. At the same time, thousands of similar vehicles move around the city, recharge their batteries when needed, and reposition themselves to be where the next passenger is likely to appear. Behind this seamless experience lies a complex puzzle. Researchers at Chalmers University of Technology have developed methods that could help thousands of vehicles plan their operations, charging and routes – even when the future is uncertain.

How can thousands of self-driving electric taxis be coordinated to pick up passengers, plan their routes and recharge at the right time? Researchers at Chalmers have developed methods that could make future electric taxi fleets more efficient and reliable, even as passenger demand and traffic conditions change. Photo: Illustrative image.

This is the vision of Electric Autonomous Mobility-on-Demand – shared, self-driving and electric mobility on a large scale. Studies point to the potential for travel costs to be reduced by up to 60 per cent, along with significantly lower emissions per journey compared with conventional taxis. But coordinating such a fleet is an extremely complex logistical challenge.

In the morning, for example, many people may want to travel towards the city centre, while later in the day travel patterns shift in the opposite direction. This creates an imbalance, requiring empty vehicles to be moved from one area to another. For an electric vehicle fleet, there is an additional challenge: the cars also need time to recharge, and charging points need to be available when they need them.

At the same time, it is impossible to know exactly what will happen during the day. Passenger demand is unpredictable, traffic conditions vary, and demand can be affected by factors such as weather and events. Every decision about where to send a vehicle now therefore affects the options available just a few minutes later.

Research at Chalmers University of Technology, conducted in collaboration with Volvo Cars, shows how such systems could be managed more intelligently despite this uncertainty.

“The common challenge in all our studies is the uncertainty about what will happen during the day – for example, travel patterns, traffic flows, travel times and energy consumption. Our work is about how to plan despite that uncertainty,” says Sten Elling, PhD student at the Department of Electrical Engineering at Chalmers.

Planning ahead – despite uncertainty

The first part of the research focuses on how empty vehicles should be positioned around a city so that they are ready when the next passenger needs them.

Instead of simply reacting when a booking comes in, the system uses forecasts of future demand – together with estimates of how uncertain those forecasts are. This allows the system to plan ahead and ensure that enough vehicles are available where they are needed. To predict travel behaviour, the researchers used historical data from real taxi journeys in San Francisco and Chicago.

The method can provide a concrete level of service: there is at least a 95 per cent probability that enough vehicles will be in the right place at the right time. For operators, this means a more reliable service – and for passengers, shorter waiting times.

Coordinating charging and driving

The second part focuses on a challenge specific to electric vehicles: charging.

A car at a charging station cannot pick up passengers at the same time. A vehicle travelling a long distance to meet demand consumes more energy and may need to recharge sooner than planned. Meanwhile, travel times, demand and energy consumption vary throughout the day.

The researchers have developed a method capable of handling several types of uncertainty simultaneously – including travel demand, travel times, energy consumption and the availability of charging stations.

The results also show that smart planning could influence how the vehicles themselves are designed. If charging can be planned efficiently, each vehicle may not need a very large battery.

“With the right control strategy, the cars can manage with smaller batteries because they can recharge during the day. A smaller battery makes the car lighter and more energy-efficient, which in turn reduces electricity consumption,” says Sten Elling.

Machine learning solves real-time matching

The third and largest part of the research focuses on what happens in real time when a passenger requests a ride.

Which vehicle should pick up the passenger? In what order should multiple passengers be picked up and dropped off? For large vehicle fleets, this is too complex to solve exactly every time.

Instead, Sten Elling has used a machine-learning-based method developed together with Attila Lischka, PhD student at Chalmers. The approach combines deep reinforcement learning, which learns decision-making through trial and error, with graph neural networks, which can represent and process the complex relationships between vehicles, passengers, and locations in the transport network. The system was trained on millions of simulated journeys and gradually learned which decisions lead to efficient solutions.

An important advantage of the method is its scalability. It can handle very large fleets while still making decisions quickly.

“Traditional optimisation methods can become very slow as systems grow larger. With machine learning, we can find good solutions in under a second,” says Sten Elling.

Self-driving taxi fleets already exist

Self-driving taxi services are no longer just a vision for the future. Autonomous taxi fleets are already operating in cities in countries including the United States and China. However, the technology still faces challenges in certain situations, for example in bad weather, changing traffic conditions and interactions with human drivers.

Balázs Kulcsár
Photographer: Anna-Lena Lundqvist

“The Chalmers method could contribute by improving traffic coordination and helping to create reliable and trustworthy autonomous mobility,” says Balázs Kulcsár, Professor at the Department of Electrical Engineering and supervisor of the study.

Possible next steps

Although Chalmers’ research is currently based on advanced simulations, the results could contribute to the future development of electric mobility systems.

One possible continuation of the research is to investigate how large fleets of electric vehicles could interact with the electricity grid through so-called vehicle-to-grid technology. Parked vehicles could then temporarily supply energy back to the grid.

– If hundreds of thousands of vehicles are available in a city, together they could provide significant power to the electricity grid while they are parked, says Sten Elling.

Challenges remain

At the same time, several questions remain to be addressed before such systems can be introduced on a large scale. For example, the research has not analysed how large fleets of autonomous vehicles themselves would affect traffic flows in a city. It is also unclear how different mobility services – such as taxi companies and ride-sharing platforms – would compete in a future system.

Despite these challenges, the research shows that combining advanced mathematics and artificial intelligence could make large electric vehicle fleets both more efficient and more reliable. In the long term, this could contribute to shorter waiting times, lower energy use and more sustainable transport in the cities of the future.