Geurts’s research is part of the AMADeuS project, funded by the Dutch Research Council (NWO), in which three PhD candidates collaborate with TNO, Ford, and Rijkswaterstaat. The project aims to develop self-driving cars that can also interact effectively with non-autonomous vehicles.

After all, the transition to fully autonomous driving won’t happen overnight. There will first be a period of mixed traffic, where human-driven and autonomous vehicles share the road — a situation that poses major challenges. While self-driving cars can communicate with each other, the behavior of human drivers is far less predictable.

Predicting behavior

“In the United States, there are already taxis driving around without a driver,” Geurts explains. “Because autonomous vehicles can’t communicate with regular cars, you have to predict their behavior.” Many people assume the hardest part of autonomous driving is the technology itself, but the real challenge lies in the unpredictability of human behavior. “We use AI and mathematical models to predict how humans will act.”

In her research, Geurts didn’t focus on making those predictions herself but used them as input to determine how self-driving cars should behave on the road. “That part is called motion planning — it’s about the actual control of the car in traffic,” she says. The ultimate goal is to create traffic situations that are safe while maintaining as much speed and efficiency as possible. “We developed algorithms using a method called model predictive control, which allows us to mathematically prove that the situations we create are always safe.”

Worst-case scenarios

Her research focused specifically on merging, both on highways and in urban settings. “At first, you’re driving side by side in different lanes, but as soon as you start merging, you need to keep enough distance by adjusting your speed.” Using algorithms, her team calculated the safe distance for each phase of merging to ensure a smooth process.

Geurts began with simulations to test how the vehicles would behave and whether the algorithms worked. Later, she moved to real-life experiments on a test track. “Once you go to the test track, you’re suddenly in an uncertain environment with all kinds of external factors that can affect the results, like the weather,” she says. “Many of your assumptions no longer hold true. The big challenge is proving that, despite those uncertainties, you can still merge safely.”

But what if a driver suddenly does something unexpected? People often brake or accelerate abruptly when they see a car entering the lane. Can self-driving cars handle that? “In the algorithms I developed, we accounted for worst-case scenarios, so the situations are always safe — no matter what the other drivers do.”

On the test track

To demonstrate that the algorithms truly work in practice, tests were carried out at the Ford Proving Ground in Lommel, Belgium. There, a Ford car with a human driver and TNO’s self-driving test vehicle drove side by side. “We first had to implement my algorithms in the autonomous car,” Geurts explains. What sounds simple was anything but.

“You have to completely rebuild and rewrite your algorithm so it can run in the car,” she says. “It also has to connect to all the sensors and run in real time, without delays.” It’s crucial to ensure that everything functions properly so you can interpret the results correctly. “If the car reacts differently than expected, it could be the algorithm — but it could also be a faulty sensor. You want to rule that out.”

Moment suprême

The tests required a regular car with a certified test driver. “Luckily, one of my supervisors had that certification and joined me to drive the Ford car,” Geurts says. The self-driving vehicle also had a safety driver who could intervene if something went wrong, much like a driving instructor during a lesson. Geurts sat in the passenger seat, observing everything closely.

“I had prepared a detailed script with the scenarios I wanted to create to see if everything worked correctly. That way, I could ‘direct’ the test and give the driver precise instructions on when to accelerate or slow down.”

For Geurts, the first test runs were tense. “You finally see your algorithm — something you’ve worked on for months — come to life, reacting to another car instead of a simulation on your screen. You can check everything a hundred times, but at the moment suprême, it’s still a matter of waiting to see if it actually works. It was stressful, but also an amazing experience.”

Taking it to the road

Fortunately, everything went better than expected, and the tests confirmed that the algorithms work. Still, there’s room for improvement. “Sometimes we saw that the computation time was a bit too long — we’d like the car to react faster. Also, it can sometimes take longer for a certain signal to be received.”

At this stage, it’s still a prototype algorithm, Geurts emphasizes, but she remains optimistic. “It gives us a clear direction and brings us one step closer to actually taking it to the road.”