How autonomous cars learn to merge safely
Merlijne Geurts defended her PhD thesis at the Department of Mechanical Engineering on October 9th.
Fully autonomous vehicles won鈥檛 appear on our roads overnight. For many years, they will share the roads with traditional, manually driven cars 鈥� a situation known as mixed traffic. This creates uncertainty, as human drivers can behave unpredictably. One of the greatest challenges in this transition is ensuring that automated vehicles can plan safe and efficient trajectories 鈥� especially during lane changes and merging, where timing and cooperation are crucial.
Predicting human drivers: does it really help?
A major part of this research of Merlijne Geurts explored whether predicting the actions of human drivers could make autonomous vehicles safer and more comfortable. Using detailed simulations of various highway merging scenarios 鈥� from long ramps to traffic jams 鈥� the study showed that having access to predictions indeed improved safety. When the system could 鈥済uess鈥� what human drivers might do next, the automated vehicle was better able to avoid risky situations.
However, comfort did not improve. The reason? Safety sometimes requires sudden braking or slower acceleration 鈥� actions that make a ride less smooth. In short, what鈥檚 safest isn鈥檛 always what feels best.
Developing safe and smart algorithms
To tackle the problem, four new lane-merging algorithms were developed. These algorithms use predictions about other vehicles鈥� movements to decide when and how to merge safely. Each version of the algorithm accounted for increasing uncertainty in human behavior 鈥� yet all of them were guaranteed to be safe.
To validate these solutions, two algorithms were tested on a real-world test track in Lommel, Belgium, using full-sized vehicles from multiple brands. The results were promising: even in unpredictable conditions, the vehicles merged safely. These real-world experiments also revealed valuable insights for improving the algorithms further, paving the way for more advanced and reliable autonomous driving systems.
A step toward a safer, autonomous future
This research demonstrates that automated vehicles can handle one of the most complex tasks in driving 鈥� merging 鈥� safely and intelligently, even in mixed traffic. While comfort still needs improvement, the ability to predict and react to human behavior is a significant milestone.
By developing algorithms that guarantee safety and validating them in both simulations and real-world conditions, this PhD work brings us one step closer to a future where humans and self-driving cars can share the road 鈥� safely, confidently, and eventually, comfortably.
This research is part of the AMADeuS (Artificially Intelligent Multi-Vehicle Automated Driving Systems) program.
Title of PhD thesis: . Supervisors: Prof. Maurice Heemels, Dr. Emilia Silvas, and Dr.ir. Alexander Katriniok.