Why smarter predictions are the key to safe autonomous vehicles
Manuel Mu帽oz S谩nchez defended his PhD thesis at the Department of Mechanical Engineering on September 17th.
Current approaches focus heavily on accuracy, often at the expense of robustness and real-world usefulness. The work shows that to build truly trustworthy autonomous systems, we must go beyond accuracy and evaluate how predictions actually affect driving behavior in realistic situations.
The trouble with today鈥檚 prediction models
Most existing research zooms in on precision: making predictions as close to the actual movements of road users as possible. While that sounds like the right target, it hides some serious flaws. Real-world driving is messy鈥攕ensors can fail, data can be incomplete, and unexpected behavior is common. When faced with these imperfections, many models break down dramatically.
The research of demonstrates that even small disruptions in the data can cause significant drops in performance. And contrary to what many believe, simply training models harder or feeding them more data isn鈥檛 enough to fix the issue.
Rethinking how we measure success
A second weakness lies in how prediction models are judged. Standard evaluation metrics鈥攖hose neat numerical scores researchers often quote鈥攄on鈥檛 always tell the full story. In fact, they can be misleading.
Some widely used metrics fail to show whether predictions actually help the car make better driving decisions. The study proposes new, more meaningful measures that reflect what really matters on the road: comfort, safety, and efficiency. This shift in evaluation makes it possible to identify which prediction methods are genuinely useful for autonomous driving.
From models to real driving behavior
The work goes one step further by exploring how trajectory predictions feed into the driving behavior of the vehicle itself. After all, a prediction that looks great on paper might not translate into safe or comfortable maneuvers on the street.
By studying how different models influence decision-making, the research identifies the kinds of predictions that are most valuable in practice. It also introduces a method to define the safe driving space around a vehicle鈥攖ested both in simulations and on real cars鈥攈elping to ensure the vehicle always has a reliable 鈥渟afety cushion鈥� to operate within.
Building trustworthy autonomous vehicles
The key takeaway is that current prediction models often lack robustness and fail when conditions are uncertain. Standard evaluation metrics can be misleading, creating a false sense of performance, while more informative alternatives provide a clearer view of what truly matters for safe and efficient driving. By addressing these gaps, this research brings trajectory prediction closer to the demands of everyday traffic. The result is not just smarter algorithms, but safer, more comfortable, and ultimately more trustworthy autonomous vehicles.
Funding: this work was supported by project SAFE-UP under EU鈥檚 Horizon 2020 research and innovation program, grant agreement 861570.
Title of PhD thesis: . Promotores: Prof. Ren茅 van de Molengraft and Dr. Emilia Silvas. Copromotores: Dr.ir. Jos Elfring.