TU Delft research makes self-driving cars safer and more human

How can a self-driving car prevent accidents? And how can an autonomous vehicle behave “naturally” among human road users? Two new studies from TU Delft provide answers to these questions.

Advanced driver assistance systems and autonomous functions are rapidly finding their way into practice. This also increases the need for demonstrably safe and predictable technology. TU Delft researchers are working on the next generation of self-driving vehicles: systems that are safer and better understand how people behave in traffic, so they can take that behaviour into account in their decision-making models.

New model predicts how people avoid collisions

Scientists at TU Delft, together with Waymo, have developed a new model that predicts with high accuracy how human drivers respond to dangerous traffic situations. For the first time, different types of collision-avoidance behaviour are brought together in a single model. The results were published last week in Nature Communications. Waymo is already using the model to compare the performance of its autonomous vehicles with that of human drivers.

When a vehicle in front of you suddenly brakes, or an oncoming vehicle unexpectedly enters your lane, you have only fractions of a second to decide whether to brake, swerve, or do both. “Existing models usually describe only part of this process, such as reaction time or steering behaviour,” says Arkady Zgonnikov, assistant professor at TU Delft. “Our new model brings all these components together.” The model integrates perception, decision-making, and execution into a single coherent structure. As a result, it can recognise when a situation becomes dangerous, predict how traffic is likely to develop, and determine the most effective evasive strategy.

To test whether the model is accurate, the researchers compared it with human behaviour in three dangerous traffic situations: a vehicle ahead braking suddenly, an oncoming vehicle unexpectedly entering the lane, and a car failing to yield. The model received exactly the same information as human drivers. “The model showed realistic braking reaction times and made similar choices between braking and steering,” Zgonnikov says. In addition, the model takes human limitations into account, which keeps its behaviour recognisably human.

The researchers see important applications for both the development and assessment of autonomous vehicles. “It can help answer the question of whether autonomous vehicles are safer than human drivers, a key issue in regulation,” says Zgonnikov. “At the same time, it becomes possible to formulate clear and measurable requirements for manufacturers.” According to Mauricio Peña, Chief Safety Officer at Waymo, the model can help the sector “move towards a shared, scientifically grounded approach to assessing collision avoidance.”

A self-driving car must feel human

On June 10, Lucas Suryana received his PhD from TU Delft’s Automated Driving and Simulation Lab. In his research, he developed an additional layer on top of the control system of a self-driving car. That layer looks not only at traffic rules and safety, but also at human reasons and expectations. This allows a self-driving car to better understand why people behave as they do and to incorporate that knowledge into its decisions.

According to Suryana, this is essential for the future of autonomous driving. “A self-driving car that always stays within the lines, never cuts anyone off and follows all the rules perfectly will sometimes actually create uncomfortable situations. In traffic, people constantly take one another into account. Automated vehicles must learn to do the same.”

One example is a car that ends up behind a cyclist on a narrow road with a solid centre line. According to the traffic rules, the cyclist cannot be overtaken safely. A self-driving car that follows only the rules will remain behind the cyclist, while a growing traffic jam forms behind the car.

“It is precisely in situations like these that you see traffic participation is about more than just following rules,” Suryana says. “People understand why a situation becomes uncomfortable and adjust their behaviour accordingly. My research shows how automated vehicles can recognise and include such human considerations in their decision-making.”

The research builds on the concept of Meaningful Human Control. Under this concept, intelligent systems should not only function safely, but also take human values, intentions and interests into account.