When a car ahead brakes hard or a vehicle swerves into the path, drivers must act within a second or two. Earlier descriptions usually covered only isolated parts of this chain, such as how long it takes to notice danger or how far the steering wheel turns. A new approach developed by researchers at Delft University of Technology links all stages into one system. It draws on active inference, a framework in which an agent continually updates its beliefs about the world in order to reduce surprise or uncertainty.
The system watches the visual expansion of nearby objects, known as looming, to judge closing speed and distance. It then forms predictions that assume other drivers will mostly obey normal traffic rules. When the situation grows more surprising than expected, the model gathers evidence until a threshold is crossed and then selects the action that best balances safety with reasonable effort.
How the model captures human behavior
The researchers ran computer simulations of three common emergencies: a lead vehicle braking suddenly, an oncoming car entering the lane, and a vehicle failing to yield at an intersection. They supplied the model with the same visual information available to human drivers in earlier simulator studies and meta-analyses of real crashes. Response times, choices between braking and steering, and the strength of braking all fell within the range observed in people. The model also built in typical human limits, such as delays in noticing distant motion and a preference for following expected road rules, which sometimes leads to hesitation in uncertain cases. These features kept the simulated behavior recognizably human rather than perfectly optimal.
The same framework can now serve as a consistent benchmark when developers and regulators assess whether autonomous or self-driving vehicles handle sudden threats at least as well as attentive human drivers. It also supplies measurable targets that manufacturers can use when designing collision-avoidance systems. By offering a shared, evidence-based reference, the work supports clearer safety standards across the industry.
This research is published in Nature Communications.