Waymo publishes human collision-avoidance benchmark in Nature

Waymo on June 10 published peer-reviewed research in Nature Communications, developed with Delft University of Technology, introducing a Reference Driver (ReD) model that simulates how a careful, competent human anticipates and avoids crashes rather than only reacting at the last second. Built on the neuroscience concept of active inference, ReD models a driver's evolving beliefs and 'surprise' during a developing conflict, giving Waymo a more human-like benchmark to measure its Driver's collision-avoidance performance at scale. Waymo said it will release ReD as open-source software under a non-commercial academic license, a move that arrives as the company faces intensifying federal and media scrutiny over real-world incidents.