Waymo Develops New Benchmark Model to Compare Autonomous Driving Safety Against Human Drivers

Waymo, in collaboration with Delft University of Technology, has developed a computer-based cognitive model called the Reference Driver (ReD) that simulates how a careful and competent human driver responds to traffic conflicts. Published Wednesday in Nature Communications, the model is built on the neuroscientific principle of active inference — the idea that humans constantly minimize surprise by imagining possible futures and adjusting behavior accordingly. Unlike previous industry models that focused only on last-second reactive maneuvers, ReD can simulate the full run-up to a potential crash, including proactive risk anticipation. It incorporates specific human cognitive traits: looming-based threat detection, a traffic norm filter that assumes rule-following until a violation is observed, and a 0.2-second foot-switching delay between gas and brake. Waymo says the model can be applied to large test sets with thousands of scenarios and is better suited to evaluating complex, real-world crash situations at scale. The company is making the research code publicly available under a non-commercial academic license and is working with regulators and standards bodies like the SAE to build consensus around the model as a shared safety benchmark. The release comes as Waymo faces heightened regulatory scrutiny, including ongoing NHTSA and NTSB investigations into a January incident in which a Waymo robotaxi struck a child in Santa Monica.