The peer reviewed ReD model, built with TU Delft, simulates split second human crash avoidance decisions. Waymo is offering it to competitors and regulators as a shared yardstick, even though that yardstick is, in part, its own.
Waymo built a digital stand-in for the average human driver and, on Wednesday, published it in Nature Communications. The model, called ReD for Reference Driver, is a cognitive simulation of how people make split-second crash-avoidance decisions, developed jointly with Delft University of Technology in the Netherlands. The company is now pitching it as something the entire self-driving industry should use to measure its own cars: a "behavioral crash test dummy" for collision-avoidance software.
That framing is the substance of the release. Physical crash-test dummies turned vehicle safety from a marketing claim into a measurable engineering target in the 1950s. Waymo's argument, laid out in a paper co-authored with researchers at Delft, is that AV collision avoidance has no equivalent common yardstick, and the company is offering one for free, according to The Verge's writeup of the announcement.
The mechanism is what makes ReD more than a marketing prop. The model layers several human decision traits: a "looming"-based judgment of how quickly a hazard is closing in, a traffic-norm filter that assumes other drivers will follow the rules until they don't, a surprise-driven plan reevaluation when conditions cross a threshold, and a 0.2-second pause that mimics the gap between moving a foot from the gas to the brake — a figure Waymo has cited in its announcement. In a surprise scenario, Waymo describes ReD as continuously calculating which future would least surprise it, in the formal language of "active inference," and steering toward that.
The model's intellectual scaffolding comes from active inference, a framework in theoretical neuroscience associated with Karl Friston, who called ReD a "technical tour de force" in a statement provided by Waymo. That endorsement is part of a coordinated release rather than independent reporting, so it carries Waymo's framing, not the journal's.
Waymo's pitch, in the words of chief safety officer Mauricio Peña, is that ReD defines a "competent human response" to surprise scenarios, and that any AV whose behavior diverges sharply from ReD in those situations is doing something worth examining. The company says it is releasing ReD as open source — the paper and an open-source release landed on the same day, according to The Verge's report of Waymo's announcement — and Waymo says it is collaborating with researchers, regulators, and SAE International on shared collision-avoidance standards.
There is a quiet circularity problem at the center of all this. The benchmark Waymo is offering the industry was designed by Waymo, in collaboration with academic partners, and the company will be among the first to measure its own vehicles against it. A yardstick built by a vendor is still a yardstick if everyone uses it the same way, but the choice of what counts as a "competent human" baseline is itself a value judgment. What ReD considers normal human braking, gap acceptance, and gap-closing behavior will shape the scorecards that follow.
The Delft co-authorship is the part most likely to blunt that concern, at least partially. Arkady Zgonnikov, an assistant professor at TU Delft and a named co-author, is a working academic in human decision-making and autonomous systems, and the work was peer-reviewed at Nature Communications rather than blog-published. That does not resolve the benchmark-design question, but it does mean the model is not a solo Waymo product.
The release fits a pattern. Waymo has been steadily producing peer-reviewed research and simulation infrastructure to argue that its safety case is structurally different from competitors', including earlier 3D world models and a 2022 study of "hyperattentive driver" perception. ReD is the next layer. Instead of showing that the cars drive a lot of miles without crashes, the company wants to show that the cars behave like an idealized human would, on a benchmark anyone can run.
What's worth watching is whether anyone outside Waymo uses it. SAE International's working groups on AV collision-avoidance benchmarks have been inching toward common methods for years, and a peer-reviewed, open-source model with named academic co-authors is a more credible seed than a vendor white paper. If a regulator cites ReD, or a competitor voluntarily tests against it, the circularity concern fades. If ReD becomes another internal Waymo metric that the rest of the industry politely ignores, it joins a long list of corporate-safety yardsticks that didn't take.
The Verge reports that the paper and an open-source release landed on the same day. The next signal is whether other AV developers, NHTSA, and SAE treat the model as a reference point or as a press release with a DOI.