On June 10, 2026, research from Waymo and TU Delft landed in Nature Communications: a behavioral benchmark model called the Reference Driver, built to answer the autonomous industry’s most sensitive question — “would a human have done better?” TechCrunch’s Sean O’Kane covered it the same day, with Waymo framing the model as an evolution of the crash-test dummy: not a physical dummy testing a car’s structure, but a virtual, behavioral one for evaluating driving software in traffic conflicts.
The timing matters. January’s Santa Monica crash put Waymo’s human-comparison claims under their first real scrutiny, with both NHTSA and NTSB investigating. Releasing a more rigorous comparison methodology now is itself a message.
What the Reference Driver Is
Waymo’s critique of the status quo is blunt: existing driver-behavior models across the industry replicate only “last-second, reactive” maneuvers. The Reference Driver instead models the entire run-up to a conflict — how a driver gradually adjusts speed and lane position to avoid falling into danger in the first place, rather than jerking the wheel a half-second before impact. Because the baseline is a runnable model rather than a statistical average, it can be dropped into simulation and replayed against any candidate planner. Waymo says the model scales to “large test sets with thousands of scenarios” and extends to behaviors beyond collision avoidance, turning “robotaxi versus human” from a one-off press claim into a repeatable, systematic evaluation.
Active Inference: A Virtual Driver That Imagines Futures
Methodologically, the Reference Driver is built on active inference: the assumption that a driver constantly simulates possible futures and acts toward the safest, most predictable outcome. Arkady Zgonnikov, an assistant professor at TU Delft, notes the model can even simulate the internal “surprise” a driver experiences during a conflict — making it a behavioral agent with internal state, not a trajectory generator. For engineering teams, that is precisely the point: if you want to validate that a planner drives like a careful human, your baseline has to exhibit gradual risk-aversion, not just panic braking.
Why Now: The Santa Monica Crash
On January 23, 2026, a Waymo robotaxi struck a child who suddenly entered the roadway from behind a tall SUV near an elementary school in Santa Monica. The vehicle braked from roughly 17 mph and made contact at about 6 mph; the child sustained minor injuries. At the time, Waymo cited its existing peer-reviewed model to claim that a fully attentive human driver would have struck the pedestrian at approximately 14 mph. That “a human would have been worse” figure was produced by the very model the new work replaces. NHTSA’s Office of Defects Investigation is examining whether the vehicle used appropriate caution given the school proximity — a crossing guard and several double-parked cars were nearby — and the NTSB opened its own probe. That scrutiny did not start there: the agency was already examining Waymo vehicles for illegally passing stopped school buses in an Atlanta case and roughly 20 incidents in Austin. Waymo’s response was not to abandon comparison, but to upgrade the comparison tool to peer-reviewed standard: applied to a future incident, the Reference Driver could well produce a different number than 14 mph.
What the Open Research Code Means
Waymo also released the research code under an academic, non-commercial license permitting research, teaching, personal experimentation, and scientific publication — an explicit invitation for outside academics to improve it. That is the smart move. As long as the comparison baseline lives on one company’s servers, every “safer than human” claim carries a referee-owns-the-whistle problem; opening the method to independent teams like TU Delft’s is what turns the numbers into a shared industry asset. Waymo’s public safety impact pages already aggregate analysis across more than 220 million fully autonomous miles — the Reference Driver fixes the weakest link in that analysis, namely how the “human control group” is defined.
For developers and product teams, the paper is a template: any claim that “AI beats humans” is only as strong as the model of the humans. Shipping the baseline as reproducible, externally reviewable tooling is worth more than another single-incident victory lap.
Sources
- Waymo says it built a better benchmark for comparing robotaxis to humans — TechCrunch
- Waymo Safety Impact data pages
- Waymo robotaxi hits a child near an elementary school in Santa Monica — TechCrunch
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.