Waymo says it built a better benchmark for comparing robotaxis to humans
Waymo unveiled a new human-driving model it says is better at judging robotaxi performance in crash scenarios. The company is releasing the research code under an academic non-commercial license.
Intelligence analysis by GPT-5.4 Mini
Waymo has published a new benchmark called the Reference Driver, built with TU Delft, to compare autonomous driving behavior with human driving more realistically. It is meant to model what a careful driver would do before and during conflict, not just last-second reactions.
Waymo made a new computer test that tries to act more like a careful human driver. It is like swapping a simple doll crash test for a pretend driver who can imagine trouble before it happens.
Analysis
What Waymo built
Waymo says it created a new computer model for human driving behavior that should be more accurate than the one it has used for years. The model, developed with TU Delft and described in a Nature Communications paper, is based on active inference, a framework that treats driving as a process of imagining possible futures and choosing the safest one.
What is different
Waymo says older models in the field mostly focused on a driver’s last-second reaction in a crash. The new benchmark, which it calls the Reference Driver, is intended to reproduce behavior earlier in the sequence, including the sense of surprise a human might feel before a conflict. According to the company, that makes it a more human-like benchmark for evaluating autonomous systems.
Why Waymo is doing this now
The timing matters. Waymo is expanding into more cities while also drawing more attention from regulators and the public. The company has already used an earlier model in a January crash in Santa Monica, California, where a robotaxi struck a child near a school. Waymo said that model suggested an attentive human driver would have hit at about 14 miles per hour; the robotaxi hit at 6 miles per hour after slowing from 17 miles per hour.
What happens next
Waymo says the new model can handle large test sets with thousands of scenarios and can be adapted beyond collision avoidance. It is also releasing the code under an academic, non-commercial license, which means researchers can use it for study, teaching, experimentation, and publication. That invitation suggests Waymo wants others to help improve the benchmark, not just use it internally.
Key points
- Waymo published a new human-driving benchmark called the Reference Driver with TU Delft.
- The model uses active inference to simulate how a careful driver might think before a crash.
- Waymo says the new approach is better than models that mainly capture last-second reactions.
- The company points to a January Santa Monica crash as an example of why better comparisons matter.
- Waymo is releasing the research code under an academic, non-commercial license.
If the model works as Waymo says, it could give robotaxi makers a better way to measure safety against human driving. Sharing the code for research could also help other people improve the benchmark and make comparisons more useful across the industry.
The benchmark may still be only a model, not a full picture of real-world driving behavior. If regulators, researchers, or the public do not trust the assumptions behind it, it may do little to reduce scrutiny of Waymo’s crash performance.
