Waymo Builds New Human Driver Benchmark to Measure Robotaxi Safety

Waymo released a new safety benchmark in June 2026 called the Reference Driver, a computational model designed to compare how its robotaxis perform against a careful, competent human driver.

The model was developed in partnership with Delft University of Technology and uses a concept called active inference, which simulates how drivers constantly anticipate potential future scenarios rather than simply reacting to immediate hazards.

Unlike previous benchmarks that focused on last-second maneuvers, the Reference Driver can reproduce human behavior in the critical moments leading up to a potential collision. It accounts for the roughly 0.2-second delay inherent in human reaction time when moving from the accelerator to the brake, and it simulates the "internal surprise" a driver feels when a conflict develops unexpectedly.

Waymo said the model acts as a virtual crash test dummy for behavioral performance, giving the company a consistent, scientifically grounded standard for evaluating its vehicles. The company has made the research code available under an academic, non-commercial license to encourage broader industry adoption.

The benchmark arrives as Waymo scales its commercial operations. The company was providing roughly 400,000 rides per week across six U.S. markets by the first quarter of 2026. It also faces increased scrutiny from federal regulators following an incident involving a child in Santa Monica.

Waymo's safety data shows an 82 percent reduction in airbag deployment events and an 81 percent reduction in injury incidents compared to the human benchmark, though researchers note that proving safety across more complex environments remains an ongoing challenge.