Waymo Unveils New Model to Benchmark Robotaxis vs. Humans

Waymo's push to demonstrate its robotaxis outperform human drivers in avoiding crashes now rests on a new computational model of attentive driving behavior. Regulators, riders, and city officials weighing expanded operations need reliable ways to judge whether the vehicles truly reduce risk, especially after incidents like the January Santa Monica collision that left a child with minor injuries.

The company, working with TU Delft researchers, developed the Reference Driver using an active inference framework. This approach models how drivers continuously predict outcomes, register surprise when expectations break, and adjust speed or steering to restore safety. Unlike earlier versions limited to last-second reactions, the new model incorporates looming perception, traffic-norm expectations, and a 0.2-second pause when switching between accelerator and brake.

Waymo states the model was described in a research paper published in Nature Communications. That specific journal placement could not be independently verified in public records, which instead point to SAE papers and an arXiv preprint. The company released the code under an academic non-commercial license and invited collaboration with standards bodies such as SAE.

Mauricio Peña, Waymo's chief safety officer, described the model as a step toward shared evaluation methods for collision avoidance. Arkady Zgonnikov of TU Delft noted it can simulate internal surprise at scale. In the Santa Monica case, Waymo previously used its older model to estimate that an attentive human would have struck the child at 14 mph; the robotaxi made contact at 6 mph after slowing from 17 mph.

The effort arrives as Waymo expands service and faces ongoing NHTSA and NTSB review of that crash. By releasing the code, the company positions the Reference Driver as a potential industry tool rather than a proprietary claim, though external validation of its predictive accuracy across thousands of scenarios remains ahead.