Waymo has created a new computer model designed to more accurately answer the fundamental question: How does self-driving software compare to humans?
The Alphabet-owned robotaxi company, which worked with TU Delft to develop a computer model of human driving ability, published a research paper on this in Nature Communications on Wednesday.
Waymo said it expects the new model to be more accurate than previous versions it has used for the past several years. The new model was built using a framework called active inference, the theory that drivers constantly imagine possible futures and take actions to reach the safest and most predictable future.
Waymo said the new model will help it better understand how humans behave in crash scenarios faced by robotaxis.
“For decades, the automotive industry has used physical and virtual crash dummies to evaluate the safety features of cars, including hardware and structural integrity,” Waymo wrote in a blog post Wednesday. “The new model advances this concept and serves as a behavioral benchmark for autonomous driving systems that can realistically represent reasonable expectations of how a careful and competent human driver would respond to a traffic collision,” Waymo said.
A more accurate model of human driving behavior is table stakes for self-driving car companies that need to understand and rate the performance of robotaxis in crashes. And this comes at a critical time for Waymo as it expands into more cities and faces greater scrutiny from regulators and the public.
When a Waymo robotaxi struck a child last January near a school in Santa Monica, California, the company used previous computer models to argue that a attentive human driver would have delivered the impact at about 14 miles per hour. A Waymo robotaxi struck the child at 6 mph after slowing from 17 mph, and the company said the child suffered minor injuries. (This accident is still under investigation by the National Highway Traffic Safety Administration and the National Transportation Safety Board.)
The biggest difference between this new model, which Waymo calls the Reference Driver, and its predecessor is that it can replicate the behavior of a human driver when a crash occurs. Previously, Waymo’s model (and other industry models) focused on replicating “last-second reaction” human operations, according to the company.
Meanwhile, Reference Driver “can simulate the internal ‘surprise’ felt by a driver during a crash, providing a more human benchmark for autonomous driving systems where large-scale automation was previously impossible,” Arkady Zgonnikov, assistant professor at TU Delft, said in a statement.
Waymo said this new driver model could be suitable for modeling “a wide range of road user behavior beyond collision avoidance” and would be better suited for application to “large-scale test sets containing thousands of scenarios.”
“This model can represent and evaluate numerous complex real-world crashes in a virtual environment and identify performance improvements with unprecedented speed and efficiency,” the company wrote.
Waymo wants others to collaborate to further develop the reference driver. The company said Wednesday that it will make research code for the model available under an academic, non-commercial license that can be used in research, teaching, personal experiments and scientific publications.
If you purchase through links in our articles, we may receive a small commission. This does not affect our editorial independence.
