Waymo has developed a new computational model aimed at better addressing a key question: how does its self-driving technology compare to human drivers?

The robotaxi firm owned by Alphabet, which collaborated with TU Delft to create the model simulating human driving abilities, released a research article on this topic in Nature Communications on Wednesday.

Waymo indicated that it anticipates the new model will outperform the previous one employed over recent years. This updated model utilizes a framework known as active inference—suggesting that a driver continuously envisions potential futures and makes choices to navigate towards the safest, most foreseeable outcome.

The company indicated that this new model will enhance its understanding of human actions in crash situations encountered by its robotaxis.

“For decades, the automotive sector has utilized physical and virtual crash dummies to assess a car’s safety attributes, focusing on both its hardware and structural soundness,” Waymo mentioned in a blog entry on Wednesday. This new model, they stated, “advances this idea, acting as a behavioral standard for autonomous driving systems that can realistically illustrate what can be expected from a careful and skilled human driver when faced with traffic conflicts.”

An accurate representation of human driving behavior is essential for autonomous vehicle firms needing to evaluate the crash performance of their robotaxis. This comes at a pivotal time for Waymo, which is expanding into additional cities and facing increased scrutiny from regulators and the general public.

In January, after a Waymo robotaxi collided with a child near a Santa Monica school in California, the company utilized its prior model to assert that a vigilant human driver would have made contact at about 14 miles per hour. The Waymo robotaxi struck the child at only 6 miles per hour, having slowed down from 17 miles per hour, and the company reported minor injuries sustained by her. (The incident is still under investigation by the National Highway Traffic Safety Administration and the National Transportation Safety Board.)

The primary distinction between this new model—the Reference Driver—and its predecessor lies in its capacity to replicate a human driver’s actions leading up to a crash. Past models by Waymo (and other industry efforts) concentrated on mimicking “last-second, reactive” human responses, according to the firm.

In contrast, the Reference Driver can “emulate the internal ‘surprise’ experienced by a driver during a conflict, offering a more human-centric standard for autonomous driving systems that was previously unfeasible to automate at scale,” stated Arkady Zgonnikov, an assistant professor at TU Delft.

Waymo claims this new driver model is adaptable to simulate a “diverse array of road user behaviors beyond just avoiding collisions,” and that it is more capable for use in “extensive test sets comprising thousands of scenarios.”

“The model can illustrate and assess various intricate, real-world collisions within a virtual setting, pinpointing performance enhancements with unparalleled quickness and effectiveness,” the company noted.

Waymo is also eager for others to join in advancing the Reference Driver. The firm announced on Wednesday that it is releasing the research code for the model under an academic, non-commercial license, enabling its usage for research, instruction, personal trials, and scientific publication.