Waymo built a virtual driver to study how humans react to surprises on the road

Waymo has a lot of experience building virtual systems to help its autonomous vehicles better understand the real world. It built realistic 3D worlds to better anticipate natural disasters and unpredictable edge cases. It created a virtual representation of a hyperattentive driver to test against its own autonomous vehicles in a series of simulated scenarios to see which is better at crash avoidance.

Now, in a new research paper published today in Nature Communications, Waymo describes a new computer-based cognitive model that explains how human drivers make split-second decisions to avoid crashes. The company thinks the new model will serve as a benchmark to compare autonomous driving systems against as a way to help move the industry toward a greater degree of shared safety standards. It’s also the latest in Waymo’s growing body of peer-reviewed research that it says sets it apart from other autonomous vehicle operators.

Waymo designed the new model, referred to as ReD for “Reference Driver,” in collaboration with the Delft University of Technology in the Netherlands. Much in the way that the auto industry uses crash test dummies to evaluate a car’s structural integrity and hardware safety, this new model works as a behavioral dummy to determine how well an autonomous vehicle can avoid dangerous situations altogether.

“Evaluating AV safety is multifaceted, and understanding how a human handles conflict is a critical piece of the puzzle,” says Mauricio Peña, chief safety officer at Waymo. “By establishing this reference model of a competent human response, we can help the industry move toward a shared, scientifically grounded approach for evaluating collision-avoidance behavior.”

ReD relies on a neuroscience framework called active inference, championed by world-leading neuroscientists like professor Karl Friston (who called the ReD model a “technical tour de force” in a statement provided by Waymo). The core principle is that human brains constantly strive to minimize surprise over time.

ReD layers together several human cognitive traits to simulate how a driver handles this stress. Humans judge longitudinal threats based on “looming,” or how fast an object expands in their field of vision. Waymo’s model replicates this by naturally struggling to judge speeds at far distances, just like a real person. It accounts for a “traffic norm” filter that biases its predictions toward rule-abiding behavior, until it explicitly observes a vehicle violating a traffic norm. And it evaluates surprises just like a human driver, triggering a reevaluation of its driving once a surprise hits a certain threshold that suggests the current plan is failing. The model also accounts for how humans operate gas and brake pedals with a single foot by introducing a 0.2-second pause when shifting between the two.

“By grounding our model in active inference, we’ve achieved a holistic representation of human collision response,” says Arkady Zgonnikov, assistant professor at Delft University of Technology, in a statement. “This allows us to simulate the internal ‘surprise’ a driver feels during a conflict, providing a more human-like benchmark for autonomous driving systems that was previously impossible to automate at scale.”

Unlike traditional safety models that only simulate emergencies, Waymo says ReD is capable of “proactive avoidance” by continuously calculating surprise while minimizing free energy. This allows it to anticipate risks early and adjust its driving before a situation ever escalates to a conflict.

Waymo says it is actively collaborating with researchers, regulators, and standards organizations like the SAE to establish a consensus around these reference models. The goal is to move the autonomous vehicle industry toward a shared, scientifically grounded definition of what constitutes a “careful and competent” human response. To that end, the company is making the ReD model open source and publicly available to anyone who wants to test it out.