Waymo Develops Virtual Driver to Analyze Human Reactions to Road Surprises
Waymo has been at the forefront of developing virtual systems that enhance its autonomous vehicles’ understanding of the real world. By creating realistic 3D environments, the company can better prepare for natural disasters and unpredictable scenarios. They’ve even built a virtual model of an exceptionally attentive driver to compare against their autonomous systems in various simulated situations, focusing on crash avoidance capabilities.
In a recent research paper published in Nature Communications, Waymo introduces a new cognitive model that sheds light on how human drivers make quick decisions to avoid accidents. The company believes this model will serve as a benchmark for evaluating autonomous driving systems, pushing the industry toward unified safety standards. This research adds to Waymo’s growing collection of peer-reviewed studies, distinguishing it from other players in the autonomous vehicle space.
Developed in partnership with Delft University of Technology in the Netherlands, the new model, named ReD for “Reference Driver,” functions similarly to crash test dummies used in the auto industry. Instead of assessing structural integrity, ReD acts as a behavioral dummy to evaluate how effectively an autonomous vehicle can navigate dangerous situations.
“Evaluating AV safety is multifaceted, and understanding how a human handles conflict is a critical piece of the puzzle,” explains Mauricio Peña, Waymo’s chief safety officer. “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 is built on a neuroscience framework known as active inference, which has been supported by leading neuroscientists like Professor Karl Friston, who praised the model as a “technical tour de force.” The fundamental idea is that human brains are constantly working to minimize surprises over time.
The model integrates various human cognitive traits to mimic how a driver copes with stress. For instance, humans assess longitudinal threats based on “looming,” or the rate at which an object grows in their field of vision. Waymo’s model mirrors this by struggling to gauge speeds at long distances, just like a real driver would. It incorporates a “traffic norm” filter that biases its predictions toward compliant behavior until it observes a vehicle breaking traffic rules. Additionally, it evaluates surprises similarly to human drivers, prompting a reassessment of its driving strategy once a surprise exceeds a certain threshold, indicating that the current plan may be failing. The model also simulates the way humans operate gas and brake pedals with a single foot, introducing a 0.2-second delay when switching between the two.
“By grounding our model in active inference, we’ve achieved a holistic representation of human collision response,” states Arkady Zgonnikov, an assistant professor at Delft University of Technology. “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 react to emergencies, Waymo asserts that ReD can engage in “proactive avoidance” by continuously assessing surprise while minimizing free energy. This capability enables it to foresee risks early and adjust its driving before a situation escalates into a conflict.
Waymo is actively working with researchers, regulators, and standards organizations to build consensus around these reference models. The aim is to guide the autonomous vehicle industry toward a shared, scientifically grounded definition of what constitutes a “careful and competent” human response. To facilitate this, the company is making the ReD model open source and available for anyone interested in testing it.
