Waymo has published research with TU Delft in the journal Nature Communications on what it calls ReD—the Reference Driver—a computational model that simulates how careful, confident human drivers avoid collisions. The breakthrough represents a significant shift in how autonomous vehicles can be evaluated: rather than testing against average driver behavior, ReD creates a benchmark based on "careful and competent" human responses to traffic conflicts.
The work matters because autonomous vehicle safety has long relied on physical and virtual crash dummies to assess hardware and structural integrity, a methodology inherited from the automotive industry decades ago. ReD evolves that concept by creating a behavioral benchmark—a digital reference point for how a skilled human driver should respond when danger emerges. This means Waymo and other autonomous systems can now measure their performance against a realistic, neuroscience-grounded standard for human driving intelligence.
ReD is built on active inference, a predictive processing framework that models human driving behavior as the minimization of surprise. The model captures the closed-loop cognitive process a driver undergoes: updating their beliefs as a situation unfolds, managing uncertainty about other road users' intentions, and selecting an evasive maneuver—braking, swerving, or a combination. What sets ReD apart from traditional models is its capacity to simulate not just last-second reactive maneuvers, but proactive avoidance: how a competent driver anticipates potential risks before entering a conflict.
Arkady Zgonnikov, assistant professor at TU Delft, noted that "by grounding our model in active inference, we've achieved a holistic representation of human collision response." The research builds on Waymo's earlier work, including NIEON (Non-Impaired driver with Eyes ON the conflict), representing the company's twelfth published paper on behavioral reference models. The continuity between these models ensures a coherent scientific approach to understanding driver behavior across different scenarios and environments—including situations with significant uncertainty where other drivers' intentions remain unclear.
The research gained validation from Karl Friston, the neuroscientist who created active inference theory itself. Reviewing the Nature Communications paper, Friston called it "a remarkable piece of work" and "a tour de force in terms of generative modeling and scenarios considered." He emphasized how the model equips autonomous systems with the situational awareness and constrained information-seeking that both driving and daily life depend upon.
Perhaps most importantly, ReD scales. Because it is built on first principles from neuroscience rather than ad hoc engineering, the model can extend beyond collision avoidance to simulate adaptive driver behavior and other road user responses. This scalability transforms it from a single-use safety tool into a foundational framework for understanding how autonomous vehicles should behave across the spectrum of driving scenarios.
Waymo's framing of itself as aiming to become the "world's most trusted driver" reflects a deeper philosophical commitment: not just to build cars that avoid crashes, but to create systems that think like the most careful, most intentional human drivers on the road. ReD provides the scientific backbone for that ambition.