Introducing Waymo’s New Reference Model for Human Collision Avoidance

Waymo and TU Delft have published joint research in *Nature Communications* introducing ReD, a new reference model designed to simulate how a careful and competent human driver avoids collisions. The model builds on Waymo’s existing predictive processing framework, expanding capabilities to represent the full cognitive process of threat detection, belief updating, and evasive action selection. Unlike traditional models focused on last-second reactions, ReD emphasizes proactive avoidance by anticipating risks before conflicts arise. The research positions ReD as a behavioral benchmark for autonomous systems, replacing physical crash dummies with a human-like standard for evaluating safety performance in complex traffic scenarios.

The ReD model is grounded in active inference theory, a neuroscience-based framework that treats human driving as the minimization of surprise. It simulates how drivers manage uncertainty about other road users’ intentions and select optimal maneuvers such as braking, swerving, or combined actions. The model’s closed-loop cognitive workflow allows it to represent realistic human responses across diverse environments, including those with unclear driver intentions. Waymo states ReD advances its decade-long safety research program, which includes over a dozen published papers on behavioral reference models, ensuring continuity with prior work like the NIEON model.

Waymo and collaborators highlight the scalability of ReD, noting its potential to model a wide range of road user behaviors beyond collision avoidance, such as adaptive driving and interactions between vehicles. The model operates fully automated, eliminating the need for manual rules or annotations, and enables rapid evaluation of thousands of real-world crash scenarios in virtual environments. Mauricio Pena, Waymo’s Chief Safety Officer, emphasizes that establishing a human reference model is critical for developing a shared, scientifically grounded approach to autonomous vehicle safety evaluation across the industry.

To foster collaboration, Waymo is releasing the research code for its active inference model under an academic license, permitting use for research, teaching, and scientific publication. The company is also working with regulators and standards bodies like SAE to establish consensus on applying driving behavior reference models in autonomous vehicle evaluations. By providing a transparent benchmark for competent human responses, Waymo aims to support a safer and more predictable autonomous driving ecosystem through open scientific contributions and industry-wide alignment.