Active inference as a model of collision avoidance behavior in human drivers

Article https://doi.org/10.1038/s41467-026-73345-0 Active inference as a model of collision avoidance behavior in human drivers Received: 2 June 2025 Accepted: 6 May 2026 1234567890():,; 1234567890():,; Check for updates Julian F. Schumann 1, Johan Engström2 , Leif Johnson2, Matthew O’Kelly2, Joao Messias2, Jens Kober 1 & Arkady Zgonnikov 1 Collision avoidance – involving a rapid threat detection and quick execution of the appropriate evasive maneuver – is a critical aspect of driving. However, existing models of human collision avoidance behavior are fragmented, focusing on specific scenarios or only describing certain aspects of the avoidance behavior, such as response times. This paper addresses these gaps by proposing a computational cognitive model of human collision avoidance behavior based on active inference. Active inference provides a unified approach to modeling human behavior: the minimization of free energy. Building on prior active inference work, our model incorporates established cognitive mechanisms such as evidence accumulation to simulate human responses in three distinct collision avoidance scenarios: front-to-rear lead vehicle braking, lateral incursion by an oncoming vehicle, and another vehicle failing to yield at an intersection. We demonstrate that our model explains a wide range of empirical findings on human collision avoidance behavior. Specifically, the model closely reproduces both aggregate results from metaanalyses previously reported in the literature and detailed, scenario-specific effects observed in two recent driving simulator studies, including response timing, maneuver selection, and execution. Our results highlight the potential of active inference as a generalizable framework for understanding and modeling human behavior in complex real-life driving tasks. Main Collision avoidance is a critical skill for human drivers. It involves the rapid detection of threats (such as a vehicle ahead suddenly braking) and deciding on an appropriate evasive maneuver (for instance, braking or swerving). These maneuvers are complex, requiring not only precise execution but also continuous adjustments as the situation evolves. Furthermore, drivers need to account for the uncertainty in the future behavior of other road users: for example, will an oncoming vehicle encroaching into my lane continue across or move back to its own lane? Understanding how humans avoid collisions in traffic can provide insights into high- stakes, split-second decision making which has substantial implications for traffic safety. Behavior models play a key role both in understanding the mechanisms of human collision avoidance and in improving traffic safety. These models are applied in diverse contexts, such as collision risk estimation1, understanding effects of driver distraction2, modeling take-over behavior3, representing human agents in simulated test environments4, and providing behavioral benchmarks for automated vehicles5,6. Besides immediate practical applications, modeling human collision avoidance behavior is an interesting subject of study in its own right. Being a highly complex and dynamic task with extremely 1 Department of Cognitive Robotics, Delft University of Technology, Delft, Netherlands. 2Waymo LLC, Mountain View, CA, USA. e-mail: Nature Communications | (2026)17:5009 1 Article high stakes, collision avoidance provides a unique testbed for theories and models of cognition previously not validated in the real world7–9. Most existing computational models of human collision avoidance are mechanistic, that is, are based on the explicit modeling of cognitive mechanisms underlying response timing and evasive maneuvering. However, they are typically fragmented, focusing either on specific scenario types (e.g., front-to-rear conflicts10–12 or merging13), specific explanatory factors (such as off-road glances1,14 or cognitive load2), or only reproduce certain aspects of human behavior (such as response times3,6,10,11,15 or the extent of steering16). Altogether, these models cover a wide range of scenarios and diverse aspects of collision avoidance behavior. Yet, each one of these mechanistic models on its own is highly specific: they are not designed to generalize to multiple scenarios or describe multiple aspects of human behavior. Recently, machine learning models based on large datasets of human driving have demonstrated the ability to generalize across a wide range of traffic scenarios17–22. Because such models typically generate full motion trajectories, they also have the potential to represent multiple aspects of collision avoidance behavior and not just a single metric of interest. However, a key challenge is that safetycritical behavior, such as collision avoidance, is often underrepresented in the datasets used for training23, which makes it hard to achieve representative human-like collision avoidance behavior solely based on learning from data24. Thus, there is currently a lack of models that can capture the key aspects of human collision avoidance behavior (response selection, timing, and execution) all at the same time and across different scenarios. This limits both practical applications (due to the need to develop a new model for every new scenario) and fundamental understanding of cognitive mechanisms underlying the behavior of humans in safety-critical situations in traffic (due to the lack of a unified explanation for multiple aspects of behavior). To address this gap, here we present a model of human collision avoidance behavior based on active inference. Originating in computational neuroscience, active inference is a versatile general framework for understanding and modeling sentient behavior in living systems25–27 that has been previously used to model human behavior in diverse contexts28–32, including the modeling of human driver behavior33 such as car following34, responses to driving automation failures15, and managing uncertainty around occlusions and nondriving-related tasks35. Building on the model of Engström et al.35, our active inference model is designed to reproduce a wide spectrum of human behavior (e.g., chosen maneuver, reaction timing, collision likelihood, etc.) in response to potential collisions. To this end, our model incorporates several well-known cognitive mechanisms to represent the dynamics of human decision making in response to sudden stimuli, such as looming perception36,37 and evidence accumulation38–41. We evaluated our model against a range of previously reported empirical findings in three paradigmatic collision avoidance scenarios: 1) the front-to-rear scenario, where a driver needs to respond to a suddenly braking vehicle in front, 2) the oppositedirection lateral incursion scenario, where an oncoming vehicle suddenly cuts across the driver’s path, and 3) the intersection right-turninto-path scenario, where a vehicle coming from the right on a perpendicular road does not yield when turning right into (...truncated)