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

Schumann, JF; Engström J; Johnson L; O'Kelly M; Messias J; Kober J; Zgonnikov A · 2026 · PubMed Central

DOI: 10.1038/s41467-026-73345-0

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the fragmentation in existing computational models of human collision avoidance, which typically focus on specific scenarios or isolated behavioral metrics like response times. To bridge this gap, the authors propose a unified computational cognitive model based on active inference, a framework that explains behavior through the minimization of expected free energy. The study aims to demonstrate that this approach can generalize across diverse driving contexts by simultaneously capturing response selection, timing, and execution. The model incorporates established cognitive mechanisms, including looming-based perception for threat detection and evidence accumulation for decision timing. It utilizes a norm-conditioned particle filter to predict other vehicles' trajectories, biasing predictions toward normative traffic behavior while allowing for norm violations when evidence suggests conflict. Policy selection is driven by minimizing expected free energy, balancing pragmatic values (e.g., avoiding collisions, maintaining velocity) and epistemic values (reducing uncertainty). The model accounts for human motor constraints, such as pedal transition times and jerk limits. The authors evaluated the model against empirical data from three distinct scenarios: front-to-rear braking, opposite-direction lateral incursion, and intersection right-turn conflicts. Parameters were fitted to the first two scenarios using meta-analyses and simulator studies, then applied without adjustment to the third to test generalizability. The results demonstrate that the model closely reproduces human behavior across all three scenarios. In the front-to-rear scenario, the model accurately replicated aggregate brake response times and deceleration magnitudes reported in literature, showing that braking intensity increases with kinematic urgency. It also correctly predicted maneuver selection, favoring braking at lower speeds and combined braking-steering at higher speeds. For the lateral incursion scenario, the model matched detailed human data regarding response timing and evasive maneuvers, including the tendency to swerve when kinematically viable. Crucially, the model successfully generalized to the unseen intersection scenario, reproducing human-like responses to vehicles failing to yield. The evidence accumulation mechanism explained response delays, while the norm-conditioned prediction prevented overly cautious behavior in non-conflict situations. The significance of this work lies in validating active inference as a generalizable framework for modeling complex, safety-critical human behavior. By unifying perception, prediction, and action under a single principle, the model overcomes the limitations of fragmented mechanistic models. This approach offers a robust tool for understanding cognitive mechanisms in driving, with implications for improving traffic safety, developing behavioral benchmarks for automated vehicles, and creating more realistic human agents in simulated test environments.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• anticipation
• mental model of traffic
• situational awareness

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model, theory or model