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

Julian F. Schumann, Johan Engström, Leif Johnson, Matthew O'Kelly, Joao Messias, Jens Kober, Arkady ZgonnikovarXiv 2025 · DOI · arXiv

Introduction

Understanding how human drivers avoid collisions is critical for improving traffic safety and developing realistic benchmarks for autonomous vehicles. However, existing computational models of human driving behavior are highly fragmented. Mechanistic models typically focus on single scenarios or specific metrics like reaction time, while modern machine learning models often struggle with the rare, safety-critical events under-represented in training data. This paper proposes a unified cognitive model based on the active inference framework, capable of capturing the full spectrum of collision avoidance—response selection, timing, and execution—across diverse traffic scenarios.

Methods

The model is built on the principle that agents act to minimize "expected free energy," effectively avoiding surprise and pursuing preferred states (e.g., avoiding collisions). To simulate human-like constraints, the authors integrated several key mechanisms:

  • Looming-based perception: Instead of exact state knowledge, the model relies on visual looming (the rate of change of an object's visual angle) with human-like detection thresholds.
  • Norm-conditioned prediction: The model initially assumes other agents follow traffic norms (e.g., lane keeping). It only considers norm-violating trajectories when evidence becomes overwhelming.
  • Accumulating surprise: Rather than reacting instantly, the model accumulates "surprise" ($E$)—evidence that the current plan will fail. A full evasive re-plan is triggered only when $E$ crosses a threshold.
  • Constrained policy sampling: Potential maneuvers are evaluated using the cross-entropy method, subject to physical human constraints like the time required to switch from the gas to the brake pedal.
The active inference cycle for collision avoidance, showing observation, norm-conditioned prediction, and surprise-based re-planning.

Results and Key Insights

The model was evaluated against empirical human data from front-to-rear conflicts and lateral incursion scenarios.

Accurate Response Timing and Maneuver Choice

In front-to-rear scenarios, the model accurately reproduced human brake response times (~1.4s) and demonstrated that deceleration magnitude scales linearly with the inverse time-to-collision. Furthermore, it correctly predicted a preference for braking at lower speeds and combined braking-and-swerving at higher speeds.

Replicating Human Failure

A critical finding occurred in "medium" angle lateral incursions. The model frequently crashed, mirroring human simulator data (~65% crash rate). This reveals that the oncoming car's trajectory is highly ambiguous in this specific geometry, preventing the agent from committing to a swerve until it is physically too late.

Model performance across different conflict scenarios, illustrating realistic response timing and maneuver selection compared to human data.

The Role of Surprise and Norms

The "surprise" mechanism explains why humans respond faster as conflicts escalate more rapidly. Additionally, the norm-conditioned filter prevents "paranoia"; without it, the model would unrealistically assume every oncoming car might swerve into its lane. Ablation studies confirmed that removing evidence accumulation or physical constraints broke the model's alignment with human timing data.

Conclusion

By unifying perception, prediction, and action under active inference, this model provides a powerful tool for understanding human driving behavior. While currently limited by hand-tuned parameters and non-reactive agents, future work aims to incorporate multi-agent communication and generative ML models to scale the framework for complex, real-world road infrastructures.

Summary updated

Abstract

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 novel 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 two distinct collision avoidance scenarios: front-to-rear lead vehicle braking and lateral incursion by an oncoming vehicle. We demonstrate that our model explains a wide range of previous empirical findings on human collision avoidance behavior. Specifically, the model closely reproduces both aggregate results from meta-analyses previously reported in the literature and detailed, scenario-specific effects observed in a recent driving simulator study, including response timing, maneuver selection, and execution. Our results highlight the potential of active inference as a unified framework for understanding and modeling human behavior in complex real-life driving tasks.
BibTeX
@misc{schumann2025active,
  author = {Julian F. Schumann and Johan Engström and Leif Johnson and Matthew O'Kelly and Joao Messias and Jens Kober and Arkady Zgonnikov},
  title = {{Active inference as a unified model of collision avoidance behavior in human drivers}},
  year = {2025},
  eprint = {2506.02215},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.02215},
}
Figure