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

Technical Summary: Active Inference as a Model of Collision Avoidance Behavior in Human Drivers

Problem Statement
Collision avoidance is a critical, high-stakes cognitive task requiring rapid threat detection and the execution of appropriate evasive maneuvers. Existing computational models of human collision avoidance are largely fragmented; they typically focus on specific scenario types (e.g., front-to-rear conflicts), specific explanatory factors (e.g., cognitive load), or isolated behavioral metrics (e.g., response times). While machine learning models trained on large driving datasets show promise in generalization, they often struggle to represent safety-critical behaviors due to the under-representation of collision scenarios in training data. Consequently, there is a lack of unified models capable of capturing the key aspects of human collision avoidance—response selection, timing, and execution—across diverse scenarios.

Methodology
The authors propose a computational cognitive model based on active inference, a framework grounded in the free energy principle which posits that agents act to minimize "surprise" (or expected free energy, EFE). The model simulates a human driver (the ego agent) interacting with other road users in a time-discrete, sequential process.

Key mechanistic components of the model include:

1. Looming-based Perception: The agent does not directly perceive kinematic states (position/velocity) but infers them from visual angle ($\phi$) and angular rate ($\dot{\phi}$, or looming). A detection threshold ($\dot{\phi}_0$) is implemented to simulate the inability to perceive small relative velocities at long distances, explaining delays in recognizing events like abrupt braking.
2. Norm-Conditioned Particle Filter: To predict the future behavior of other agents, the model uses a particle filter biased by traffic norms (e.g., staying in lanes, yielding right-of-way). This prevents overly pessimistic predictions in non-conflict situations while allowing the model to relax these assumptions and sample "long-tail" norm-violating trajectories when evidence of a conflict arises.
3. Evidence Accumulation for Re-planning: The agent continuously accumulates "surprise" (defined as the residual negative pragmatic value of the current policy) as evidence for a full policy re-plan. If accumulated evidence exceeds a threshold, the agent re-plans the entire policy; otherwise, it extends the current policy. This mechanism models human response timing dynamics.
4. Constrained Policy Sampling: New policies are generated using the cross-entropy method, limited by bounded planning capacity (a fixed number of evaluated policies). Crucially, acceleration constraints mimic human pedal operation, including a 0.2-second holding time at near-zero acceleration to simulate the physical delay of moving a foot between pedals.
5. Policy Selection: The agent selects the policy that minimizes Expected Free Energy (EFE), balancing pragmatic value (avoiding collisions, maintaining speed, staying in lane, minimizing control effort) and epistemic value (reducing uncertainty). In collision scenarios, pragmatic value is the primary driver.

Evaluation and Results
The model was evaluated against empirical human data across three distinct collision avoidance scenarios:

1. Front-to-Rear: A lead vehicle brakes suddenly. The model reproduced meta-analytic findings on brake response times and deceleration magnitudes, showing a dependency on scenario kinematics (e.g., shorter response times and harder braking in more urgent scenarios). It also predicted maneuver selection (braking vs. swerving) based on speed and time gaps.
2. Opposite-Direction Lateral Incursion: An oncoming vehicle encroaches into the ego lane. The model was tested against simulator data with varying incursion "steepness." It successfully reproduced human tendencies to steer left in steep scenarios, steer right in shallow scenarios, and collide in "medium" scenarios due to high uncertainty about the other vehicle's path.
3. Intersection Right-Turn-Into-Path: A vehicle fails to yield at an intersection. The model, tuned only on the first two scenarios, generalized to this unseen scenario, reproducing human collision rates and response times (faster responses in more critical "not-stopped" variants).

Ablation Studies
Systematic removal of model mechanisms (e.g., removing evidence accumulation, noise in prediction, or norm-conditioning) demonstrated that each component is essential for capturing specific aspects of human behavior. For instance, removing evidence accumulation resulted in unrealistically fast responses, while removing norm-conditioning led to overly conservative behavior or premature reactions.

Significance and Claims
The paper claims to present the first published computational model offering a detailed account of human collision avoidance behavior across multiple conflict scenarios based on first principles. Key contributions include:

• Unified Framework: Demonstrating that active inference can serve as a generalizable framework for modeling human behavior in complex, safety-critical driving tasks, unifying response timing, maneuver selection, and execution.
• Generalization: Showing that a model tuned on specific scenarios can accurately predict human behavior in unseen scenarios without re-parameterization.
• Mechanistic Insight: Providing a computational account for why human drivers sometimes fail to avoid collisions (e.g., in medium incursion scenarios), attributing these failures to the interplay of bounded planning capacity, prediction uncertainty, and the bias toward norm-compliant behavior.
• Closed-Loop Dynamics: Unlike many existing models that use "one-shot" open-loop maneuvers, this model implements dynamic, closed-loop re-planning similar to human decision-making.

The authors position active inference not merely as a computational tool but as a theoretical framework for understanding road user behavior, bridging the gap between cognitive theory and practical traffic safety applications. They acknowledge limitations, such as the assumption of non-reactive other agents and the need for future work on cultural norm variations and more complex perceptual variables, but assert that the model provides a robust foundation for simulating human agents in automated vehicle testing and safety research.