Learning Active Inference MODELs of Perception and Control: Application to Car Following Task

Garcia, Alfredo · 2023 · ROSA P / Safety through Disruption (Safe-D) University Transportation Center (UTC)

archive: archived pipeline: cataloged verified

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

Summary

This paper introduces the Active Inference Driver Agent (AIDA), a novel computational model of driver behavior that integrates perception and control through the framework of active inference. The research addresses the need for interpretable, theoretically grounded models of human driving that can match the predictive performance of data-driven neural networks while offering greater transparency into decision-making mechanics. The authors motivate this approach by framing driving as a process where agents jointly perceive and act to maximize the match between perceived and preferred states of the world, minimizing free energy. The methodology builds upon Partially Observable Markov Decision Process (POMDP) theory, modeling a Bayesian agent that maintains beliefs about hidden environmental states and selects actions to maximize expected rewards minus information processing costs. The authors formulate a learning procedure to estimate the model’s primitives—specifically the internal generative model of perception (state transitions and observation likelihoods) and the reward function representing preferences—from observed sequences of vehicle trajectories. This estimation is structured as a bi-level optimization problem. The application utilizes the INTERACTION dataset, which provides high-frequency (10 Hz) time-indexed trajectories of vehicle positions, velocities, and headings. The model was evaluated on car-following tasks under two conditions: same-lane and new-lane scenarios. The results demonstrate that AIDA significantly outperforms the rule-based Intelligent Driver Model (IDM) across all evaluation metrics. Furthermore, AIDA performs comparably to state-of-the-art data-driven neural network benchmarks in terms of prediction accuracy, as measured by offline and online evaluation metrics such as MAE-IQM and ADE-IQM. Visualizations of generated trajectories confirm the model’s ability to replicate realistic driving behaviors, though some instances of rear-end collisions were noted in online evaluations. Crucially, the study highlights that AIDA offers superior interpretability compared to neural network models, allowing for a clearer understanding of the input-output mechanics driving the agent’s decisions. The significance of this work lies in bridging the gap between theoretical cognitive models and practical autonomous driving applications. By demonstrating that an active inference model can achieve performance parity with complex neural networks while retaining structural interpretability, the paper suggests that such frameworks are viable for safety-critical systems where understanding agent behavior is essential. The authors conclude that future work should expand training data to include more diverse driving environments and explore extensions to capture heterogeneity across different drivers, further validating the model’s generalizability and utility in transportation safety research.

Key finding

The active inference-based driver agent significantly outperformed the rule-based Intelligent Driver Model on all metrics and performed comparably to neural network benchmarks while offering superior interpretability.

Methodology

modeling

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. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

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

• 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