Open Access

ARTICLE

Structured Future Interpretation for Predictive and Explainable Autonomous Driving

Rihem Farkh¹, Ghislain Oudinet², Alaeddine Moussa³, Yasser Fouad^4,*

1 LabISEN, KLaIM, ISEN Méditerranée, Toulon, France
2 LabISEN, VISION-AD, ISEN Méditerranée, Toulon, France
3 Laboratoire LIS UMR CNRS 7020, Aix Marseille Université, Marseille, France
4 Department of Applied Mechanical Engineering, College of Applied Engineering, Muzahimiyah Branch, King Saud University, Al Muzahimiyah, Saudi Arabia

* Corresponding Author: Yasser Fouad. Email:

Computers, Materials & Continua 2026, 89(2), 47 https://doi.org/10.32604/cmc.2026.086607

Received 02 June 2026; Accepted 08 July 2026; Issue published 15 September 2026

Abstract

Autonomous driving systems must reason not only about the current scene but also about how the environment may evolve under alternative actions. Although predictive world models can generate future latent rollouts, these rollouts are often consumed directly by planners or explanation modules without an explicit and auditable interpretation stage. This paper presents a predictive and explainable driving framework centered on a Future Interpretation Module (FIM), which transforms action-conditioned future rollouts into structured descriptors, including risk trend, peak risk, time-to-critical, minimum clearance, predicted collision, dominant predicted event, and confidence. An aligned latent interface, trained with feature-alignment and structured-consistency objectives, reduces semantic drift when predicted future states are interpreted by the structured perception head. The resulting summaries support safety-aware action selection and prediction-conditioned natural-language explanations grounded in the same future evidence used by the planner. The evaluation protocol covers seven scenario families, three difficulty levels, five independent seeds, and ten episodes per scenario-difficulty-seed combination, yielding 1050 matched episodes per agent and 9450 episodes in the principal nine-agent comparison. Compared with the strongest non-FIM baseline, risk-aware MPC, Full FIM increases success from 0.84 to 0.89, reduces collision from 0.08 to 0.05, improves risk-anticipation accuracy from 0.83 to 0.89, and reduces the Decision Optimality Gap from 0.56 to 0.32. Descriptor ablations, sensitivity tests, noise stress tests, latency measurements, explanation-grounding metrics, and failure analysis further characterize the method. The results support explicit future interpretation as a useful mechanism for controlled predictive driving scenarios, while simulation-based evaluation remains insufficient to establish real-world safety or universal superiority.

---

Keywords

---