Predictive and Explainable UAV Navigation via World Models, Structured Future Reasoning and LLM-Based Explanation

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, Riyadh, Saudi Arabia
* Corresponding Author: Yasser Fouad. Email: email

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.085610

Received 14 May 2026; Accepted 13 August 2026; Published online 16 September 2026

Download PDF

Abstract

Autonomous UAV navigation in safety-critical environments requires more than accurate prediction of future states; it also requires the ability to interpret action-conditioned futures as actionable safety risk. This paper presents a predictive and explainable UAV navigation framework that combines a latent world model with a Future Interpretation Module (FIM), a safety-constrained decision layer, and a hybrid explanation module. The world model predicts action-conditioned latent futures, while the FIM converts these rollouts into interpretable descriptors including time-to-collision, minimum vertical clearance, route-commitment risk, and risk trend. These descriptors are used jointly for decision-making, safety filtering, and explanation grounding. In a delayed-hazard diagnostic benchmark with 150 episodes per agent, the FIM policy achieved 100.0% success and 0.0% collisions, compared with 97.3% success and 2.7% collisions for the prediction-only world-model baseline. In descriptor-level ablation, Full FIM achieved the highest success rate (71.3%) while eliminating collisions and route traps, whereas individual descriptors reduced collisions but often produced route-trap failures. Across six cross-environment validation scenarios and 900 Full-FIM episodes, the method achieved 98.0% success and 0.0% collisions. Because the prediction-only baseline was also near saturation in this broad suite, these cross-environment results are interpreted as robustness evidence rather than as a claim of uniform superiority. Overall, the results show that FIM improves diagnostic delayed-risk reasoning and provides robust collision-free behavior across the evaluated suite, while broad superiority over all prediction-only baselines remains outside the scope of the current evidence. Secondary comparisons with adapted YUME, Matrix-Game, NWM, and ANWM remain diagnostic references under a unified three-action UAV interface, not complete re-evaluations of those methods in their native settings.