Etudiant : OUEDRAOGO WENDE BENEDO ARIEL

Filière : Master Machine Learning Avancé et Intelligence Multimédia (MLAIM)

Encadrant : Pr. BOUMHIDI JAOUAD

Annèe : 2026

Résumé : Autonomous Driving Systems (ADS) have historically relied on modular pipelines in which perception, prediction, planning, and control are handled by separate, task-specific models. While effective in structured environments, these architectures exhibit fundamental limitations in generalization, particularly in rare, ambiguous, or semantically complex driving scenarios. The emergence of Foundation Models (FMs), large-scale models pretrained on broad, heterogeneous data through self-supervised learning has opened a new research direction, offering richer semantic representations, chain-of-thought reasoning, and multimodal fusion capabilities that traditional approaches cannot provide. This work presents a comprehensive study of Foundation Model integration across the ADS stack. We examine how FMs are applied to perception, prediction, planning, and end-to-end driving, reviewing representative architectures, datasets, and benchmarks from the recent literature. Special attention is given to the practical conditions of real-world deployment, analyzing both the promises demonstrated by current systems and the structural limitations that constrain their operational use. To evaluate the practical integration of Foundation Models in Autonomous Driving Systems, two closed-loop simulation studies are conducted in controlled driving environments. The first evaluates a Large Language Model as the primary decision-making agent in highway-env, where driving actions are selected through prompt-based reasoning and tool-augmented ReAct interactions. The second investigates a hybrid autonomous driving architecture in CARLA, where a Large Language Model operates as a high-level planner that generates semantic driving goals, while a conventional PID controller executes the corresponding low-level vehicle control. Both simulations include representative driving scenarios such as overtaking, merging, emergency braking, and dense traffic to assess decision quality, safety, reasoning consistency, and real-time feasibility. Results indicate that Foundation Models provide strong semantic reasoning capabilities and interpretable decision-making while achieving safe behavior across a variety of closed-loop driving scenarios. However, their deployment remains constrained by inference latency, limited long-horizon reasoning, and the lack of formal safety guarantees. The findings support the use of Foundation Models as high-level planning and reasoning modules within hybrid Autonomous Driving Systems rather than as complete replacements for conventional control pipelines, and highlight low-latency inference, robust safety verification, and standardized closed-loop evaluation as key challenges for real-world deployment