Research & Writing

Muhtasim Ahmed(Audri), Ashfaq Habib Rafi, Salman Farsi

This is an ongoing group research project and undergraduate thesis currently in active development. The work sits at the intersection of Vision-Language-Action architectures, autonomous driving, and AI reliability. VLA models have shown significant promise in enabling agents to perceive multimodal environments and follow natural language instructions to produce control outputs. However, a class of fundamental problems remains open around whether the reasoning these models produce is actually connected to the actions they take, and whether their behavior holds up under varied, imperfect, or ambiguous instructions. Our research investigates these problems in the context of autonomous driving, where the gap between what a model reasons and what it physically executes has direct safety implications. The work combines elements of reinforcement learning, imitation learning, and multimodal instruction following, with an evaluation framework covering instruction adherence, policy robustness, and generalization across different input conditions. A secondary objective is to demonstrate feasibility on a physical robotic vehicle as a simulation-to-real transfer example, connecting simulation-based VLA research to deployable autonomous systems. The full methodology and results will be disclosed upon completion of the research.