30 credits - Evaluating and Improving Explainable Vision-Language Models for Autonomous Driving Introduction A Master's thesis is an excellent way to get closer to TRATON Group R&D and build relationships for the future. Background Vision-language models VLMs are increasingly used in autonomous-driving research to interpret traffic scenes, answer questions, describe relevant objects, and provide natural-language explanations for driving decisions. These explanations could support model development, debugging, safety analysis, and communication between automated vehicles and human operators. However, a plausible explanation is not necessarily a faithful explanation. A VLM may produce fluent reasoning that is not grounded in the visual observations, omit safety-critical objects, hallucinate scene elements, or provide an explanation that is inconsistent with its predicted decision. Conventional language-generation metrics may reward similarity to a reference answer without revealing whether the explanation reflects the model’s actual reasoning. The thesis will initially rely on public autonomous-driving datasets, open-source VLMs, and language-based driving benchmarks such as DriveLM and LingoQA. Depending on the selected direction, the work may investigate models that explain scene understanding, behavior prediction, trajectory planning, or end-to-end driving decisions. Objective The thesis will investigate how the explainability of vision-language models for autonomous driving can be evaluated and improved. The work should distinguish between explanations that are merely plausible and explanations that are grounded in the observed scene and faithful to the model’s output. Possible research directions include: defining measurable properties of useful driving explanations, such as correctness, visual grounding, faithfulness, completeness, uncertainty awareness, and decision consistency; evaluating whether generated explanations refer to the correct objects, road users, traffic rules, and spatial relationships; comparing free-form natural-language explanations with structured explanations based on scene graphs, object references, concepts, or causal relations; grounding explanations in image regions, object tracks, bird’s-eye-view representations, or other intermediate model features; comparing intrinsic explanation mechanisms with post-hoc methods applied to existing VLMs; using visual perturbations or counterfactual scenes to test whether explanations respond to safety-relevant changes; detecting hallucinated, unsupported, or internally inconsistent explanations; developing training objectives that improve agreement between visual evidence, predicted behavior, and generated explanations; investigating calibrated explanations that communicate uncertainty or abstain when the available evidence is insufficient; or studying whether explanations help engineers or human operators identify model failures more efficiently. The selected approach should be
