When Capabilities Become Actionable: Capability Grounding in Driving Foundation Models

Driving foundation models increasingly augment planning with richer representations, auxiliary supervision, and predictive objectives. Yet a model can benefit from such information without necessarily using it in the intended way when generating a trajectory. To address this problem, we formulate *capability grounding* as the question of whether capability-relevant information actually shapes the planned action in a way that is appropriate for the driving situation. We introduce **Capability Grounding Attribution (CGA)**, a controlled evaluation framework for testing this relationship. CGA constructs controlled comparisons that change capability-relevant evidence or supervision while holding alternative factors fixed, and examines how the planned action changes. Using CGA on four representative driving policies, we find that evidence of sensitivity, learning, or planning benefit can diverge from evidence that the intended information actually guides action. We further compare CGA diagnostics with closed-loop outcomes and find that the two capture complementary aspects of driving behavior: CGA identifies how information shapes planning, whereas closed-loop evaluation measures what happens when those plans are executed.