This thesis investigates how coverage-driven data collection can improve the robustness and generalization of vision–language–action (VLA) models. The student will define coverage criteria for relevant task and environment characteristics and use them to guide the selection or collection of training data. The resulting models will be evaluated on both familiar situations and unseen combinations of these characteristics.

Collecting diverse robot demonstrations is costly, and increasing dataset size alone may leave important situations underrepresented. How can coverage criteria guide data collection to achieve better generalization within a limited data budget? The thesis will compare coverage-driven approaches with conventional sampling under equal data budgets, examining task success, robustness to variations, and limitations of the chosen coverage criteria.

• Committed to producing research suitable for scientific publication.
• Good coding skills, including effective use of AI coding assistants.
• Hardworking and willing to turn ambiguous problems into concrete research tasks.
• A background in formal methods or combinatorial testing is a plus.