A robot may execute its program while failing the task: grasping empty space, for example, or proceeding before a drawer opens. RoboAuto puts physical verification into program search. It stores knowledge of objects and possible interactions, grounds it in the current scene and runs small experiments when competing explanations remain.

The authors report 89.1% success on short LIBERO-PRO tasks and 73.4% on LIBERO-PRO Long, against ASPIRE's 30.8%. Three real-robot tasks are also evaluated. The appeal is inspectable, repairable behavior, but it relies on perception, reusable skills and safeguards for running generated code.