A machine learning technique for automating the traditional controller tests process that evaluates autonomous-vehicle software controllers is discussed. In the proposed technique, a controller is subjected to an adaptively chosen set of fault scenarios in a vehicle simulator, and then a genetic algorithm is used to search for fault combinations that produce noteworthy actions in the controller. This approach has been applied to find a minimal set of faults that produces degraded vehicle performance and a maximal set of faults that can be tolerated without significant performance loss. >