Thumbnail Image


Publication or External Link





In this dissertation, I propose a software-in-the-loop testing architecture that uses adaptive sampling to generate test suites for intelligent systems based upon identifying transitions in high-level mission criteria. Simulation-based testing depends on the ability to intelligently create test-cases that reveal the greatest information about the performance of the system in the fewest number of runs. To this end, I focus on the discovery and analysis of performance boundaries. Locations in the testing space where a small change in the test configuration leads to large changes in the vehicle's behavior. These boundaries can be used to characterize the regions of stable performance and identify the critical factors that affect autonomous decision making software. By creating meta-models which predict the locations of these boundaries we can efficiently query the system and find informative test scenarios. These algorithms form the backbone of the Range Adversarial Planning Tool (RAPT): a software system used at naval testing facilities to identify the environmental triggers that will cause faults in the safety behavior of unmanned underwater vehicles (UUVs). This system was used to develop UUV field tests which were validated on a hardware platform at the Keyport Naval Testing Facility. The development of test cases from simulation to deployment in the field required new analytical tools. Tools that were capable of handling uncertainty in the vehicle's performance, and the ability to handle large datasets with high-dimensional outputs. This approach has also been applied to the generation of self-righting plans for unmanned ground vehicles (UGVs) using topological transition graphs. In order to create these graphs, I had to develop a set of manifold sampling and clustering algorithms which could identify paths through stable regions of the configuration space. Finally, I introduce an imitation learning approach for generating surrogate models of the target system's control policy. These surrogate agents can be used in place of the true autonomy to enable faster than real-time simulations. These novel tools for experimental design and behavioral modeling provide a new way of analyzing the performance of robotic and intelligent systems, and provide a designer with actionable feedback.