AI-Assisted V&V for Adaptive Space Systems: A Search-in-the-Loop Approach

As robotic space missions push into increasingly remote and poorly characterized environments, onboard autonomy is becoming essential. Yet the verification and validation (V&V) methods used to certify autonomous systems before launch were designed for largely preprogrammed spacecraft, and they do not scale to systems whose behavior evolves in response to unknown conditions. This paper addresses one of the central challenges in this transition: how to systematically identify scenarios that stress autonomous mission performance without exhaustive enumeration or arbitrary worst-case assumptions. We introduce Gremlin, an AI-based adversarial framework for simulation-based V&V of autonomous spacecraft. Rather than defining test cases upfront, Gremlin acts as an adversary within the simulation, adaptively injecting disturbances during execution and using Monte Carlo Tree Search (MCTS) to identify scenarios that degrade mission performance. Two contributions distinguish it from existing approaches such as Adaptive Stress Testing. First, disturbances are modeled as continuous correlated functions through a Gaussian-process (GP) prior, rather than discrete predefined values. Second, a global Mahalanobis distance constraint on the joint disturbance sequence (informally, the insanity budget) ensures that identified scenarios remain statistically plausible under the assumed uncertainty model, rather than unconstrained worst cases. The framework is evaluated on the descent phase of NASA’s Uranus Orbiter and Probe (UOP) concept. Gremlin identifies disturbance sequences that degrade mission performance by reducing the science data returned to the orbiter by up to 19.7%. Under the same disturbance model, MCTS-guided search consistently identifies more adversarial scenarios than a single random draw from the same admissible set, and the global constraint produces a meaningful trade-off between adversarial depth and statistical plausibility. Gremlin is broadly applicable to any autonomous system that can be evaluated in simulation under uncertainty, and represents a step toward AI-driven V&V tools capable of supporting the design and certification of autonomous systems for missions to environments that cannot be fully characterized before launch.