Adversary-Aware Neuroevolutionary Navigation: Extending Uncertainty-Conditioned Skill Libraries to Contested ISR Under Jamming, Spoofing, and Denied Sensing

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Autonomous ground and aerial robots are increasingly considered for intelligence, surveillance, and reconnaissance (ISR) in contested environments—settings in which an adversary actively degrades a robot’s ability to perceive, localize, and communicate. Most learning-based navigation systems, including my own prior uncertainty-conditioned skill library (UC-MESL), treat sensing degradation as passive and stationary: the environment obscures perception but does not strategically adapt to the robot’s behavior. In this paper, I remove that assumption. I present A-UC-MESL, an adversary-aware extension in which degradation is active, reactive, and deceptive: a jammer may deny GNSS and communications in response to the robot’s emissions or position, and a spoofer may inject false observations that corrupt the very uncertainty estimates on which a skill selector relies. My main contributions are: (i) a three-tier adversary model (static jammer, reactive jammer, deceptive spoofer) instantiated via custom ROS 2 and Gazebo plugins; (ii) a cross-modal sensor integrity layer that combines an instantaneous χ2 test on extended Kalman filter (EKF) innovations with a windowed cumulative-sum (CUSUM) statistic, so that slow-drift attacks designed to evade step-change detection are still exposed as a per-modality trust vector; and (iii) an emissions-discipline dimension in the quality-diversity behavior archive, yielding skills that trade exposure for speed. Empirical validation over 1,500 evaluation episodes—30 configurations run for 50 episodes each (5 fixed random seeds × 10 episodes)—shows that A-UC-MESL outperforms the strongest single-policy baseline on target confirmation (p < 10−6, Holm-corrected), reduces spoof susceptibility by 77–78% relative to the adversary-unaware ablation, and degrades gracefully under a deceptive adversary where single-policy baselines collapse. A drift-rate sensitivity study locates the detection boundary of the integrity layer: the cumulative statistic extends usable detection from 1.2ms−1 down to 0.03ms−1, forcing an undetected adversary into drift rates requiring 100s to accumulate a mission relevant position error. The scope is strictly non-lethal: reconnaissance and situational awareness, not target engagement.