Lessons from RAITE: real-world evaluation of robust multimodal target detection and tracking under adversarial attacks
Baker Herrin, Aditya Penumarti, Andres Pulido, Nikhil Iyer, You-Long Wu, Kaleb Lewellen, Tyler Fitzsimmons, Jane Shin · 2025
This paper presents lessons learned from the University of Florida team’s deployment of a robust, multi-modal perception and tracking system during the 2024 Robust Artificial Intelligence Test Event (RAITE), a field exercise designed to assess AI system resilience under adversarial conditions, hosted by NSWC Crane. The deployed system integrated RGB, EO/IR, and radar sensing with deep learning-based object detectors, homography-based fusion, and Kalman filter-based tracking. We incorporated a real-time retraining pipeline using nightly-collected adversarial data and employed evaluation metrics such as failure rate and time-to-recovery to quantify system resilience. Over three days of red/blue team testing, we encountered diverse physical attacks, including occlusion, visual degradation, and distractor interference. Our findings highlight how sensor failed over di!erent types of attacks and how sensor redundancy, multi-view fusion, and adaptive retraining contributed to system robustness. We share practical insights into failure modes, recovery dynamics, and design considerations for deploying resilient AI perception systems in complex and adversarial field environments.