Utilizing Machine Learning for Robustness Testing of Autonomous Systems

Engineering notes

Key topics: autonomous driving system, autonomous driving, reinforcement learning, large language model, deployment. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Autonomous systems—such as Autonomous Driving Systems (ADSs) and Uncrewed Aerial Vehicles (UAVs)—are increasingly deployed across transportation and service applications. Because they operate in safety-critical settings, rigorous pre-deployment robustness testing is imperative. Yet existing approaches still struggle to expose diverse, realistic corner cases at scale, leaving residual risks in real-world operation. This thesis investigates machine-learning–driven robustness testing for autonomous systems in both simulation and real-world settings, with an emphasis on advancing test generation and automation. For search-based testing, we introduce GARL, which couples genetic algorithms with reinforcement learning to synthesize diverse failure cases; the approach is validated in simulation and on a physical UAV auto-landing system. We then turn to ADS testing and develop MARL-OT, a multi-agent RL–guided online fuzzing framework that reveals cooperative safety violations arising from interactions among vehicles in dense traffic. To further improve coverage under complex road conditions, we propose PtoP, a hybrid framework that combines adaptive random seed generation with Stein Variational Gradient Descent and provides a plug-in API for online testing methods; this design efficiently produces failure-inducing initial scenarios for ADS evaluation. Finally, moving to real-world data via metamorphic testing, we present AutoMT, a multi-agent framework powered by large language models that automatically extracts metamorphic relations from traffic regulations and generates valid follow-up test cases. This automation increases scenario diversity and fault detection on real-world datasets while reducing human effort. Collectively, these contributions yield scalable, automated, and practical pipelines that integrate machine learning, search-based methods, and large language models. The resulting toolchain advances the robustness testing of autonomous systems and supports safer, more trustworthy deployment of next-generation autonomy.

Links and sources

Comments