Verification and Design of Robust and Safe Neural Network-enabled Autonomous Systems

摘要

Neural networks are being applied to a wide range of tasks in autonomous systems, such as perception, prediction, planning, control, and general decision making. While they may improve system performance over traditional physical model-based methods, pressing concerns have been raised on the uncertain behaviors of neural networks under varying inputs, especially for safety-critical systems such as autonomous vehicles and robots. In this paper, we will discuss the challenges in ensuring the safety and robustness of neural network-enabled autonomous systems, and present our recent work in addressing these challenges. These include methods for certifying the robustness of neural networks, verifying the safety of neural network-controlled systems, designing these systems with safety assurance, and conducting safety-assured runtime adaptation.