Search for: All records

It remains a challenge to provide safety guarantees for autonomous systems with neural perception and control. A typical approach obtains symbolic bounds on perception error (e.g., using conformal prediction) and performs verification under these bounds. However, these bounds can lead to drastic conservatism in the resulting end-to-end safety guarantee. This paper proposes an approach to synthesize symbolic perception error bounds that serve as an optimal interface between perception performance and control verification. The key idea is to consider our error bounds to be heteroskedastic with respect to the system’s state — not time like in previous approaches. These bounds can be obtained with two gradient-free optimization algorithms. We demonstrate that our bounds lead to tighter safety guarantees than the state-of-the-art in a case study on a mountain car.