Closed-Loop Online Inference Assurance for Cyber-Physical Systems with Neural Network Controllers

Yufeng Yang, Mengyu Liu

Keywords: cyber-physical systems, AI, real-time

TL;DR: Closed-loop framework that estimates inference time budgets online and adapts early-exit neural networks to meet dynamic deadlines, improving control reliability in time-critical CPS under varying hardware workloads.

Abstract: Neural network (NN) policies deployed in time-critical cyber-physical systems (CPS), such as autonomous robots and intelligent transportation platforms, often experience time-varying inference latency when executed on embedded hardware. This variability arises from factors including input-dependent computational complexity, shared resource contention, operating system scheduling, and fluctuating workload conditions. As a result, the actual inference completion time may deviate from its nominal value, leading to delayed actuation, missed control deadlines, and degraded closed-loop performance. In safety-critical CPS, such timing uncertainty can accumulate over successive control cycles, potentially compromising stability, safety, and task completion guarantees. In this work, we demonstrate a closed-loop inference assurance framework that couples online inference budget estimation with time-adaptive NN inference. The experiment results show the superiority of the proposed framework in terms of control performance and inference performance compared with baselines.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 11