Data-Driven Adaptive ML-Enabled Edge-Cloud System Framework for Safe and Efficient Autonomous Systems

Abstract

Machine learning (ML)-enabled systems like autonomous driving systems (ADSs) face challenges meeting safety and performance requirements in diverse environments, especially in resource-constrained, latency-sensitive edge-cloud settings. These challenges often arise from the ML models' limitations, including poor generalization to unseen conditions. Static ML models often struggle to generalize to unseen scenarios, particularly under the latency and resource constraints of edge-cloud infrastructure. Adaptive algorithms using ML system switching have been proposed, but existing approaches frequently lack generalizability, support for common black-box systems, and effective use of distributed edge-cloud resources. This paper presents a novel adaptive ML-enabled edge-cloud system framework to address these shortcomings. Our framework combines cloud-based pre-runtime analysis, which leverages simulation for behavioral understanding and scenario-to-system mapping, with collaborative edge-cloud runtime adaptation featuring dynamic ML model switching. It supports black-box systems and aims to balance safety and efficiency by utilizing appropriate edge and cloud resources situationally. Preliminary CARLA-based evaluation of the edge runtime component suggests our framework can potentially improve the safety-efficiency tradeoff compared to single-model ADSs in some scenarios. Moreover, extensive experiments using the MetaDrive simulator with 100,000 randomized driving scenarios demonstrate that the adaptive system improves safety by 2.6% while doubling computational efficiency compared to a single-model baseline. These results validate the framework's scalability and the feasibility of data-driven scenario-system mapping for adaptive ML-enabled autonomous systems operating across edge and cloud environments.