Cloud-on-Demand: Automating Safety-Aware Cloud DNN Invocation for Edge-Cloud Hybrid CPS

Abstract

Hybrid edge-cloud perception DNNs combine fast on-board inference with more accurate but delayed cloud assistance for learning-enabled CPS. The key challenge is deciding when cloud inference should be invoked: always-on cloud usage improves information quality but increases latency, bandwidth, and monetary cost, while fixed schedules ignore time-varying control difficulty. We formulate this problem as runtime design automation for safety-aware cloud invocation in delayed edge-cloud CPS. To this end, we propose Cloud-on-Demand, a lightweight scheduling framework based on the Shift-Response Policy (SRP) and its causal refinement SRP-o. The method uses self-normalized feature-shift and control-response-shift signals with cloud-freshness safeguards to adapt cloud invocations online. We also compare against an approximate Bayesian neural network (BNN) baseline using Monte Carlo dropout to convert predictive uncertainty into a cloud-query signal. We evaluate the approach on a delayed F1Tenth edge-cloud autonomous racing stack using cloud-call rate, maximum cross-track error, RMSE, and failure-related outcomes. Results show that SRP-family methods substantially reduce cloud usage while remaining competitive with fixed-interval, probabilistic, and BNN-based baselines in worst-case tracking performance; at Dcloud = 20, SRP operates at about 4.4% cloud-call rate, while the BNN baseline achieves lower maximum cross-track error only by using substantially more cloud budget (≈ 15.9% CCR) and worse RMSE. These results demonstrate that automated, safety-aware cloud invocation improves the cost-performance tradeoff in CPS with hybrid edge-cloud perception DNNs.