ElasticInfer: Safe Contextual Model Routing under Dynamic Edge-Resource Constraints

On-device intelligence for embodied agents must operate under highly dynamic resource constraints caused by shifting concurrent workloads and thermal throttling. Static deployment strategies either risk severe memory and latency violations or conservatively sacrifice task accuracy. We present ElasticInfer, an adaptive framework that decouples pre-execution resource admission from online model selection. By combining a Nano-first feasibility filter with explicit delivery accounting and cooperative allocation leases, ElasticInfer prevents catastrophic out-of-memory (OOM) and deadline failures, while dynamically optimizing perception quality across a portfolio of detectors. Evaluated on 5,000 COCO validation images under simulated resource dynamics, ElasticInfer achieves 0.406 mAP at a mean latency of 12.3 ms, utilizing contextual Gaussian Thompson sampling to provide label-free exploration of the remaining accuracy-cost trade-off. In a physical GPU evaluation featuring concurrent compute, managed allocation, and warmed input shapes, the guarded admission protocol successfully delivers 100% of its 180 queries within a strict 50 ms deadline (averaging 21.31 ms) with zero observed system violations. These results demonstrate that explicit resource admission and selective context acquisition are the primary drivers of safe edge inference, allowing embodied agents to gracefully defer queries rather than suffer unregulated system crashes.