Compute-Aware Deployment at the Edge: How Test-Time Routing, Temporal Asymmetry, Inference Efficiency, Force-Sensor Surrogates, and Safety Certification Jointly Constrain Real-Time Robot Policy Execution

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This version corrects two citation errors found by an automated check and confirmed by hand. In the Selection Process, HANDOFF was cited as arXiv:2606.06491 (TempoVLA) and now cites arXiv:2606.06493, and RoboNaldo was cited as arXiv:2606.11091 (QUIET, a network-neuroscience paper) and now cites arXiv:2606.11092. A reference to an earlier internal synthesis and labels pointing to an internal review rubric are removed, and typographic dashes are removed. This version has not had a full claim-by-claim audit. The correction note is at the top of the PDF. Deploying learned robot policies on physical hardware imposes constraints that laboratory benchmarks routinely ignore: bounded computational budgets, closed-loop latency requirements, the absence of expensive sensing modalities, and the need for certifiable safety under uncertainty. This paper synthesises five recent findings from the cs.RO and eess.SY arXiv corpus to argue that real-time robot policy execution is jointly constrained by a cluster of deployment-time engineering decisions, decisions about when to spend compute, how to schedule temporal streams, how to meet inference-speed targets on edge silicon, how to recover missing sensing modalities cheaply, and how to certify safety without compromising task efficiency. This is a heuristic reading, not a formal derivation: the five findings do not share a common mathematical framework, but they converge on a shared structural problem, the gap between what a policy can do at training time and what it can do within the hard real-time envelope of physical deployment. We draw on DIRECT's test-time compute routing (arXiv:2606.12402), AHA-WAM's asynchronous temporal decoupling (arXiv:2606.09811), RhinoVLA's edge-codesign methodology (arXiv:2606.07383), FACTR 2's sensor-surrogate approach (arXiv:2606.12406), and the belief-space safety filter certification of (arXiv:2606.02562). Together, these sources suggest that the deployment gap is not primarily a model-capacity problem but a resource allocation problem: latency, token count, sensor cost, and safety conservatism are all currencies that can be traded against one another if the right architectural interfaces are in place. Falsification paths are identified for each claim. The abstract-only reading methodology is disclosed as a limitation throughout. The BeliefSF result (arXiv:2606.02562) connects to the deployment-efficiency thesis through conservatism reduction rather than through a shared mechanism with the other four papers; it is treated as a weakly connected but plausible component and should be read accordingly. Authorship: Saluca Agentic AI Research Team (Saluca LLC). AI-drafted synthesis from an arXiv preprint corpus, originally drafted 2026-06-15, produced under the direction of Cristian Ruvalcaba, the accountable human author. Not peer-reviewed. AI disclosure. This work was produced with an agentic AI research apparatus operated by Saluca Labs. The apparatus drafted, searched and analysed under direction. Cristian Ruvalcaba is the human author and is accountable for the content. No AI system is listed as an author or contributor, because authorship entails accountability that a model cannot hold; this disclosure is the credit, and it is deliberately the whole of it.