Microsoft shows offloading robot inference to edge/cloud boosts success rates and battery life

Microsoft Research · rss · 2026-09-24

Microsoft Research systematically challenges the assumption that physical AI must run on onboard GPUs: in mobile manipulation workloads, small onboard GPUs slowed mapping/planning by up to 383% vs an A100 and cut VLA accuracy by 50%, while offloading inference to edge/cloud GPUs improved task success rates. Replacing onboard GPUs with a Raspberry Pi 5 extended battery life dramatically (Jetson Thor drained batteries by up to 160%). Microsoft also shipped an industry-first offloading capability in its open-source Physical AI Toolchain, using Kubernetes to containerize and orchestrate inference across robot, edge, and cloud, with examples for SO-101, UR10e, and Mobile Aloha.

Original post →

More from Embodied

Embodied channel →