Armory: Serve large robot policies to 10+ robots on a single GPU, boosting throughput by 18%

danfei_xu · x · 2026-08-03

Stanford researchers introduce Armory, a system for serving large robot policies (VLA/WAM) from cloud/edge. By real-time scheduling of action chunks, Armory enables a single GPU to serve 10+ robots simultaneously without pausing. On heterogeneous fleets, it boosts task throughput by up to 18% over baselines. The work frames policy serving as a scheduling problem with real-time constraints, where robot waiting for inference can cause failures.

Original post →

More from Embodied

Embodied channel →