Open-Source System Serves VLA Models to 10+ Robots on a Single GPU

danfei_xu · x · 2026-08-03

Georgia Tech robotics researchers have released a new system capable of serving large Vision-Language-Action (VLA) models like π0.5 to fleets of 10+ robots using just a single GPU or cloud/edge setup.

The core of the system uses real-time scheduling of action chunks to prevent execution pauses. Unlike LLM serving, robotics control demands ultra-low latency; otherwise, robots exhaust precomputed actions, leading to jerky motion or task failure.

The system introduces a Lookahead scheduler that models batching as a closed-loop MDP to track robot states and delays. In tests with a real-world cluster of 10 AgileX PiPER arms, it achieved up to 18% higher throughput compared to baselines like Earliest-Deadline-First (EDF). The engine is fully open-sourced.

Related event: Armory System Serves 10+ Robots on a Single GPU(3 posts)→

Original post →

More from Embodied

Embodied channel →