Armory boosts robot fleet throughput 18% with smart action chunk scheduling

Deploying robot foundation models like Vision-Language-Action (VLA) models at scale is a major goal for general-purpose robotics, but these models are computationally heavy, and on-device computing is limited by power and space. A new paper from researchers at Georgia Tech introduces Armory, a serving system that runs robot policies on a remote GPU and streams action chunks to multiple robots simultaneously. This turns the problem into a scheduling challenge: how to batch requests efficiently while meeting the real-time, closed-loop requirements of robot control.

The team found that naive scheduling heuristics work well when all robots are identical, but degrade when robots consume action chunks at different rates—a common real-world scenario with heterogeneous fleets. Armory's algorithm accounts for this variability, improving overall throughput by up to 18% in experiments with both simulated and physical robots. This is an early step toward making remote, shared compute a viable option for fleets of intelligent robots, potentially lowering the cost and hardware requirements for deploying VLA-based automation in warehouses, factories, and other environments.