Researchers built an AI system that lets delivery drones double as flying cloud servers - and used a language model just to write the math problem.
The setup comes from a cloud manufacturing scenario: drones fly between factory stations, picking up finished products and hauling them to a central depot. Along the way, sensors at those stations generate computing jobs - process locally, hand them to a passing drone, or bounce them up to the cloud. Researchers built two pieces to manage this: an agentic AI that uses large language models, retrieval-augmented generation, and chain-of-thought reasoning to turn a plain-English problem description into a formal mathematical formulation, and a two-layer reinforcement learning system where one layer plans drone routes and the other assigns computing tasks in real time.
The interesting part isn't the drones - it's using an LLM as a translator between human intent and the rigid math that optimization algorithms actually need, a step usually hand-coded by an engineer. In simulations, the reinforcement learning half collected every product in 99.6% of its last 500 test runs and hit every processing deadline 100% of the time, beating a more common baseline algorithm called advantage actor-critic.
It's a lab result on a simulated warehouse, not a fleet of drones over a real factory floor, and the paper is light on how the LLM's formulations get checked for errors - the kind of detail that matters before anyone lets this set delivery routes for real machinery.