A research team tests plain-language mission orders for fleets of underwater robots in simulation

A preprint describes a framework for coordinating several autonomous underwater vehicles from instructions written in plain language. A large language model turns the instructions into assignment and ordering constraints, an optimiser assigns tasks using each vehicle's capabilities, and each vehicle runs its own behaviour tree on board so work continues when communication is limited.

In a simulator built on the Robot Operating System, the authors compared two small language models. Qwen-3-8B extracted the constraints more accurately than Llama-3.1-8B, and all tasks were completed in the tested sensor-degradation and vehicle-fault scenarios. The results are simulation only and the preprint has not been peer reviewed, so real-water performance is untested.