GovAI report: frontier LLMs are not yet able to support autonomous attack drones
In a recent report, "General-Purpose AI and the Future of Drone Autonomy in Contested Environments," GovAI researcher Jake Steckler assesses that current LLMs are not yet sufficient to reliably complete most drone tasks and are too slow; frontier LLMs, even using data-center compute, can only be used for the most time-lenient tasks such as planning, smaller LLMs remain too slow under on-board compute, and VLA models run faster but have not yet demonstrated the required capabilities. GovAI introduced the report on X, and Steckler's post added that VLA models in particular still fall short of requirements in real environments and are still not fast enough.
How feasible is it today to use frontier AI models to enable autonomous drones on the battlefield?
GovAI researcher @JakeASteckler addresses this and many other questions in his recent report, "General-Purpose AI and the Future of Drone Autonomy in Contested Environments".
Read it here: https://www.governance.ai/research-paper/general-purpose-ai-and-the-future-of-drone-autonomy-in-contested-environments
Takeaway 2: Today’s LLMs aren't yet good enough for most drone tasks, and they're far too slow.
For autonomous attack drones to perform reliably on the battlefield, they’d have to do a variety of things that require a combination of low- and high-level reasoning. They’d have to avoid obstacles, track targets, verify that those targets meet mission criteria, and do a variety of planning tasks like changing a mission when unexpected conditions arise.
Frontier LLMs can't yet reliably accomplish these tasks, and even with ample compute from data centers, they're are much too slow for anything except the most time permissive tasks, such as planning, as you can see in the chart below.
Much smaller LLMs, when run with less-powerful onboard compute, are still too slow. Many tasks like flight control and obstacle avoidance would likely be passed to lower-level controllers using more traditional robotics policies, but if general-purpose AI models are involved, they’ll likely be a rate limiter regardless.
Vision-language-action (VLA) models offer another possible approach beyond LLMs, and they can currently run much faster than LLMs. But these models have also not yet shown the requisite capabilities (especially in real-world conditions), and they are still not quite fast enough for the most time constrained tasks.
3/View the quoted post on X