An AI agent can now chart its own drone flight path and correct course mid-flight without new training.
Researchers built LAPF, a framework that wires a large language model into a closed-loop system with perception, memory, planning, and action modules, then tested it on the UAVScenes dataset in town-scale outdoor settings. Across three trials per method, LAPF flew average paths of 512.83 meters and 506.37 meters against a straight-line optimum of 497.33 meters, in open-field and obstacle-injected scenarios respectively. That works out to path efficiencies of 97.1% and 98.1%, and reductions of 17.2% and 15.6% compared to a standard chain-of-thought prompting baseline. LAPF also tied every detected hazard to a bounded corrective action with zero clamp events in both tests, while the chain-of-thought baseline racked up 9.7 to 14.0 clamp events.
Most autonomous-navigation systems still lean on optimization models or reinforcement learning trained for specific tasks, which struggle when conditions shift. Earlier LLM-based navigation added reasoning but lacked real memory or planning, so agents repeated the same mistakes instead of learning from them. LAPF's real contribution is less the language model itself than the scaffolding wrapped around it, which is where most of the recent gains in agentic AI have actually come from.
Three trials per method on a benchmark dataset is a proof of concept, not a flight-ready product, and "zero clamp events" measures internal stability, not real-world safety in traffic, wind, or rain.