A clear edge in simulation
An LLM-guided UAV, using a large language model to interpret the search task, located a banana, a computer mouse and a toy zebra in every reported guided run across the simulated scenarios. The conventional lawn-mower scan had a lower success rate in each case. The guided runs also reported shorter search times and paths than the baseline.
For the banana scenario, the LLM-guided method had a 100% success rate, compared with 65% for the lawn-mower baseline. Among trials where both approaches found the target, mean search time was 36.02% lower and mean path length was 52.26% lower with the guided method. The reported variability figures were 30.52 for search time and 38.16 for path length.
In the computer-mouse scenario, the LLM-guided method showed a 100% success rate versus 75% for the baseline. Its mean search-time reduction was 24.83%, and its mean path-length reduction was 43.50%. For the toy zebra, success was 100% versus 95%, while mean search time was 29.24% lower and mean path length 51.93% lower with the guided approach.
The simulation used 120 flights in a 10-by-10-meter environment. The three scenarios began from 20 randomly selected starting positions each, with 60 flights using the LLM-guided method and 60 using the lawn-mower strategy.
How the drone chose a route
The system’s central idea was to use context to decide where to look first. It combined real-time object detection, three-dimensional spatial mapping, LLM semantic reasoning and polynomial spline interpolation for trajectory planning. The research question was whether that semantic navigation could help a UAV locate a user-specified object with less time and path length than a conventional lawn-mower scan while maintaining high success.
Once the system selected a relevant object, it used A* path planning in three-dimensional space, followed by seventh-order polynomial splines to generate a smooth trajectory. The initial language-model reasoning step averaged 9.97 seconds, while later reasoning ran concurrently with the flight.
The evaluation defined success as localizing the target. Its other main measures tracked the percentage reduction in elapsed time from takeoff to successful localization and the percentage reduction in distance traveled to localization. The paper reported mean reductions with standard deviations, but the time and path comparisons were calculated only for trials in which both methods successfully localized the target.
The real-world test was narrower
The real-world portion was a described validation in which semantic reasoning directed a Crazyflie 2.1 nano-UAV to search around a monitor and successfully locate a computer mouse. It used an 8-by-10-meter flight arena and OptiTrack localization.
The report does not give the number of real-world trials or quantitative comparison measures. It therefore offers a demonstration of successful localization, not a measured real-world advantage over lawn-mower scanning.
The reported experiments also relied on a pre-mapped setup. Environments were mapped externally, camera images were processed offboard with YOLO11, OctoMap ray-casting was used for spatial grounding, and the implementation used OpenAI’s o3-mini language model. This means the evaluation did not test full onboard mapping in an unknown environment or establish that the complete system would run onboard in real time.
A promising result with a narrow reach
The authors summarize the evaluated framework as achieving 100% mission success, with average mission duration up to 36% lower and path length more than 50% lower than conventional lawn-mower patterns. That is an author-level aggregate summary; the supplied analysis does not provide pooled uncertainty estimates or the calculation behind that wording.
The evidence covers three simulated object-search scenarios and 120 quantified flights, plus one described real-world test in pre-mapped indoor space. It does not establish broad deployment, performance in unknown environments, or outcomes such as safety, collision, energy use, computing load or human benefit.
The document is identified as arXiv:2608.28270v1, dated 28 Aug 2026, and no funding source is reported in the supplied text.
Paper data and sources
Original title: Spatial-Semantic Reasoning using Large Language Models for Efficient UAV Search Operations
Authors: Marin Maletic, Marijana Peti, Tamara Petrovic, Stjepan Bogdan
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: 10.1109/ecmr65884.2025.11163229
Original paper · Full text