A Joint LLM and Optimization Agent Solution for Drone Scheduling in Victim Search Missions
Authors: Yinghao Zhang, Yalin Liu, Hong-Ning Dai, Yan Zhang, Qubeijian Wang, Bishenghui Tao, Yaohan Li
Organizations: Hong Kong Metropolitan University · Hong Kong Baptist University, Hong Kong, China · University of Electronic Science and Technology of China · Northwestern Polytechnical University, China
Victim search in wilderness areas is time-critical. Drones enable rapid area search, but scheduling them is challenging because victims may continue moving, detection results are uncertain, and effective decisions must combine structured mission information with unstructured contextual clues and domain knowledge. To address this, we propose an agent-aided drone scheduling system. The system employs an agent at the ground control station (GCS) to coordinate multiple drones over successive search cycles. In each cycle, the agent applies a scheduling policy that combines a large language model (LLM) for reasoning about where to search with an optimizer for determining how to dispatch the drones under energy constraints. Simulation results show that the proposed system consistently completes the victim-search mission, whereas none of the six non-LLM baselines does so and direct LLM scheduling is considerably less reliable. The source code is publicly available at https://github.com/1643204431/llm-uav-victim-search.
Figures & tables
Fig. 1: The agent-aided drone scheduling system for victim search.
Method / Variant
Found
Compl.
Cycles
Tcomp (s)
E/vic (kJ)
Tok/run
Tagent/cyc. (s)
Illegal
Fb.
Ours (DS V4 Pro)
5.0±0.0
5/5
3.8±1.8
4661±457
323±128
32k±16k
22±1
26.3%
0
LLM-Direct
4.0±1.2
2/5
28.0±11.0
6223±1083
1160±729
273k±127k
9±0
52.1%
6
w/o Knowledge Retrieval
4.6±0.9
4/5
9.6±12.5
4654±264
556±560
95k±130k
24±9
38.8%
0
w/o Mobility Reasoning
3.2±1.5
1/5
26.0±12.5
4682±0
1854±1654
289k±162k
17±7
33.8%
5
w/o Bottleneck
5.0±0.0
5/5
3.8±1.8
4703±453
327±136
32k±16k
22±1
26.3%
0
TABLE I: Architecture and component ablation results. Bold marks the best result in columns that permit direct comparison.
Backbone
Found
Compl.
Cycles
Tcomp (s)
E/vic (kJ)
Tok/run
Tagent/cyc. (s)
Illegal
Fb.
DeepSeek V4 Pro
5.0±0.0
5/5
3.8±1.8
4661±457
323±128
32k±16k
22±1
26.3%
0
DeepSeek V4 Flash
4.8±0.4
4/5
5.0±6.7
4399±253
406±428
91k±131k
54±21
12.5%
1
Qwen Flash
1.6±0.9
0/5
40.0±0.0
—
2201±65
566k±9k
19±2
93.0%
0
Qwen Plus
5.0±0.0
5/5
5.6±3.2
6426±1536
470±242
114k±70k
239±17
50.0%
0
Qwen3 Max
5.0±0.0
5/5
3.2±1.6
4907±623
274±83
52k±33k
142±21
29.4%
0
TABLE II: Cross-backbone results for Ours .
Method
Found
Compl.
Cycles
Tcomp (s)
E/vic (kJ)
Lawnmower
0.0±0.0
0/5
14.0±0.0
—
—
Greedy-Coverage
0.0±0.0
0/5
32.0±0.0
—
—
Bayesian Search
0.0±0.0
0/5
33.0±0.0
—
—
Clue-Prior Sweep
2.6±0.9
0/5
14.0±0.0
—
2186±1396
Clue-Prior Greedy
2.8±0.4
0/5
31.0±0.0
—
1690±344
Clue-Prior Bayes
3.0±0.0
0/5
28.0±0.0
—
1559±0
TABLE III: Main comparison results.
Fig. 2: Initial priority maps generated by five LLM backbones for seed 42 before any search observation. Cell color represents the inferred priority, white squares mark the top- B ( B=10 ) high-priority reachable cells, and stars mark the simulated victim-subgroup locations (hidden from the agent).
Fig. 3: Case study of the proposed agent solution. (a) shows the terrain map with four annotated landmarks. (b)–(d) show the inferred cell priorities at three reference moments. Heatmap intensity encodes wk , cyan traces show victim paths, and blue triangles mark previously dispatched search cells.
Fig. 4: Sensitivity to dispatch batch size B with N=10 . Top: completion time, with each label showing completed seeds. Bottom: total fleet energy. Curves show mean values and shaded bands show one standard deviation. The dashed line marks the default B=10 .