Year: 2026
Type: preprint
Abstract: Vision-language systems can identify mission targets in aerial imagery, but the order in which a model emits those targets is not necessarily a usable flight route. We present ARIES-Mission, a fixed routing layer that generates four complete candidates---the raw recognition order and routes produced independently by ant colony optimization, differential evolution and particle swarm optimization---and returns the shortest feasible closed loop under a common geodesic objective. Unlike an oracle analysis, this terminal choice uses only route lengths available at planning time. On 30 satellite-image missions and 20 stochastic runs per task, the portfolio reduced total distance from 79.65 to 60.37~km (24.21\%). Among 27 missions admitting Held--Karp verification, its mean gap to the exact optimum was 0.042\% and the maximum was 0.468\%. ACO supplied 87.3\% of selected routes, while DE, PSO and the raw route remained useful on a minority of runs. The portfolio was also modestly shorter than nearest-neighbour, insertion and 2-opt baselines. A candidate-subset ablation attributed most of the reduction to ACO, with DE and PSO narrowing the residual exact-optimality gap and the raw route retained as a no-cost fallback. These results delimit the conditions under which transparent route competition improves recognition-order geometry within a modular vision-language mission pipeline.
Source: Research Square
Institution: Macao Polytechnic University
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7172000826