UAV-VLPA*: Vision-Language Guided Global-Local UAV Mission Planning from Satellite Imagery

Abstract

We present UAV-VLPA* (Visual-Language-Path-and-Action), a vision-language system for autonomous UAV mission generation with the primary goal of achieving humanlevel path-planning capability using satellite imagery and natural language commands. The system's architecture combines: fewshot no training VLM generalization capabilities for language processing and waypoint generation, global waypoint optimization via the Traveling Salesman Problem (TSP), and local obstacle-aware path refinement through comparative analysis of A*, Dijkstra, Bidirectional A*, and discrete RRT-Connect algorithms. This work establishes that language-guided autonomous systems can match human expertise in complex path-planning scenarios, particularly in unstructured environments. Evaluated on the UAV-VLPA-nano-30 benchmark, UAV-VLPA* achieves human-competitive performance: its TSP+A* configurations generate optimized trajectories averaging 49.7 km, surpassing manual planning by 20.9 % by length while operating 5 times faster than expert human operators.

Related publications

• UAV-VLA: Vision-Language-Action System for Large Scale Aerial Mission GenerationACM/IEEE International Conference on Human-Robot Interaction · 2025 · Citations: 32 · Same author: Айбек Ахметказы
• Reliable Learning-Enabled Multi-Robot Coverage with Masked Entity-Attention MAPPO: A Controlled Multi-Seed StudyAI · 2026 · Shared topic: Robotic Path Planning Algorithms
• Adaptive Replay Strategies Stabilize Multi-Agent Reinforcement Learning for Differential-Drive Robot CoordinationJournal of Robotics and Control · 2026 · Citations: 5 · Shared topic: Robotic Path Planning Algorithms
• Autonomous agents and multi-agent systems: a comprehensive reviewArtificial Intelligence Review · 2026 · Shared topic: AI-based Problem Solving and Planning