Rapid and Hierarchical UAV Exploration via Adaptive Regional Viewpoint Generation
Xiaolu Zhang, Jiangjian Xiao, and Zhiqiang Li
AI summary
Problem
Autonomous UAV exploration in large-scale environments is hindered by excessive long-distance back-tracking and frequent low-speed flight, caused by greedy local planning and the high computational cost of global optimization.
Approach
The method online partitions the environment into subregions using hgrid decomposition, optimizes their traversal order via a non-closed traveling salesman problem, and rapidly selects safe candidate viewpoints near frontiers using a global loss function that balances coverage and velocity constraints.
Key results
- Online hgrid spatial decomposition for adaptive subregion partitioning
- Non-closed TSP formulation for globally optimized subregion traversal
- Velocity-aware loss function for rapid, safe viewpoint selection and trajectory refinement
- Validated superior exploration efficiency and flight velocity over state-of-the-art methods in simulation and real-world tests
Why it matters
Enables faster, safer, and more efficient autonomous mapping for UAVs in large-scale or dynamic environments, benefiting search-and-rescue, inspection, and 3D reconstruction applications.
Abstract
UAVs exploring large-scale environments rapidly and autonomously face significant challenges. Exploration effi- ciency is mainly limited by two factors: excessive long-distance back-tracking and frequent low-speed flight. To address the two issues, we propose a spatially hierarchical exploration method with fast regional viewpoint generation. The exploration area is partitioned online into subregions using an hgrid-based spatial decomposition, and the exploration sequence of these subregions is determined by solving a non-closed traveling salesman prob- lem. Within each subregion, viewpoints are selected by optimizing a global frontier-related loss, yielding a set of globally consistent targets. This hierarchical partitioning and global path opti- mization reduce back-tracking distance. Meanwhile, the chosen viewpoints encourage smaller turns, facilitate obstacle avoidance, and improve coverage, which further decreases low-velocity motion and unnecessary revisits. In addition, we introduce a velocity-loss constraint to refine local trajectories and promote high-speed flight. We validate the proposed approach in both simulation and real-world experiments, where it consistently improves exploration efficiency over several state-of-the-art meth- ods, especially in large-scale environments.