ViT-CoRRT*for Multi-UAV Path Planning with Path Compression and Sequential Dispatching

Multi-UAV path planning in complex 3D urban environments is affected by large search spaces, obstacle constraints, redundant waypoints, and inter-UAV spatiotemporal conflicts. To address these issues, this paper proposes ViT-CoRRT*, a planning framework that integrates start-goal-conditioned ViT-guided RRT* search, dynamic programmingbased path compression, and sequential dispatching. The 3D Vision Transformer generates a task-dependent guidance map from the voxelized environment and uses it as a probabilistic sampling prior for RRT*. The generated paths are then compressed by selecting collision-checked shortcuts between non-adjacent waypoints, after which UAV departure times are adjusted according to spatiotemporal conflict checks. Across 50 baseline trials, ViT-CoRRT* reduced the mean total path length from 113.72 to 92.99 and the mean number of retained waypoints from 46.38 to 13.80 compared with standard RRT*. The mean task makespan also decreased from 35.00 to 28.33. The differences in these three metrics are statistically significant (p < 0.001). Additional experiments under different obstacle configurations and UAV fleet sizes achieved solution success rates of 98%–100%. No post-dispatch conflicts were detected in successful trials under the adopted discrete-time checking resolution.