Title: AgenticDiffusion: Leveraging Agentic Diffusion for Vision-Centric UAV Path Planning

Abstract: Navigating Unmanned Aerial Vehicles (UAVs) indoors demands robust capabilities in scene comprehension, efficient exploration, and precise trajectory control, particularly when constrained by a restricted field of view. Current vision-based navigation systems largely depend on single-view inputs, which hinders their capacity to interpret occlusions, assess target visibility, and grasp the broader spatial layout. To address these limitations, this study introduces AgenticDiffusion, a comprehensive multi-view navigation architecture. This framework integrates language-driven reasoning, open-vocabulary target localization, vision-based diffusion planning, and Nonlinear Model Predictive Control (NMPC) into a cohesive aerial navigation pipeline.

By processing synchronized first-person-view (FPV) and top-view imagery alongside natural language commands, the system identifies the most advantageous viewpoints for navigation and formulates a mission strategy before executing any movement. Specifically, an open-vocabulary grounding model pinpoints targets, enabling viewpoint-specific diffusion planners to construct navigation trajectories for the UAV. By leveraging the synergies of complementary camera perspectives, AgenticDiffusion minimizes redundant target searches and enhances overall navigation efficiency within complex, cluttered indoor spaces.

The proposed framework underwent validation across four distinct real-world UAV navigation tests, covering adaptive viewpoint selection, multi-stage mission execution, long-horizon navigation, and the identification of safe landing zones. Experimental data from 40 real-world trials revealed an overall mission success rate of 80%, while the diffusion planners demonstrated a perfect 100% success rate in trajectory generation.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC