AgenticDiffusion framework achieves 80% success in drone navigation using dual cameras

AgenticDiffusion is a new AI-driven navigation framework for indoor UAVs that addresses the limitations of single-view systems. Built by researchers Faryal Batool, Muhammad Ahsan Mustafa, and colleagues, the framework coordinates language-guided reasoning, open-vocabulary target grounding, vision-based diffusion planning, and Nonlinear Model Predictive Control (NMPC) into a unified pipeline. Given a natural language instruction, it processes synchronized FPV and top-view camera feeds to determine the most informative viewpoint, localize targets using an open-vocabulary grounding model, and then uses viewpoint-specific diffusion planners to generate collision-free trajectories. This dual-camera approach allows the drone to reason about occlusions and scene structure, dramatically cutting redundant exploration.

In rigorous real-world testing across four scenarios—adaptive viewpoint selection, multi-stage missions, long-horizon navigation, and safe landing-site selection—AgenticDiffusion achieved an overall mission success rate of 80% over 40 trials. Notably, the diffusion planners hit a perfect 100% trajectory generation success rate. The framework's ability to leverage complementary viewpoints significantly improves navigation efficiency in cluttered indoor spaces, making it a promising step toward fully autonomous indoor drone operations for tasks like inspection, search-and-rescue, and logistics.