Last-Meter Precision Navigation for UAVs: A Diffusion-Refined Aerial Visual Servoing Approach

2026-07-05

AI Summary

This work addresses the challenges of scale ambiguity, rotational discontinuity, and fine-grained spatial reasoning in monocular vision–based autonomous navigation for unmanned aerial vehicles (UAVs) within the final ten meters. To tackle these issues, the authors propose DreamNav, a coarse-to-fine two-stage framework. It first employs a triangle-parameterized rotation representation to robustly estimate an initial pose, effectively mitigating periodic ambiguities. Subsequently, it leverages a pretrained world model for visual imagination and integrates a diffusion mechanism to refine action trajectories, selecting the path that minimizes visual discrepancy with the target. Evaluated on the newly introduced large-scale PairUAV benchmark—comprising 4.8 million image pairs across 72 scenes—DreamNav substantially outperforms existing visual servoing and foundation model approaches, achieving high-precision navigation and zero-shot generalization across unseen environments.

Abstract

In this work, we study the last-meter precision navigation for UAVs, e.g., autonomously reaching a target within the final 10 meters using monocular vision. This task is challenging due to scale ambiguity, rotation discontinuities, and the need for fine-grained spatial reasoning. Existing methods often fail under large viewpoint changes or lack generalization to unseen environments. To this end, we propose DreamNav, a coarse-to-fine diffusion-refined aerial visual servoing framework. In the first coarse-estimation stage, a robust regression policy employs a trigonometric parameterization to predict rotation by jointly modeling sine and cosine components, effectively mitigating optimization instabilities caused by angular periodicity. Given this coarse estimate, the second diffusion-refined stage utilizes a pre-trained world model to simulate future visual observations for candidate actions, selecting the trajectory that minimizes visual discrepancy with the target through a process of visual imagination. To support rigorous evaluation, we contribute PairUAV, a large-scale benchmark comprising 4.8 million image pairs across 72 scenes, curated from the University-1652 dataset. Extensive experiments show DreamNav outperforms strong visual servoing and foundation model baselines in accuracy and generalization, with zero-shot transfer to unseen scenes.

Problem

• precision navigation
• UAV
• visual servoing
• monocular vision
• last-meter navigation

Innovation

• diffusion-refined visual servoing
• trigonometric rotation parameterization
• visual imagination
• zero-shot generalization
• aerial navigation