Perceptual Aliasing and Uncertainty in Repetitive Environments for Monocular UAV Navigation Using Perception-Aware Reinforcement Learning

Abstract

Robust aerial navigation in complex environments remains a central challenge for autonomous uncrewed aerial vehicles (UAVs), particularly under GPS-denied conditions. Many real-world environments, including plantations, forests, orchards, and industrial facilities, exhibit repetitive spatial structures that complicate visual perception. Under such conditions, similar observations may correspond to different physical locations, leading to perceptual aliasing and degraded navigation stability for monocular systems. Existing approaches remain insufficient in handling perceptual aliasing, monocular depth uncertainty, and temporal belief uncertainty under partial observability, as perception outputs are often treated deterministically and are not integrated into reinforcement learning. To address these challenges, this paper proposes PEARL-Nav (Perception-Enhanced and Uncertainty-Aware Reinforcement Learning for Navigation), a unified framework that incorporates perception reliability into decision-making. The architecture combines a convolutional neural network (CNN) for local feature extraction, a Vision Transformer with Hierarchical Aliasing-Aware Attention and Uncertainty Quantification (HAA-UQ) for modeling structural ambiguity, and a Long Short-Term Memory (LSTM) module for temporal belief aggregation. Furthermore, the Physics-Informed Cross-Modal Depth Refinement (PI-CMDR) module enhances monocular depth estimation using motion-consistent geometric constraints, while a Learned Perceptual Curriculum (LPC) and Perception-Guided Episode Truncation (PGET) improve training stability under perception uncertainty. Experiments show that PEARL-Nav achieves a navigation success rate of up to 96.1%, reduces monocular depth error on average by approximately 13% in Absolute Relative error, 10% in RMSE, and 25% in scale consistency error, and requires approximately 30% fewer training episodes, outperforming state-of-the-art methods and confirming the value of modeling perceptual uncertainty for robust navigation in repetitive environments. The source code is available at https://github.com/Aqshalikhsan/PEARL-Navigation