Learning to Explore A Curiosity Aware Zero-Shot Framework for UAV Navigation in Indoor Environments

Abstract

Unmanned Aerial Vehicles (UAVs) require robust exploration strategies to operate effectively in unknown indoor environments. Traditional methods often rely on prior training data or environment-specific models, limiting their adaptability in novel scenarios. In this paper, we propose a Curiosity-Aware Zero-Shot Framework that integrates an Intrinsic Curiosity Module (ICM) with a domain-randomized Zero-Shot planner to enable efficient and autonomous UAV exploration without retraining. Our framework is trained in simulated environments with randomized layouts to promote generalization, and eval- uated in unseen 3D indoor scenes. Experimental results show that our method significantly outperforms baselines such as Random Walk, Greedy Frontier, ICM-only, and Zero-Shot-only planners, achieving 89.7% coverage, 1.6 path efficiency, 328 seconds exploration time, and a 94.5% success rate. The ablation study highlights the complementary role of both ICM and Zero- Shot components. This work presents a scalable solution for real-time UAV navigation and contributes to the development of intelligent aerial systems capable of learning to explore novel environments autonomously.