Deep Reinforcement Learning for Autonomous Navigation in GPS-Denied Environments: A Comprehensive Review

Abstract

Autonomous navigation in GPS-denied environments represents a critical challenge for unmanned systems operating in contested, indoor, underground, and extraterrestrial domains. This comprehensive review examines the state-of-the-art deep reinforcement learning (DRL) approaches that enable robust navigation without reliance on global positioning systems. The paper systematically analyzes the evolution of DRL algorithms from value-based methods (DQN, Rainbow) to policy gradient approaches (PPO, SAC, TD3) and their applications in vision-based, LiDAR-based, and multi-modal navigation frameworks. We critically evaluate the performance of model-free versus model-based DRL techniques across diverse operational scenarios including urban canyons, dense forests, subsurface environments, and planetary exploration. The review identifies key architectural innovations such as attention mechanisms, recurrent neural networks for temporal reasoning, and hierarchical reinforcement learning structures that enhance navigation robustness. Particular emphasis is placed on sample efficiency challenges, sim-to-real transfer gaps, and safety guarantees during learning and deployment phases. We analyze benchmark datasets, simulation environments, and evaluation metrics used in the literature, highlighting inconsistencies that hinder comparative assessment. The paper also examines hybrid approaches combining DRL with classical path planning, SLAM integration strategies, and multi-agent coordination in GPS-denied scenarios. Emerging trends including meta-learning for rapid adaptation, curiosity-driven exploration, and physics-informed neural networks are discussed as promising directions. This review concludes by identifying critical research gaps including generalization across environmental conditions, computational efficiency for embedded deployment, and formal verification methods for safety-critical applications, providing a roadmap for future investigations in DRL-based autonomous navigation.