Autonomous unmanned aerial vehicles (UAVs) promise transformative advances in logistics, inspection, and emergency response if they can reliably navigate cluttered, dynamic airspace. We review peer-reviewed studies in deep reinforcement learning (DRL), a subfield of reinforcement learning (RL), for UAV navigation and path planning and organize them into six themes: single-UAV control, simulation-to-real transfer, safe and constraint-aware learning, multi-agent swarms, model-based and Model Predictive Control (MPC) hybrids, and vision-centric methods. Across these themes, single-aircraft success rates rose from ~ 75% in early simulators to task completion on hardware. In particular, end-to-end actor–critic methods, especially proximal policy optimization (PPO), have become the default for high-dimensional control. Meanwhile, safety filters, notably control-barrier-function (CBF) shields, reduce indoor collisions, though they lengthen training. Graph attention-based multi-agent reinforcement learning (MARL) scales to 225 simulated UAVs, yet outdoor deployment still struggles with wind, onboard compute limits, and decentralized communication. We present a unified taxonomy Figures, tables, textboxes, or e-components should not be cited in the article abstract; hence, this citation has been removed. Please check the changes made if appropriate. and a cross-benchmark comparison and conclude by outlining five open challenges: verifiable safety, robust sim-to-real transfer, scalable coordination, compute-efficient learning, and vision-based fault recovery.