Analytical Assessment of Deep Reinforcement Learning for Autonomous UAV Path Planning in Congested Nigerian Urban Airspace

The deployment of commercial and public safety drones in Nigerian cities faces significant path planning challenges in congested low-altitude urban airspace, and deep reinforcement learning represents a promising approach for autonomous collision-avoidance navigation, but no empirical assessment exists for Nigerian urban airspace topology. This study analytically assessed DRL-based path planning for autonomous UAV operations in a simulated representation of Lagos Island low-altitude airspace. A 3D urban environment model was constructed in AirSim simulator using building footprint data extracted from OpenStreetMap for Lagos Island. Twin Delayed Deep Deterministic Policy Gradient and Soft Actor-Critic algorithms were trained as continuous-action DRL agents for 3D UAV navigation with obstacle avoidance, targeting a delivery drone mission profile (take-off, waypoint navigation, landing). Rewards were defined to balance mission completion speed, obstacle clearance distance, and energy efficiency. TD3 achieved 89.4 percent mission success rate after 2 million training steps, compared to 85.7 percent for SAC under equivalent training budget. Mean path length efficiency (ratio of actual path to straight-line distance) was 1.24 for TD3 and 1.28 for SAC. Both algorithms significantly outperformed A* graph search (63.2 percent success in dynamic obstacle environments) due to their ability to handle unplanned moving obstacles. Transfer of TD3 policy from AirSim to a physical DJI F450 quadcopter in a controlled outdoor environment showed an 18 percent success rate degradation, identifying simulation-to-real transfer as the primary research gap requiring further investigation.

Keywords: deep reinforcement learning, UAV path planning, urban airspace, Lagos, autonomous navigation