AI in Autonomous Drone Systems

Abstract

Autonomous drone systems -- unmanned aerial vehicles that perceive, navigate, plan, and execute tasks without continuous human control -- are transitioning from research prototypes to operational deployments in inspection, agriculture, logistics, search and rescue, and environmental monitoring. AI is central to each capability layer: perception (detecting obstacles, identifying targets), navigation (path planning, obstacle avoidance, SLAM), mission planning (task decomposition and execution), and fleet coordination (multi-drone task assignment and collision avoidance). This study evaluates nine AI systems for autonomous drone operations across five task domains: indoor obstacle avoidance, outdoor GPS-denied navigation, agricultural crop inspection, infrastructure inspection, and multi-drone coordination. Systems evaluated include classical path planners with learned obstacle avoidance (RRT-NN), deep RL navigation (PPO-UAV, DreamerV3-UAV), vision-based SLAM (ORB-SLAM3-DL), semantic segmentation for target detection (Mask2Former-UAV), LLM-guided mission planning (MissionGPT), multi-drone coordination (MARL-Drone), a proposed Unified Drone AI System (UDAIS), and a safety monitor framework (DroneGuard). UDAIS achieves the highest task completion rate across all five domains (78.4%). DreamerV3-UAV achieves the best GPS-denied navigation success (84.2%). MissionGPT reduces mission planning time by 42.4% vs. human operators for complex multi-step tasks. DroneGuard achieves zero unsafe incidents during 2,840 flight hours across all tested systems. EU airspace regulation (U-Space) compliance and a certification framework for autonomous drone AI are proposed.