Vision-Guided Control Using Proximal Policy Optimization for Safe Low-Altitude UAV Navigation

Autonomous operation of Unmanned Aerial Vehicles (UAVs) in complex low-altitude environments requires navigation strategies that ensure safety, adaptability, and reliability under dense obstacle conditions. This paper proposes an end-to-end intelligent navigation framework that integrates a real-time object detection network (YOLO) with an adaptive control policy based on Proximal Policy Optimization (PPO) to achieve robust, collision-free obstacle avoidance. The approach is validated in a simulated low-altitude 10×10×10 m 3 three-dimensional workspace containing 25 static obstacles arranged in a deterministicstochastic distribution, representing highly cluttered operational scenarios. Perception inputs derived from YOLO detections are transformed into a structured observation vector and used to guide continuous control actions learned via a shaped reward function that accounts for goal convergence, altitude stability, motion efficiency, and safe obstacle standoff distances. To better reflect realistic sensing conditions, perceptual uncertainty is modeled by injecting Gaussian noise and false-positive detections into the perception pipeline during training. Experimental results over 500,000 training timesteps demonstrate stable policy convergence, achieving a 90% success rate with zero collisions across multiple test episodes and maintaining consistent performance under varying noise levels. The resulting trajectories show improved safety-aware maneuvering and reduced path length, highlighting the frameworks potential for low-altitude engineering applications requiring dependable autonomous navigation under imperfect perception.