AgenticRL: Self-Refining Agentic Reinforcement Learning for Vision-Conditioned UAV Navigation

The paper presents AgenticRL, a novel reinforcement learning framework designed for UAV navigation that enhances autonomy in reward design and policy refinement. Utilizing a multimodal generative pre-trained transformer (GPT) agent, AgenticRL generates task-specific rewards and employs Proximal Policy Optimization (PPO) for policy training, achieving a 71% improvement in policy behavior through a closed-loop self-improvement process. The framework demonstrates robust performance with a real-world success rate of 91% and sim-to-real accuracy of 94%, making it significant for practitioners seeking to reduce manual tuning in autonomous navigation tasks.