Abstract
Adverse weather severely impairs real-world visual perception, while existingvision models trained on synthetic data with fixed parameters struggle togeneralize to complex degradations. To address this, we first constructHFLS-Weather, a physics-driven, high-fidelity dataset that simulates diverseweather phenomena, and then design a dual-level reinforcement learningframework initialized with HFLS-Weather for cold-start training. Within thisframework, at the local level, weather-specific restoration models are refinedthrough perturbation-driven image quality optimization, enabling reward-basedlearning without paired supervision; at the global level, a meta-controllerdynamically orchestrates model selection and execution order according to scenedegradation. This framework enables continuous adaptation to real-worldconditions and achieves state-of-the-art performance across a wide range ofadverse weather scenarios. Code is available athttps://github.com/xxclfy/AgentRL-Real-Weather