In practical applications, single-sensor SLAM algorithms struggle to adapt to complex and dynamic environments, often leading to a heavy reliance on the hardware precision of that single sensor for the final positioning result. Conversely, multi-sensor fusion SLAM algorithms offer superior positioning accuracy and robustness, playing a significant role across various domains. However, existing multi-sensor fusion positioning technologies lack environmental awareness. This paper proposes a dynamic environment adaptive fusion localization method based on reinforcement learning, integrating Extended Kalman Filters (EKF) to construct an environmental perception model. It determines the credibility of different sensor data based on environmental context and dynamically adjusts observations accordingly. The objective is to infer fusion strategies based on experience or physical models and establish confidence for each sensor data. Ultimately, this method provides effective fusion strategies for multimodal sensor data. The paper designs and constructs a deep reinforcement learning model to identify environmental factors, determine key influencing factors, and make optimal decisions for multimodal sensor fusion localization. These decisions guide multimodal sensor fusion, significantly improving accuracy. Experimental results demonstrate a 38% increase in accuracy compared to other methods using the same sensor combination on public dataset simulations.