<p>Recent years have witnessed a significant increase in demand for robots capable of autonomously performing object navigation tasks in cluttered indoor environments. Existing approaches to object navigation either focus solely on navigating to a single target object or adhere to a predefined sequence when multiple targets are involved. Both approaches often result in inefficient paths and low success rates. To overcome these limitations, we propose a modular framework for efficient Multi-object navigation that 1) dynamically optimizes the target visitation order and 2) adaptively adjusts path-planning costs based on the current exploration context. First, a pretrained semantic map-based inference network predicts spatial distributions of potential target locations. These predicted distributions are then combined with geodesic distances to calculate target values, enabling the agent to dynamically select the most beneficial next target. Navigation toward the chosen goal is executed by a Fast Marching Method planner operating on a cost map, whose costs are adaptively updated based on predicted probabilities and visitation history, thereby balancing exploration and efficient navigation. Experiments conducted on the MP3D and HM3D datasets demonstrate that the proposed method significantly improves success rates and path efficiency compared to fixed-order baselines, showcasing its practical applicability to real-world robotic missions.</p>