<p>Magnetic microrobots, given their unique characteristics, hold great potential in biomedical applications such as targeted therapy and microscale operations and are receiving widespread attention. Research on the autonomous navigation of magnetic microrobots is highly focused, as it is an essential prerequisite for achieving functions such as targeted delivery in medical settings. The success of autonomous navigation determines the level of intelligence and precision in the motion of magnetic microrobots. However, uncertainties stemming from environmental changes and time-varying disturbances in electromagnetic systems adversely affect the control accuracy of magnetic microrobots. Additionally, the random appearance of dynamic obstacles along expected trajectories challenges their autonomous navigation. In this study, we demonstrate a method for the exact autonomous navigation of magnetic microrobots in fluid environments, successfully avoiding dynamic obstacles that suddenly appear in predefined trajectories. Improved versions of the A* algorithm and dynamic window approach are integrated as path planners, that can generate smooth and collision-free trajectories that adhere to kinematic constraints in fluid environments with dynamic obstacles. A learning-based model predictive control strategy is employed, where radial basis function neural networks are used to effectively predict and compensate for fluid disturbances and inevitable errors introduced by electromagnetic system coupling, thereby ensuring the control accuracy of the magnetic microrobot in a flowing environment. Experiments in a constructed microfluidic environment validate the effectiveness of our navigation approach in motion control, autonomous navigation, and replanning, with an average error of less than 8% of the body length of the microrobot.</p>