Abstract: The environment of coal mine tunnels is narrow and complex, containing a large number of unstructured obstacles, resulting in multiple uncertainties in the target perception calibration decision chain. It is difficult for tracked robots to achieve balanced and high-precision obstacle avoidance. Therefore, a dynamic obstacle avoidance method for tracked robots in narrow and unstructured coal mine tunnels is proposed. By utilizing high-resolution imaging in the visible light band and thermal radiation sensitivity in the infrared band, multimodal perception of areas with abnormal brightness and hidden targets due to thermal radiation can be achieved; In response to obstacles in narrow and unstructured coal mine tunnels, the Mean Shift algorithm is introduced to estimate the kernel density of the calibration object. By iteratively converging to the pose model points of the probability density function, the three-dimensional spatial coordinates of the brightness anomaly area and thermal radiation hidden targets are solved to alleviate multiple uncertainties in space; Considering the relative position relationship of pose model points, combined with the sphere envelope method, a safe potential field boundary corresponding to three-dimensional spatial coordinates is constructed. The actuator pose optimization reward, ontology collision avoidance reward, and task termination reward are repeatedly optimized to generate and update obstacle avoidance strategies. The experimental results show that the fluctuation of cross modal structural similarity mean is relatively balanced in high dust concentration environments or strong and weak light environments, and the obstacle calibration accuracy is higher. The minimum distance is always greater than the safety threshold, and the ability to maintain safety margin throughout the entire obstacle avoidance task is excellent.