Abstract: Existing robot obstacle avoidance methods mostly rely on a single sensor, which leads to large obstacle calibration errors and insufficient safety margins in complex unstructured roadway environments where dynamic obstacles appear randomly. To address these problems, a dynamic obstacle avoidance method for tracked robots based on multi-sensor perception and deep reinforcement learning was proposed for narrow unstructured coal mine roadways. The high-resolution imaging capability in the visible spectrum and the sensitivity to thermal radiation in the infrared band were used to perceive roadway environments with low illumination and high reflectivity. The Mean Shift algorithm was introduced to perform kernel density estimation on the occurrence probability of obstacles in the roadway, and the three-dimensional spatial coordinates of obstacles were calibrated to overcome the limited field of view and occlusion caused by the narrow roadways. The spherical envelope method was used to construct the safety potential field boundary corresponding to the three-dimensional spatial coordinates of obstacles as the constraint condition for the obstacle avoidance reward in deep reinforcement learning, and the robot obstacle avoidance behavior was optimized according to the reward to achieve dynamic obstacle avoidance. Experimental results showed that under conditions of high dust concentration, strong light, and weak light, the mean of cross-modal structural similarity between the visible light images and the infrared images of the perceived results was higher than 55%, enabling accurate perception of the roadway environment. The maximum error between the calibrated obstacle position and the actual position was only 0.4 m. During movement, the minimum distance between the robot using the proposed method and obstacles was greater than the safety threshold, and no collision occurred, indicating a sufficient safety margin for obstacle avoidance.