Abstract: To solve the problems such as low inspection efficiency and low recognition accuracy existing in traditional underground coal mine inspection robots, improve the accuracy and efficiency of underground coal mine target detection, and ensure mine safety, a target detection method for underground coal mine inspection robots based on the improved YOLOv8 is studied and proposed. Based on YOLOv8, aiming at the complex environmental characteristics such as uneven illumination, dust and fog interference, multi-scale targets and occlusion in underground coal mines, the convolutional block attention module (CBAM) is introduced. The model is optimized by CBAM and weighted bidirectional feature pyramid network (BiFPN). The CBAM attention mechanism is embedded in the neck network of YOLOv8. At the output end of the upper sampling layer, the end of the feature stitching operation, and the multi-scale feature fusion stage, dual attention weighting in the channel and spatial dimensions is implemented on the feature maps to enhance the ability of model to capture key features. The original path aggregation network (PANet) structure is replaced by the BiFPN structure. Through the bottom-up and top-down bidirectional feature flow and dynamic weighted feature aggregation, the fusion efficiency of multi-scale features is enhanced, and the detection ability of the model for small targets and occlusions is improved. Furthermore, a spatio-temporal adaptive feature enhancement mechanism is constructed, a gated cross-layer attention transfer module is designed, a learnable dual-domain gating coefficient is introduced to achieve the dynamic coupling of attention weights and feature pyramids, and the feature fusion evaluation criterion based on energy entropy is used to suppress the interference of low-information content features. Meanwhile, a lightweight reasoning framework is designed. The backbone network is deployed in the edge computing unit, and the neck and head networks are deployed in the robot embedded platform. The reasoning delay is reduced through dynamic pipeline parallelism. The research uses the multimodal data set from the working faces of typical coal mines in the main coal-producing areas of Shanxi, Shaanxi and Inner Mongolia for experiments. This data set covers core scenarios such as underground roadway excavation, comprehensive mining equipment inspection, and personnel safety monitoring, and follows relevant environmental parameter standards, which can represent the actual detection difficulties. The results show that the improved YOLOv8 model performs excellently in both detection accuracy and efficiency. Its accuracy rate, recall rate and F₁ value reached 96.1%, 93.4% and 95.8% respectively, which were significantly higher than those of comparison models such as YOLOv8 and YOLOv5. The average detection time is 10.55 ms, which is superior to other comparison algorithms. Under different environmental conditions, when there is no occlusion and normal brightness, the detection time of a single target is as low as 2.4 ms, and the missed detection frequency is only 0.5%. The longest detection time is 18.7 ms when there is severe occlusion and low brightness, and the highest missed detection frequency is 4.0%. In multi-object detection, the detection time is 3.4 ms when the number of objects is 5, and 24.1 ms when the number of objects reaches 25. In terms of positioning accuracy, the average error between the actual coordinates and the positioning coordinates is within 2 mm, meeting the inspection requirements in underground coal mines.