In robotic grasp detection tasks, the difference in information properties between depth images and color images makes it difficult to fully fuse them, which affects the improvement of grasp detection accuracy. To address this issue, a grasp detection network based on channel-spatial adaptive fusion strategy is proposed. The network employs a dual-branch Yolo v5 backbone as the feature extraction module to separately extract features from color images and depth images. To fully fuse the information from both modalities during the feature extraction process, a channel-spatial multi-stage fusion strategy is proposed. This strategy combines channel fusion and spatial fusion to perform two-stage feature fusion, solving the problem of fusion difficulty due to attribute differences between the two modalities. Furthermore, to capture the differences among samples during the fusion process, a sample adaptive attention module is proposed, allowing the fusion strategy to dynamically adjust the fusion process based on sample differences. Experimental results show that the proposed network achieves accuracies of 98.0% and 95.6% on the Cornell and Jacquard datasets, respectively, showing improvements compared to baseline networks. In physical robot experiments, the network is tested in both multi-object and single-object scenarios, achieving a grasp success rate of 94.5%, which validates the practical value of the proposed network.