Abstract: A mass spectrometer experimental instrument robotic arm automatic recognition and detection method based on improved YOLOv5s is proposed to address the issues of improving recognition accuracy and speed, as well as reducing system costs in the recognition technology of grabbing samples for mass spectrometer robotic arms. Firstly, the CA attention mechanism is introduced, which can strengthen the feature learning and feature extraction of the sample tray image, while reducing the impact of the sample surface background on the detection results, especially the background similar to the sample tray. Secondly, a path is added to the original backbone network structure that directly connects the shallow feature map to the deep layer to enhance the model’s detection ability of small target objects; finally, the original C3 module is replaced with the DSConv module to reduce the complexity of the model, and maintain its accuracy. The experimental results show that in the custom dataset, compared to YOLOv5s, the improved model has slightly increased parameter and computational complexity, significantly improved detection accuracy. In the COCO public dataset, the model proposed in this paper outperforms other models with high accuracy of 0.723 mAP, while maintaining a compact weight size of 20.2 MB and a real-time processing speed of 35.81 FPS.