Objective To improve the accuracy and robustness of the robotic arm food sorting system in sorting tasks, a new sorting method combining machine vision and force sensing is studied.Methods On the basis of the food sorting system, a mechanical arm food sorting method integrating an improved YOLOX model and improved admittance control is proposed. By introducing a convolutional attention mechanism module (CBAM) to focus on target features, adopting a deep separable convolutional optimization network structure, and combining with a new backbone network, the recognition and localization ability of the YOLOX model for food targets is enhanced. Meanwhile, by utilizing an active compliant control method based on improved admittance, stable sorting of different foods can be achieved. An experimental platform is built to evaluate he application effect of the proposed method.Results The proposed improved YOLOX model improves the detection accuracy of food targets to over 99%, which is over 3% higher than that of conventional methods. In addition, the system demonstrates significantly enhanced robustness in various fragile food sorting tasks, with a sorting success rate increasing by over 5%.Conclusion The proposed method effectively improves the accuracy and robustness of robotic arm food sorting, demonstrating promising application prospects.