In order to achieve rapid recognition and selection of stacked parts by robots in complex industrial environments,an improved YOLOv8s instance segmentation model was constructed and applied to real-time recognition and segmentation of stacked parts.To address the issue of difficult segmentation of stacked industrial parts,the original model′s backbone network was replaced with a PoolFormer backbone network with stronger feature extraction capabilities to improve the edge segmentation effect of stacked parts;in order to better filter out excess background information and retain key information,a better CARAFE upsampling module was introduced.The experimental results show that the average segmentation accuracy and prediction box accuracy of the improved model are 93.57% and 97.47%,respectively,which are 1.89% and 1.23% higher than the original model,and far higher than the YOLACT++and SOLOv2 models with the same type,verifying the effectiveness of the improved model.