Personalized products have gradually become the main business model and core competencies of many enterprises. Large differences in components and short delivery cycles of such products, however, require industrial robots in cloud manufacturing (CMfg) to be smarter, more responsive and more flexible. This means that the deep learning models (DLMs) for smart robots should have the performance of real-time response, optimization, adaptability, dynamism, and multimodal data fusion. To satisfy these typical demands, a cloud-edge-device collaboration framework of CMfg is first proposed to support smart collaborative decision-making for smart robots. Meanwhile, in this context, different deployment and update mechanisms of DLMs for smart robots are analyzed in detail, aiming to support rapid response and high-performance decision-making by considering the factors of data sources, data processing location, offline/online learning, data sharing and the life cycle of DLMs. In addition, related key technologies are presented to provide references for technical research directions in this field.