Cleft lip and palate is the most common congenital malformation in oral and maxillofacial region. As a kind of facial plastic surgery, the most important factor for the success of cleft lip and palate repair surgery is the design of surgical markers and incisions. However, general hospitals especially in rural areas lack dependable medical resources, which makes the effect of the surgery hard to guarantee. To solve this problem, we propose a novel robotic surgery assistant technology based on deep learning to help reduce the technical threshold and improve the overall effect of cleft lip and palate repair surgery. For the first time, a robust dataset of cleft lip and palate cases is established, which can be used to train the model to locate surgical markers and incisions. Secondly, we build a strong baseline on this dataset by using state-of-the-art Hourglass architecture and residual learning, with two neoteric block designs, one of which enables stronger capability of generalization, while the other greatly reduces the complexity of the model, thus making efficient application possible. Finally, by comparing with other facial feature extraction methods, our models achieve the best results on multiple metrics, showing their strong superiority and adaptability on this task.