The regular pushing of feed is an essential part of feeding process of dairy cows. Aiming at the problem that the existing feedpushing robots have single function, which cannot collect and transport the feed according to the cows' position to meet their needs, an intelligent feedpushing robot for dairy cows was developed. Firstly, YOLACT instance segmentation model was used to identify cows, feed, and square rod and obtain the mask. Secondly, dynamic objects were removed at ORB-SLAM3 by using the mask to improve the positioning accuracy, and then the real-time robot position was obtained. Thirdly, the location of foraging cows was calculated by combining the mask, stereo camera depth image and the robot position, and the distance between the robot and the cattle barn was calculated by using mask and depth image with the square rod as reference. Finally, during the working process of the robot, the distance between the robot and the cattle barn was kept unchanged, and the independent decisions were made by the robot according to the foraging cows position and the feeding time, so as to realize the multi-mode feeding functions of push, collect-transport and clean, so it can improve the feed utilization efficiency and meet the free feeding needs of cows. The research and experimental results showed that on the TUM RGB-D dataset, compared with ORB-SLAM3, the proposed algorithm can effectively reduce the positioning error in dynamic environments; the foraging cows position calculation accuracy was ±0.1m, and each cow can be recognized; the distance calculation accuracy between the robot and the cattle barn was ±0.8cm; the working mode selection accuracy was 100%; and the algorithm running rate was 12f/s. The robot met the requirements of intelligent feeding of robots in complex environments.