Existing push robots mainly use magnetic induction technology. These devices are susceptible to external electromagnetic interference and have a low degree of intelligence. To make up for the insufficiency of the existing material pushing robots, and at the same time solve the problems of labor-intensive, labor-intensive, and inability to push material in time at night, etc., in this study, an autonomous navigation pusher robot based on 3D lidar is designed, and an obstacle avoidance strategy based on the improved artificial potential field method is proposed. Firstly, the 3D point cloud data of the barn is collected by the self-designed pushing robot, the point cloud data of the area of interest is extracted using a direct-pass filtering algorithm, and the 3D point cloud of the barn is segmented using a height threshold. Secondly, the Least-Squares Method (LSM) and Random Sample Consensus (RANSAC) were used to extract fence lines, and then the boundary contour features were extracted by projection onto the ground. Finally, a target influence factor is added to the repulsive potential field function to determine the principle of optimal selection of the parameters of the improved artificial potential field method and the repulsive direction, and to clarify the optimal obstacle avoidance strategy for the pusher robot. It can verify the obstacle avoidance effect of the improved algorithm. The experimental results showed that under three different environments: no noise, Gaussian noise, and artificial noise, the fence lines were extracted using RANSAC. Taking the change in the slope as an indicator, the obtained results were about −0.058, 0.058, and −0.061, respectively. The slope obtained by the RANSAC method has less variation compared to the no-noise group. Compared with LSM, the extraction results did not change significantly, indicating that RANSAC has a certain resistance to various noises, but RANSAC performs better in extraction effect and real-time performance. The simulation and actual test results show that the improved artificial potential field method can select reasonable parameters and repulsive force directions. The optimized path increases the shortest distance of the obstacle point cloud from the navigation path from 0.18 to 0.41 m, where the average time is 0.059 s, and the standard deviation is 0.007 s. This shows that the optimization method can optimize the path in real time to avoid obstacles, basically meet the requirements of security and real-time performance, and effectively avoid the local minimum problem. This research will provide corresponding technical references for pusher robots to overcome the problems existing in the process of autonomous navigation and pushing operation in complex open scenarios.

The pushing robot working in the complex farming environment encounters several problems. For example, the precision of its navigation path extraction is low, and its working quality is greatly affected by the weather. In view of this, a method of farm operation path extraction based on machine vision is proposed in this study in order to solve the problems above and realize the autonomous and intelligent operation of the robot. First of all, the RGB images of the working area in front of the robot are obtained by using an RGB camera installed on the machine. Then, the collected images are preprocessed by means of sky removal, denoising and grayscale transformation. After that, the image is segmented to obtain the front fence, feed belt and ground data. Finally, the navigation path is obtained by extracting the features of the feed belt. The test results show that the absolute deviation of the pushing robot at different initial lateral distances is less than ±15 cm, and the deviation between the actual navigation route and the target route is within the expected range. The absolute value of the maximum lateral deviation in five test areas is 8.9 cm, and the absolute value of the average maximum lateral deviation is 7.6 cm. These experimental results show that the pushing robot can work stably without disturbing the feeding of cows. Particle swarm optimization is used to optimize the parameters of the PID and find the optimal parameters. This makes the system balanced and more responsive. Through this test, it is found that the initial direction of the robot will have a certain impact on the path production and tracking efficiency, and this effect is more significant when the robot changes the working area or turns. In which case, the trajectory of the robot should be in such a way that it immediately faces the next row at a small angular deviation, thus ensuring smoother motion. The method proposed in this study can provide support for the automatic navigation of pushing robots in dairy farms.