Global navigation satellite system for localization of agricultural machineries has been successfully used for many years in a broadacre environment. However, it will fail to work in a close planting orchard because navigation satellite signals are often interfered by dense tree canopies. Moreover, localization by only odometers can lead to a large accumulated error. Therefore, firstly the position of an agricultural robot before entering interrow of the close planting orchard was taken as an original point and a world coordinate system was established. A laser radar was applied to scan trunks on the both sides of the robot and the center points of trunks were extracted through circular clustering. The position values of these center points were saved in the world coordinate system. Then during the moving of the robot, the center points of trunks detected by the laser radar in real time were matched with those which had been saved previously in the world coordinate system. Matching results were applied to correct the robot position and heading angle which had been measured with only odometers. In this way the agriculture robot was able to locate itself accurately in the interrow of the close planting orchard. Repeated experimental results showed that the robot localization standard errors in both x and y directions in the world coordinate system were about 0.08m. The localization accuracy can meet requirements when the agricultural robot is applied in the close planting orchard. Moreover, the successful interrow localization is the premise of global positioning and navigation of the robot working in the orchard.