Abstract: Abstract: In Greenhouses, the management of fruits and vegetable production involves many tasks, such as picking, spraying, cutting and so on. However, the agricultural robots currently available mainly focus on one specific task. This research focused on a new multifunctional agricultural robot with three arms. Each of its robotic arms has its own function for a different end use. The robot not only can complete one task by collaboration of two different robotic arms but also can fulfill spraying, cutting and picking tasks in the same area. So by using such robot, output productivity can be improved effectively and the cost would be reduced. The robot comprises caterpillar chassis, height-adjustable torso, mechanical structure with three arms, machine-vision system and control system. When at the state of working, the robot uses its machine-vision system to recognize targets. At the moment when targets are found, the robot adjusts its position in order to do tasks correspondingly. Setting up kinematics model of the agricultural robot is the precondition of realizing the system multi-functions. In this paper, we set up the link coordinate system of agricultural robot based on D-H approach. Homogeneous transforming matrixes of adjacent link were established according to the linkage parameters and joint variables. Kinematics equations of visual system, cutting end effector, picking end effector and spraying end effector were deducted and solved, respectively. So the relationship between the position and posture of end effects and linkage parameters and joint variables were determined. The closed-form inverse kinematics was presented by algebraic method and the value of joint variables was calculated according to the certain position and posture of end effects. Kinematics experiments were carried out to verify the correctness of kinematics algorithm and accuracy in application. Matlab software was used to help the calculation of the enveloping space of end effects in the mode of picking and spraying. A series of 12 points already known were picked up as test samples. In the test platform, end effect was moved to location one by one in the given order by driving motors and real coordinate data was collected. The results showed that the robot could carry out specific action under the guidance of kinematics model and the maximum error was only 8 mm. These researches lay foundation for further study of dynamic performance, trajectory planning and accurate motion control of multifunctional agricultural robot.