By driving the joint to reduce the drag force of the robot，it is an effective way to improve the flexibility of the robot．In the process of dragging teaching，the accurate and real-time calculation of the torque compensation of each joint of the robot is the key problem to realize the dragging assistance．The dynamic modeling of the dragging teaching spraying robot was carried out，and the relationship between joint torque compensation and inertia force，gravity and other factors was analyzed，and a BP neural network torque control algorithm based on unsupervised learning was proposed to compensate the direct teaching of the robot．Finally，experimental verification was carried out on a ６DOF spraying robot．The experimental results show that the calculation efficiency of the torque compensation algorithm is increased by 70%，the average calculation error is 9%，and the assistance effect is obvious．