Towards the zero-gravity simulation test requirements of spacecraft for on-orbit service missions, the zero-gravity motion simulation technology based on industrial robot was studied. In order to realize the accurate force sensing of load on robot, the BP neural network was used to establish the force prediction model. The model uses the robot pose, acceleration, angular velocity, and angular acceleration as input layer parameters, and uses the data from force/torque sensor on robot wrist as the output layer parameter. In order to realize the accurate force perception in whole working space of the robot, the orthogonal experiment design method was adopted to determine the path points of the robot for sample data collection. Based on the established force prediction model, the accurate force sensing of the end load of the robot under moving conditions is realized. In experiment, for the load with 100kg mass on robot, the maximum error of force sensing is 39.8N, and the root mean square error of force sensing is 9.3N. The maximum error of torque sensing is 18.7Nm, and the root mean square error of torque sensing is 4.4Nm. Furthermore, according to the real-time force sensing results of the robot load, the motion speed of the load in zero-gravity state is calculated by dynamics theory. Then the robot is driven to execute the corresponding motion of the load, and the zero gravity motion simulation of the load is realized.