In view of the problem found in one-to-many mapping between input and output while dealing with the inverse kinematics multiple solution of robot with neural network, a method has been proposed of multiple module neural network based on geometric sub-region to solve the inverse kinematics multiple solutions of planar 3R manipulators. This method seeks to divide the planar 3R manipulators into two sub-spaces with multiple solutions by separating and reassembling the connecting rods. Meanwhile, the azimuth angle planning of the end effector is to be added, so that there is a unique mapping relationship between input and output in the BP neural network for two subspace training. Two trained BP neural networks are used to predict the same planned circular trajectory respectively, with the predicted trajectories being smooth and almost consistent with the planned trajectory. The results show that this method can obtain high precision inverse kinematics solutions of planar 3R manipulators.