We propose a transfer learning method to estimate the motion time of a 6 -axis robot arm obtained by 3D simulation using a machine learning method. The model is constructed from two perspectives: one without obstacles and one with obstacles. The motion planning method is Rapidly exploring Random Trees Star (RRT^*), where the initial and target postures of the robot are set randomly within its range of motion, and the resulting motion times are used for learning. The model in an environment without obstacles can be constructed with high speed and high accuracy. We develop a transfer learning method for a heterogeneous 6 -axis robot arm to verify the speed-up of model construction. For the machine learning model with obstacles, the problem setup was extended to adapt to obstacles based on the former model. Computational results show that both the accuracy and reliability of the model were improved by the proposed model.