Due to the complexity and variety of module robot, the classic planning method can hardly give out a general algorithm that fits most combination. In order to maintain the advantage of module robot’s rapid deployment and avoid the problem of complex solution, this essay give out an automatic solution to easily generate certain combination from text description of a module robot combination. During the setup phase of the physical environment, different environmental parameters are studied and considered for their impact on the planning methods obtained by the algorithm, comparing the adaptability of different planning strategies to the environment, then start to train the combination to do various tasks like moving and posing with fixed hyper-parameter, the research focuses on designing observation functions and reward functions as evaluation mechanisms and using reinforcement learning to obtain corresponding planning models. After training complete under Isaac Gym, the network was stored in a library for fast fetch. By applying this method to minimize the time consumption for locomotion planning, managed to accomplish just-in-time planning for each module robot combination that has been given.