Robot motion planning involves finding a feasible path for a robot to follow while satisfying a set of constraints and optimizing an objective function. This problem is critical for enabling robots to navigate and perform tasks in realworld environments. However, it can become challenging, especially when the environment is cluttered or uncertain, or when the task requirements are complex. A range of approaches have been developed to tackle the robot motion planning problem. Mixed Integer Programming (MIP) is one of them. MIP is an optimization technique consisting of optimising an objective function subject to constraints, where some of the variables are integer valued. This framework is well-suited for finding feasible paths for robots because it allows to incorporate logical constraints such as avoiding obstacles and also continuous constraints such as robot’s dynamics. However, it can be computationally intensive and may not scale well to large or complex environments. To address these limitations, we propose a two-pronged approach. First, we explore the use of machine learning techniques to learn a part of the solution, in order to reduce the size of the MIP problem and improve its scalability. We condlude that state-of-the-art approaches are not directly applicable in a realistic robot motion planning setup. Second, we adopt a receding-horizon approach, in which the global problem is broken down into a series of local problems, each of which can be solved more efficiently. Machine learning techniques are used to learn the solutions of these local problems. We demonstrate the effectiveness of this novel approach through a series of experiments on a range of robot motion planning problems. Our results show that our approach can significantly reduce the computation time of the MIP solution, and make it more scalable to large and complex environments.