摘要：AbstractWe propose a centralized multi-robot motion planning approach that leverages machine learning and mixed-integer programming (MIP). We train a neural network to imitate optimal MIP solutions and, during execution, the trajectories predicted by the network are used to fix most of the integer variables, resulting in a significantly reduced MIP or even a convex program. If the obtained trajectories are feasible, i.e., collision-free and reaching the goal, they can be used as they are or further refined towards optimality. Since maximizing the likelihood of feasibility is not the standard goal of imitation learning, we propose several techniques aimed at increasing such likelihood. Simulation results show the reduced computational burden associated with the proposed framework and the similarity with the optimal MIP solutions.
关键词：KeywordsMachine LearningPath PlanningMotion ControlOptimal Control
