TR2021-134

Fast Multi-Robot Motion Planning via Imitation Learning of Mixed-Integer Programs


    •  Srinivasan, M., Chakrabarty, A., Quirynen, R., yoshikawa, N., Mariyama, T., Di Cairano, S., "Fast Multi-Robot Motion Planning via Imitation Learning of Mixed-Integer Programs", IFAC Modeling, Estimation and Control Conference (MECC), October 2021.
      BibTeX TR2021-134 PDF
      • @inproceedings{Srinivasan2021oct,
      • author = {Srinivasan, Mohit and Chakrabarty, Ankush and Quirynen, Rien and yoshikawa, nobuyuki and Mariyama, Toshisada and Di Cairano, Stefano},
      • title = {Fast Multi-Robot Motion Planning via Imitation Learning of Mixed-Integer Programs},
      • booktitle = {IFAC Modeling, Estimation and Control Conference (MECC)},
      • year = 2021,
      • month = oct,
      • url = {https://www.merl.com/publications/TR2021-134}
      • }
  • MERL Contacts:
  • Research Areas:

    Control, Machine Learning, Optimization

Abstract:

We 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-is 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 optimal MIP solutions.