The Holy Grail of Multi-Robot Planning: Learning to Generate Online-Scalable Solutions from Offline-Optimal Experts

Many multi-robot planning problems are burdened by the curse of\ndimensionality, which compounds the difficulty of applying solutions to\nlarge-scale problem instances. The use of learning-based methods in multi-robot\nplanning holds great promise as it enables us to offload the online\ncomputational burden of expensive, yet optimal solvers, to an offline learning\nprocedure. Simply put, the idea is to train a policy to copy an optimal pattern\ngenerated by a small-scale system, and then transfer that policy to much larger\nsystems, in the hope that the learned strategy scales, while maintaining\nnear-optimal performance. Yet, a number of issues impede us from leveraging\nthis idea to its full potential. This blue-sky paper elaborates some of the key\nchallenges that remain.\n