A Machine Learning Method for Improving Task Allocation in Distributed Multi-Robot Transportation

Machine learning is a means of automatically generating solutions that perform better than those that are hand-coded by human programmers. We present a general behavior-based algorithm that uses reinforcement learning to improve the spatio-temporal organization of a homogeneous group of robots. In this algorithm each robot applies the learning at the level of individual behavior selection. We demonstrate how the interactions within the group affect the individual learning in a way that produces group-level effects, such as lane-formation and specialization, and improves the group's performance. We also present a model of multi-robot task allocation as resource distribution through vacancy chains, a distribution method common in human and animal societies, and an algorithm for multi-robot task allocation based on that model. The model explains and predicts the task allocation achieved by our algorithm and highlights its limitations. We present experimental results that validate our model and show that our algorithm outperforms pre-programmed solutions. Last, we present an extension of our algorithm that makes it applicable to heterogeneous groups of robots