Memory-augmented multi-robot teams that learn to adapt

As multi-robot teams increasingly permeate real world environments like factories, homes and other extrater-restrial surfaces, the ability to form joint strategies that can effectively adapt to new observations and changes in teammates' policy becomes more vital. Memory is a crucial part of adaptive behavior since it is the mechanism by which knowledge from past observations can be used to modify future behavior. Previous works have studied adaptive behaviors primarily with the tools of memory but have largely been limited to single robot approaches. In this paper, we formulate the Extended T-Maze domain that marries the difficulties of a task requiring explicit adaptive behavior, to the complexities introduced by concurrent actions of multiple robots acting in the same environment. We develop a memory-based learning approach using Gated Recurrent units with Memory Block as robot policies, and demonstrate its efficacy in diverse sets of experiments in the Extended T-Maze domain that vary across multiple axes of difficulty including team size and task depth. Our results show that the memory-based multi-robot controllers are able to form effective joint strategies significantly outperforming feedforward controllers trained without memory.