Abstract
Communication between agents in collaborative multi-agent settings is ingeneral implicit or a direct data stream. This paper considers text-basednatural language as a novel form of communication between multiple agentstrained with reinforcement learning. This could be considered first stepstoward a truly autonomous communication without the need to define a limitedset of instructions, and natural collaboration between humans and robots.Inspired by the game of Blind Leads, we propose an environment where one agentuses natural language instructions to guide another through a maze. We test theability of reinforcement learning agents to effectively communicate throughdiscrete word-level symbols and show that the agents are able to sufficientlycommunicate through natural language with a limited vocabulary. Although thecommunication is not always perfect English, the agents are still able tonavigate the maze. We achieve a BLEU score of 0.85, which is an improvement of0.61 over randomly generated sequences while maintaining a 100% maze completionrate. This is a 3.5 times the performance of the random baseline using ourreference set.