MAIN: A Multi-agent Indoor Navigation Benchmark for Cooperative Learning

Fengda Zhu, Siyi Hu, Yi Zhang, Haodong Hong, Yi Zhu, Xiaojun Chang, Xiaodan Liang

Keywords: Multi-agent, Reinforcement learning, Embodied Navigation

TL;DR: A new multi-agent indoor navigation benchmark for cooperative multi-agent learning.

Abstract: Previous works have proposed many multi-agent reinforcement learning methods to study this problem in diverse multi-agent environments. However, these environments have two limitations, which make them unsuitable for real-world applications: 1) the agent observes clean and formatted data from the environment instead of perceiving the noisy observation by themselves from the first-person perspective; 2) large domain gap between the environment and the real world scenarios. In this paper, we propose a Multi-Agent Indoor Navigation (MAIN) benchmark, where agents navigate to reach goals in a 3D indoor room with realistic observation inputs. In the MAIN environment, each agent observes only a small part of a room via an embodied view. Less information is shared between their observations and the observations have large variance. Therefore, the agents must learn to cooperate with each other in exploration and communication to achieve accurate and efficient navigation. We collect a large-scale and challenging dataset to research on the MAIN benchmark. We examine various multi-agent methods based on current research works on our dataset. However, we find that the performances of current MARL methods does not improve by the increase of the agent amount. We find that communication is the key to addressing this complex real-world cooperative task. By Experimenting on four variants of communication models, we show that the model with recurrent communication mechanism achieves the best performance in solving MAIN.

URL: http://main-dataset.github.io/