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

The ability to cooperate and work as a team is one of the ‘holy grail’ goals of 1 intelligent robots. Previous works have proposed many multi-agent reinforcement 2 learning methods to study this problem in diverse multi-agent environments. How3 ever, these environments have two limitations, which make them unsuitable for 4 real-world applications: 1) the agent observes clean and formatted data from the 5 environment instead of perceiving the noisy observation by themselves from the 6 first-person perspective; 2) large domain gap between the environment and the 7 real world scenarios. In this paper, we propose a Multi-Agent Indoor Navigation 8 (MAIN) benchmark1, where agents navigate to reach goals in a 3D indoor room 9 with realistic observation inputs. In the MAIN environment, each agent observes 10 only a small part of a room via an embodied view. Less information is shared 11 between their observations and the observations have large variance. Therefore, 12 the agents must learn to cooperate with each other in exploration and communi13 cation to achieve accurate and efficient navigation. We collect a large-scale and 14 challenging dataset to research on the MAIN benchmark. We examine various 15 multi-agent methods based on current research works on our dataset. However, 16 we find that the performances of current MARL methods does not improve by the 17 increase of the agent amount. We find that communication is the key to addressing 18 this complex real-world cooperative task. By Experimenting on four variants of 19 communication models, we show that the model with recurrent communication 20 mechanism achieves the best performance in solving MAIN. 21