Communicative and Cooperative Learning for Multi-agent Indoor Navigation

The ability to cooperate and work as a team is one of the “holy grail” goals of intelligent robots. To address the importance of communication in multi-agent reinforcement learning (MARL), we propose a Cooperative Indoor Navigation (CIN) task, where agents cooperatively navigate to reach a goal in a 3D indoor room with realistic observation inputs. This navigation task is more challenging and closer to real-world robotic applications than previous multi-agent tasks since each agent can observe only part of the environment from its first-person view. Therefore, this task requires the communication and cooperation of agents to accomplish. To research the CIN task, we collect a large-scale dataset with challenging demonstration trajectories. The code and data of the CIN task have been released. The prior methods of MARL primarily emphasized the learning of policies for multiple agents but paid little attention to the communication model, resulting in their inability to perform optimally in the CIN task. In this paper, we propose a MARL model with a communication mechanism to address the CIN task. In our experiments, we discover that our proposed model outperforms previous MARL methods and communication is the key to addressing the CIN task. Our quantitative results shows that our proposed MARL method outperforms the baseline by 6% on SPL. And our qualitative results demonstrates that the agent with the communication mechanism is able to explore the whole environment sufficiently so that navigate efficiently.