Emergent Language based Dialog for Collaborative Multi-agent Navigation

Han Yu, Hengtong Lu, Caixia Yuan, Xiaojie Wang

Primary Area: applications to robotics, autonomy, planning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: emergent communication

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: This paper aims to provide an empirical study on how to build agents that can collaborate effectively with multi-turn emergent dialogues. Recent research on emergent language for multi-agent communications mainly focuses on single-turn dialogue and simple settings where observations are static during communications. Here, we propose a multi-agent navigation task, a representative task with multi-turn communications and dynamic environment observations: the Tourist (the embodied agent) who can observe its local visual surroundings, and the Guide who has a holistic view of the environment but no foreknowledge of the Tourist's location. The objective of the task is to guide the Tourist to reach the target place via multi-turn dialogues emerging from scratch. To this end, we propose a collaborative multi-agent reinforcement learning method that enables both agents to generate and understand emergent language, and develop optimal dialogue decisions with a long-term goal of solving the task. We also design a real-world navigation scene with the matterport3D simulator. The result shows that our proposed method highly aligns emergent messages with both surroundings and dialogue goals, hinting that even though without human annotation or initial meaning, the agents can learn to converse and collaborate under task-oriented goals.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5358