September 30, 2024
ISP student Huao Li's paper, "Language Grounded Multi-Agent Communication for Ad-hoc Teamwork," has been accepted by NeurIPS 2024.
This work proposes a novel approach that allows Multi-Agent Reinforcement Learning (MARL) agents to learn human-interpretable communication protocols, enabling them to collaborate more effectively with unseen teammates. The communication space of MARL agents is aligned with an embedding space of human language by leveraging synthetic data from embodied Large Language Model (LLM) agents. This language grounding not only improves interpretability but also accelerates the emergence of useful communication among agents. The authors demonstrate zero-shot generalization capability of the proposed method in ad-hoc teamwork scenarios where MARL agents successfully collaborate with LLM agents using natural language in novel situations on which they were not trained. This research takes an important step towards enabling AI systems that can flexibly communicate and coordinate with both artificial agents and humans in real-world teamwork settings. Check out the full paper on arXiv: https://arxiv.org/abs/2409.17348
This paper extends Huao's previous work on emergent communication presented at NeurIPS 2021 and Theory of Mind modeling presented at EMNLP and SMC this past year. He plans to test the effectiveness of these methods in human-agent teamwork by introducing human participants into previously trained agent teams.