Language Grounded Multi-agent Reinforcement Learning with Human-interpretable Communication

Furthermore, the learned communication protocols exhibit zero-shot generalization capabilities in ad-hoc teamwork scenarios with unseen teammates and novel task states.