Embodied and Multi-Agent Reinforcement Learning: Advances, Challenges and Opportunities

Embodied and Multi-Agent Reinforcement Learning (MARL) lies at the intersection of artificial intelligence, robotics, and complex systems theory, enabling multiple agents whether physical or virtual to learn coordinated behaviors through direct interactions with their environments. By leveraging advances in deep reinforcement learning, decentralized decision-making, and communication protocols, MARL has shown promise in a range of applications such as cooperative robotics, swarm intelligence, autonomous driving, and large-scale simulations. Unlike single-agent reinforcement learning, the multi-agent paradigm introduces new layers of complexity: each agent must learn to navigate both the environment and the dynamic behavior of peers or competitors, often under conditions of partial observability and limited communication. This paper offers a comprehensive review and analysis of key topics driving progress in MARL. We begin by exploring social learning and emergent communication, focusing on how agents learn to share information or signals that enhance teamwork. We then delve into Sim2Real transfer approaches, critical for bridging the gap between simulation-based training and real-world deployments, particularly in safety-critical domains. Hierarchical reinforcement learning serves as a powerful framework to handle tasks at varying levels of complexity and abstraction, improving interpretability and sample efficiency. Lastly, we examine safety and robustness challenges, including adversarial interactions, non-stationarity, and explicit constraints that must be integrated into multi-agent systems. By highlighting the underlying mathematical formalisms, empirical methods, and open research questions, this paper aims to map out current trends and future directions in Embodied and Multi-Agent Reinforcement Learning.