Researchers have developed a new approach called Shared Recurrent Memory Transformer (SRMT) for multi-agent lifelong pathfinding. SRMT enables agents to share information about their behavior and intentions without relying on communication protocols or centralized training. This allows for decentralized training and execution, making it scalable to large agent populations. The authors evaluate SRMT on the Partially Observable Multi-Agent Pathfinding problem and demonstrate its ability to enable emergent coordination among agents, even with minimal guidance from reward functions. They also show that SRMT outperforms other approaches in certain scenarios. --- Why it matters: This matters because it addresses a significant challenge in multi-agent reinforcement learning: enabling coordination among decentralized agents without relying on communication protocols or centralized training. This has implications for developing more scalable and efficient AI systems for applications like autonomous vehicles, robotics, and smart homes. Source: https://arxiv.org/abs/2501.13200