"SRMT: Shared Memory for Multi-agent Lifelong Pathfinding"

The paper addresses the challenge of coordinating multiple agents in pathfinding tasks.

It proposes a novel architecture called Shared Recurrent Memory Transformer (SRMT). SRMT allows agents to implicitly share information, enhancing cooperation and leading to improved performance in complex navigation scenarios.

📌 SRMT eliminates the need for explicit inter-agent communication. Instead, it leverages a shared recurrent memory, allowing agents to access relevant historical information through cross-attention. This significantly reduces computational overhead compared to traditional message-passing methods while maintaining high coordination efficiency.

📌 The architecture scales well in complex environments. By pooling individual agent memories into a global context, SRMT enables efficient information retrieval. This improves generalization to unseen scenarios where rigid rule-based approaches typically fail.

📌 SRMT's performance gains in bottleneck and sparse-reward settings highlight its robustness. Traditional MARL systems struggle in these conditions due to delayed rewards and limited feedback. SRMT mitigates this by enabling agents to make informed decisions using shared long-term memory, effectively reducing deadlocks.

-----

https://arxiv.org/abs/2501.13200

Original Problem 🤖:

→ Coordinating multiple agents in decentralized Multi-Agent Pathfinding (MAPF) is difficult.

→ Explicit prediction of other agents' behavior is needed for cooperation.

→ Existing methods often struggle with deadlocks and poor generalization in new environments.

-----

Solution in this Paper 💡:

→ Shared Recurrent Memory Transformer (SRMT). extends memory transformers to multi-agent settings.

→ It enables agents to implicitly exchange information through shared memory.

→ Each agent maintains a personal recurrent memory.

→ SRMT pools these individual memories into a shared global memory.

→ Agents use cross-attention to access and utilize this shared memory.

→ This allows implicit coordination without explicit communication protocols.

-----

Key Insights from this Paper ✨:

→ Shared recurrent memory enhances coordination in decentralized multi-agent systems.

→ SRMT outperforms baselines, especially with sparse rewards and in long corridors.

→ Implicit information sharing via memory is effective for multi-agent cooperation.

→ SRMT demonstrates good generalization to unseen environments and scales well.

-----

Results 📊:

→ SRMT achieves higher Cooperative Success Rate in Bottleneck tasks, especially with Sparse rewards.

→ SRMT shows top performance in Individual Success Rate and lower Sum-of-Costs in Bottleneck tasks.

→ SRMT is competitive with MARL, hybrid, and planning-based algorithms on POGEMA benchmarks.