Agentic AI
New AI framework enables robots to coordinate without central control using shared memory
Researchers introduce a transformer-based system that lets multiple agents learn to work together by broadcasting their learned states to a shared memory workspace.
A team of computer scientists has developed a novel approach to help groups of robots or autonomous agents coordinate their actions in real-time without relying on a centralized controller. The method, called the Shared Recurrent Memory Transformer (SRMT), represents a step forward in decentralized multi-agent reinforcement learning.
The coordination challenge#
In traditional multi-agent systems, coordination typically requires either restrictive communication protocols tied to specific domains, or a centralized training approach that struggles to handle large numbers of agents. Neither solution scales well or offers the flexibility needed for real-world deployments.
How SRMT works#
SRMT enables agents to share information through a global memory workspace where each agent can broadcast its learned working memory states and access those of other agents. This allows teams to exchange information and coordinate their behavior during both training and execution, all while maintaining fully decentralized operations.
The system uses transformer architecture to allow agents to query and interpret shared memory representations without explicit communication protocols or domain-specific constraints.
Testing on pathfinding tasks#
Researchers evaluated SRMT on the Partially Observable Multi-Agent Pathfinding (PO-MAPF) problem, where multiple agents must find optimal routes while avoiding collisions and deadlocks. Results showed that the shared memory approach fostered emergent coordination even when reward signals were sparse or minimal.
On a specially designed Bottleneck task requiring negotiation between agents, SRMT consistently outperformed baselines that used communication or memory augmentation. The system also generalized successfully to longer corridors it had never encountered during training.
Real-world scalability#
When tested on POGEMA benchmark maps, SRMT scaled effectively as the number of agents and map complexity increased, achieving competitive performance with recent alternatives—including other multi-agent reinforcement learning methods, hybrid approaches, and planning-based solutions—without needing domain-specific heuristics.
Researchers have made their code freely available on GitHub, enabling the community to build on the work.
