SLIM architecture lets drone swarms share info with minimal bandwidth

Multi-agent reinforcement learning (MARL) systems—like drone swarms searching for survivors—need to communicate to coordinate, but real-world bandwidth is often severely limited. Existing architectures suffer from a coupled bottleneck: the same latent representation serves both policy execution and inter-agent messaging. Reducing message size directly shrinks the policy's capacity, causing sharp performance drops. Researchers Alexi Canesse, Benoît Goupil, Jesse Read, and Sonia Vanier tackle this with two contributions. First, they introduce β, a normalized per-agent budget that unifies sparsity, communication rounds, and message dimension into one comparable constraint. Second, they propose SLIM (Simple Latent Independent Messaging), a minimal architecture that cleanly separates the communication pathway from the policy's latent space. This decoupling allows agents to compress messages without restricting their internal reasoning.

Evaluated on several partially-observable MARL benchmarks where communication is essential, SLIM achieves state-of-the-art results. Crucially, as bandwidth is reduced—simulating real-world constraints—the performance degradation is marginal, unlike prior methods that collapse. The approach scales well with the number of agents and remains robust across different communication limits. This means real-world deployments (e.g., search-and-rescue robot teams, autonomous vehicle fleets) can maintain high coordination quality even when each agent can only send a few bytes per step. The work provides both a theoretical framework (β budget) and a practical architecture (SLIM) for building reliable, communication-efficient multi-agent systems.