LDT-Coord cuts LLM agent team communication by 70x using digital twins

Embodied agent teams powered by heterogeneous large language models (LLMs) are increasingly deployed in smart factories, warehouses, and service robotics. These agents typically coordinate through multi-round natural-language conversations, which creates three intertwined challenges: communication overhead that scales with team size, coordination quality constrained by the weakest LLM, and action delays from iterative negotiation. To solve this, researchers introduce LDT-Coord (Lightweight Digital Twin Coordination). Each agent independently selects its intended action and sends both the decision and a structured temporal constraint to a central digital twin server, decoupling coordination from language reasoning ability.

The digital twin executes a training-free, rule-based orchestrator that resolves cross-agent conflicts and returns instructions. To further minimize bandwidth, agent reporting is modeled as a constrained partially observable Markov decision process (C-POMDP) and solved with the PPO-Lagrangian algorithm. In simulation, LDT-Coord matches the task success rate of conventional dialogue-based coordination while slashing communication overhead by more than 70x. The system also maintains robustness when agents have wildly different LLM capabilities. This opens the door for deploying large-scale robot teams in bandwidth-constrained environments like remote warehouses and industrial IoT networks.