Abstract
Teams of embodied agents powered by heterogeneous large language models (LLMs) are being widely deployed in physical artificial intelligence applications such as smart factories, warehouses, and service robotics. To enable collaboration among such agent teams, efficient coordination mechanisms that operate reliably under limited network resources are required. However, existing heterogeneous LLM-agent coordination frameworks that rely on multi-round natural-language-based conversations introduce three coupled challenges:
- Inter-agent dialogue incurs communication overhead that grows rapidly with team size.
- The quality of coordination is constrained by the heterogeneous capabilities of the agent team's LLMs.
- Agents may suffer from action delays due to iterative negotiation.
To address these challenges, we propose LDT-Coord, a networked coordination framework built upon a lightweight digital twin (DT). Specifically, each agent independently selects its intended action and reports both the action decision and a structured temporal constraint over shared resources to the DT server, thereby decoupling coordination performance from natural-language reasoning ability. The DT then executes a training-free, rule-based orchestrator algorithm to resolve cross-agent conflicts and return coordination instructions to prevent such conflicts. To further reduce communication overhead, we formulate agent reporting control as a constrained partially observable Markov decision process (C-POMDP) and solve it with the PPO-Lagrangian algorithm. Simulation results show that LDT-Coord achieves a task success rate comparable to conventional coordination methods while reducing communication overhead by more than 70x and maintaining robustness under LLM heterogeneity.
Blogger's Review: LDT-Coord significantly enhances communication efficiency among heterogeneous LLM agents by decoupling coordination mechanisms from natural language processing abilities. This innovative framework not only reduces communication costs but also improves the overall robustness of the system, offering new insights for future intelligent agent collaborations. Its potential in practical applications deserves attention.