Digital Twin Boosts LLM Agent Coordination, Reduces Communication Overhead.

Teams of embodied AI agents, powered by diverse large language models, are increasingly deployed in physical environments like smart factories and robotics. Coordinating these agents efficiently, especially with limited network resources, is a significant challenge. Traditional methods relying on multi-round natural language conversations suffer from high communication overhead, limitations due to varying LLM capabilities, and action delays. A new framework, LDT-Coord, addresses these issues by employing a lightweight digital twin. Each agent independently decides its action and reports it, along with resource constraints, to a central digital twin server. This server then uses a rule-based orchestrator to resolve conflicts and issue coordination instructions, bypassing the need for complex natural language negotiation. To further optimize communication, the system models agent reporting control as a constrained partially observable Markov decision process, solved using the PPO-Lagrangian algorithm. Simulations demonstrate that LDT-Coord achieves comparable task success rates to conventional methods while drastically cutting communication overhead by over 70 times and maintaining performance despite LLM heterogeneity.