Communication-Efficient Digital-Twin Coordination for Heterogeneous LLM Embodied Agents over Computing Power Networks
06:00 · July 13, 2026 · arXiv cs.AI RSS
Embodied agent teams powered by heterogeneous large language models (LLMs) are being widely deployed in physical artificial intelligence such as smart factories, warehouses, and service robotics. To enable collaboration among such an agent team, 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. First, inter-agent dialogue incurs communication overhead that grows rapidly with team size. Second, the quality of coordination is constrained by the heterogeneous capabilities of the agent team's LLMs. Third, 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. Then, DT executes a training-free, rule-based orchestrator algorithm to resolve cross-agent conflicts and returns 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.
Summary
LDT-Coord addresses coordination among teams of embodied agents whose large language models differ in capability and resource footprint. Conventional approaches rely on repeated natural-language exchanges to negotiate task assignments and resolve conflicts over shared resources. These exchanges impose communication loads that scale with team size, expose overall performance to the weakest model in the group, and introduce latency because agents must complete dialogue before acting.
The proposed framework inserts a lightweight digital twin as coordination middleware. Each agent first selects its intended action locally and transmits only an action tuple together with typed temporal constraints on shared resources. The digital twin applies a training-free, rule-based orchestrator that detects mutual-exclusion, synchronization, and dependency conflicts, then returns compact downlink instructions that enforce a maximal consistent execution set. Because coordination now operates over structured primitives rather than natural-language text, performance is decoupled from the language-generation quality of any individual agent.
To limit uplink traffic further, the decision of which agents report at each time step is cast as a constrained partially observable Markov decision process. The PPO-Lagrangian algorithm solves this formulation under an explicit per-step latency bound. Simulation results on a Confined-Space Sorting task show that the resulting system matches the task success rate of full natural-language dialogue while cutting communication volume by more than seventy times and preserving robustness across heterogeneous team sizes.
Why it matters
This research is highly relevant for Dutch AI researchers and practitioners in smart manufacturing, logistics, and robotics. It offers a mathematically rigorous, communication-efficient solution for deploying multi-agent LLM systems in resource-constrained industrial environments.
