Researchers propose LDT-Coord, a networked coordination framework for teams of embodied agents powered by heterogeneous large language models (LLMs). The system utilizes a lightweight digital twin server to decouple coordination from natural-language reasoning, thereby addressing high communication costs and action delays inherent in existing multi-round dialogue methods.
- Agents report intended actions and structured temporal constraints to the DT server instead of engaging in iterative negotiation.
- A training-free, rule-based orchestrator resolves cross-agent conflicts over shared resources.
- Agent reporting control is formulated as a constrained partially observable Markov decision process (C-POMDP) and solved using the PPO-Lagrangian algorithm.
- Simulation results show task success rates comparable to conventional methods while reducing communication overhead by more than 70x.
The framework maintains robustness under LLM heterogeneity, offering a reliable coordination mechanism for physical AI deployments in environments with limited network resources.