Large Language Models Empowering Multi-Robot Systems: A Comprehensive Technical Survey

Original Author/Maintainer: Zhourobotics (Peihan Li, Zijian An, Shams Abrar, Lifeng Zhou) Source Platform: GitHub Original Title: LLM-MRS-survey: Repository for the Large Language Models for multi-robot systems survey paper Original Link: https://github.com/Zhourobotics/LLM-MRS-survey Publish Time: 2026-06-14

This paper is a survey published by the Drexel University research team in the journal Autonomous Robots. It systematically reviews the current applications of Large Language Models (LLMs) in Multi-Robot Systems (MRS), covering core levels such as high-level task allocation, mid-level motion planning, low-level action generation, and human-robot interaction, providing researchers with a panoramic overview and direction for technology selection.

Multi-Robot Systems (MRS) are a core research direction in robotics, but traditional rule-based or optimization algorithm-based methods have limitations in dealing with dynamic environments, uncertain tasks, and complex human-robot interactions. The capabilities of Large Language Models (LLMs) in natural language understanding, reasoning, and generation provide new possibilities for the intelligent upgrade of robot systems. The Drexel University team reviewed the current applications of LLMs in MRS and published it in Autonomous Robots to provide a reference for researchers and practitioners.

The survey adopts a hierarchical classification framework, dividing the applications of LLMs in MRS into four levels:

1. High-level Task Allocation: Responsible for task decomposition, allocation, and coordination. Typical works include DART-LLM (Dependency-Aware Task Decomposition), EMOS (Heterogeneous Multi-Robot Operating System), REBEL (Rule and Experience Enhanced Learning), etc.
2. Mid-level Motion Planning: Convert natural language descriptions into motion commands. Typical works such as Co-NavGPT, CAMoN, MARLIN explore the integration of LLMs with traditional planning algorithms.
3. Low-level Action Generation: Generate specific commands to control robot actuators. Typical works include LLM2Swarm, FlockGPT, ZeroCAP, but need to solve physical feasibility and safety issues.
4. Human-Robot Interaction and Intervention: Enhance the naturalness of human-robot interaction, convert human instructions into robot task descriptions, and provide feedback.

Key technical trends:

1. Architecture Trade-offs: Centralized (a central LLM coordinates all robots) vs. distributed (each robot has an independent LLM instance and collaborates via communication).
2. Multimodal Fusion: Combine sensor data such as vision and force to make up for the shortcomings of pure text input.
3. Safety and Interpretability: Ensuring instruction safety and improving the interpretability of model decisions are key to technology implementation.

Practical Significance: Provide researchers with a complete technical map to help quickly locate relevant research and tools; compare the advantages and disadvantages of methods at different levels to provide a basis for technology selection in application scenarios. Application Prospects: Expected to play a role in fields such as warehouse logistics, disaster rescue, smart agriculture, and industrial manufacturing, especially with significant advantages in scenarios of quickly adapting to new tasks and natural human-robot interaction.

Limitations: The hallucination problem of LLMs may lead to infeasible plans, computational delays affect real-time performance, and there is a lack of standardized evaluation benchmarks. Future Directions: To address the above limitations, the team will continuously update the literature database to track domain progress and provide a resource collection for cross-domain researchers.