Large Language Models for Multi-Robot Systems: A Survey

📝 Paper Summary

Multi-Robot Systems (MRS) Embodied AI Human-Robot Interaction

This survey systematically categorizes how Large Language Models (LLMs) are integrated into Multi-Robot Systems (MRS) across high-level task allocation, mid-level planning, and low-level action execution.

Core Problem

Traditional Multi-Robot Systems (MRS) struggle with natural language communication, adaptive coordination in dynamic environments, and intuitive human-robot interaction due to rigid, predefined protocols.

Why it matters:

MRS scale and complexity often exceed single-robot capabilities, requiring robust coordination which traditional rigid protocols stifle
Operators lacking technical expertise need natural language interfaces to command robot swarms effectively
Prior surveys focus on single-robot systems or virtual multi-agent systems, missing the physical constraints and specific coordination challenges of real-world embodied MRS

Concrete Example: In a search and rescue mission, a traditional MRS requires specific coded commands to reallocate tasks. With LLMs, an operator can simply say 'Robot A, help Robot B move the debris,' and the system autonomously re-plans and executes the coordination.

Key Novelty

Comprehensive Taxonomy of LLM-MRS Integration

Establishes a three-level hierarchy for LLM usage in MRS: high-level task allocation, mid-level motion planning, and low-level action execution
Analyzes distinct communication architectures (Centralized, Decentralized, Hybrid) specifically for LLM-driven robot teams
Bridges the gap between virtual multi-agent system literature and physically embodied multi-robot research

Architecture

Comparison of four communication architectures for LLM-based Multi-Agent Systems: Centralized (CMAS), Decentralized (DMAS), and two Hybrid variants (HMAS-1, HMAS-2)

Evaluation Highlights

Reviews four distinct communication architectures (Centralized, Decentralized, HMAS-1, HMAS-2) for LLM-based MRS
Identifies Hybrid Multi-Agent Systems (HMAS-2) as superior to fully centralized or decentralized approaches for complex tasks involving >6 agents
Categorizes applications across 5 domains: household robotics, construction, formation control, target tracking, and robot games

Breakthrough Assessment

7/10

A timely and necessary survey that defines the taxonomy for a rapidly emerging field. While it doesn't propose a new model, it structures the research landscape for future work.

⚙️ Technical Details

Problem Definition

Setting: Integration of Large Language Models into Multi-Robot Systems to enhance autonomy and collaboration

Inputs: Natural language instructions, environmental observations, inter-robot messages

Outputs: Coordinated task plans, motion trajectories, executable robot actions, communication text

Pipeline Flow

Review of Communication Architectures
Review of Task Planning Levels (High/Mid/Low)
Review of Real-world Applications

System Modules

Communication Architecture Analysis (Taxonomy)

Categorize how LLMs interact within the team (Centralized vs. Decentralized)

Model or implementation: Conceptual Framework

Hierarchical Planning Categorization (Taxonomy)

Classify LLM utility at different abstraction levels

Model or implementation: Conceptual Framework

Novel Architectural Elements

Introduction of a structured taxonomy specifically for LLM-MRS integration, distinct from general Multi-Agent Systems (MAS) taxonomies
differentiation between 'virtual agent roles' (MAS) and 'physical robot interactions' (MRS)

Modeling

Base Model: Survey paper covering various models (GPT-4, Llama, PaLM, etc.)

Comparison to Prior Work

vs. Guo et al.: Focuses on physical embodiment and MRS constraints rather than virtual/abstract agents
vs. Zeng et al.: Focuses specifically on multi-robot coordination/collaboration rather than single-robot manipulation/navigation
vs. Hunt et al.: Focuses on collective intelligence and system architecture rather than just language as an interaction medium

Limitations

Survey depends on the current state of rapidly evolving literature
Does not propose a new algorithm or model itself
Mathematical reasoning limitations of current LLMs hamper complex MRS coordination
Hallucination remains a critical safety issue for physical robot deployment

Reproducibility

Code: https://github.com/Drexel-DISCO/LLM-MRS-Survey

Survey paper; code repository provided (https://github.com/Drexel-DISCO/LLM-MRS-Survey) contains the list of reviewed papers and resources.

📊 Experiments & Results

Evaluation Setup

Literature review and qualitative analysis of existing methods

Benchmarks:

Literature Analysis (Survey)

Metrics:

Qualitative categorization
Architecture comparison
Application coverage
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Hybrid architectures (HMAS-2) often outperform fully centralized or decentralized approaches for complex multi-robot tasks involving larger teams (>6 agents)
LLMs are transforming MRS by enabling natural language interfaces for non-experts to control robot swarms
Current challenges preventing full adoption include high latency, lack of robust benchmarks for physical coordination, and safety concerns due to hallucinations
Future directions lie in Vision-Language-Action (VLA) models for closed-loop control and fine-tuning specialized models for MRS coordination

📚 Prerequisite Knowledge

Prerequisites

Fundamentals of Multi-Robot Systems (MRS)
Basics of Large Language Models (LLMs)
Control paradigms (Centralized vs. Decentralized)

Key Terms

MRS: Multi-Robot Systems—systems consisting of multiple autonomous robots working collaboratively to complete tasks

LAAs: LLM-augmented Autonomous Agents—agents that use LLMs for reasoning, planning, and decision-making

CMAS: Centralized Multi-Agent System—a framework where a single central controller/LLM directs all robots

DMAS: Decentralized Multi-Agent System—a framework where each robot operates its own LLM for decision-making

HMAS: Hybrid Multi-Agent System—a framework combining centralized oversight with decentralized local execution

RAG: Retrieval-Augmented Generation—enhancing LLM responses by retrieving relevant information from external databases during runtime

LoRA: Low-Rank Adaptation—a parameter-efficient fine-tuning technique that freezes pre-trained weights and injects trainable rank decomposition matrices

VLM: Vision-Language Model—models that combine visual perception with language reasoning

VLA: Vision-Language-Action Model—models that link perception and reasoning directly to executable robotic actions

homogeneous MRS: A team where all robots are identical in hardware and functionality

heterogeneous MRS: A team consisting of different types of robots with specialized capabilities