Technical Summary: Large Language Models for Multi-Robot Systems: A Survey

Problem Statement

While Large Language Models (LLMs) have demonstrated significant potential in enhancing single-robot intelligence and multi-agent systems (MAS), their integration into Multi-Robot Systems (MRS) presents unique and distinct challenges. Unlike MAS, which often operate in abstract or virtual environments, MRS involves physical robots interacting with the real world, requiring coordination, scalability, and adaptability to dynamic, uncertain environments. Existing surveys on LLMs in robotics or MAS often overlook the specific constraints of embodied multi-robot collaboration, such as physical embodiment, continuous-time dynamics, imperfect perception, and lossy communication. Furthermore, current benchmarks and simulation environments often fail to account for the latency, hallucination, and reasoning limitations of LLMs when applied to physical multi-robot fleets. There is a critical need for a dedicated review that systematically categorizes LLM applications in MRS, identifies the gap between virtual agent benchmarks and real-world deployment, and outlines the specific challenges and opportunities for this emerging field.

Methodology

This paper conducts a comprehensive survey of recent literature integrating LLMs into MRS. The methodology involves:

1. Systematic Categorization: The authors organize existing works into four functional levels:
   • High-Level Task Allocation and Planning: Strategic decision-making, task decomposition, and resource allocation.
   • Mid-Level Motion Planning: Navigation, path planning, and collision avoidance.
   • Low-Level Action Generation: Direct control of actuators and formation control.
   • Human Intervention: Mechanisms for human-robot interaction, oversight, and feedback.
2. Comparative Analysis: The survey compares different communication architectures (centralized, decentralized, and hybrid) and evaluates their performance across various domains (household, construction, formation control, target tracking, and games).
3. Benchmark and Simulation Review: The authors analyze existing simulation environments (e.g., AI2-THOR, PyBullet, RoCoBench) and benchmarks, highlighting their limitations regarding LLM-specific metrics (e.g., hallucination rates, latency, token consumption).
4. Gap Analysis: The paper contrasts the assumptions common in MAS benchmarks (perfect observability, lossless communication, atomic actions) with the realities of MRS deployment (noisy sensors, communication delays, continuous dynamics).

Key Contributions

• First Dedicated Review: This is the first survey specifically focused on the integration of LLMs into Multi-Robot Systems, distinguishing it from broader robotics or MAS surveys.
• Structured Framework: The paper introduces a four-tier framework (High-Level, Mid-Level, Low-Level, Human Intervention) to classify LLM applications in MRS, providing a clear taxonomy for researchers.
• Architectural Insights: It identifies common design patterns, such as the "hybrid" approach where a central LLM handles initial allocation while local agents manage validation, and the trend of using LLMs to generate executable artifacts (e.g., code, signed distance functions) rather than querying them at every control step to mitigate latency.
• Critical Gap Identification: The survey explicitly details the "Sim-to-Real" gap in LLM-based MRS, noting that current simulators often treat LLM inference as zero-cost and instantaneous, failing to model latency variance, communication packet loss, and perceptual noise.
• Benchmark Critique: The authors critique existing benchmarks for lacking LLM-specific metrics (e.g., hallucination rates, reasoning chain quality) and for often being limited to small-scale or single-agent scenarios, proposing the need for unified metrics covering communication overhead, latency scaling, and robustness.

Results and Findings

• Performance Variability: The effectiveness of LLMs in MRS varies significantly based on the communication architecture. Hybrid frameworks (e.g., HMAS-2) often outperform purely centralized or decentralized approaches in complex tasks, balancing scalability with coordination.
• Limitations in Reasoning: LLMs struggle with precise mathematical reasoning required for trajectory optimization and spatial planning. They are also prone to hallucinations, which can lead to faulty decision-making in collaborative tasks.
• Latency and Scalability: High latency in LLM inference (ranging from seconds to minutes) poses a significant barrier to real-time control. Most current systems are validated on small fleets (typically fewer than 10 robots), and scalability to larger teams remains an unvalidated challenge.
• Domain Applications: Successful applications have been demonstrated in household tasks (navigation, object manipulation), construction (excavation, transportation), formation control (swarm coordination), and target tracking. However, many of these rely on simulation or small-scale physical tests.
• Human Role: Human intervention is currently primarily reactive (e.g., error recovery, pre-execution approval) rather than collaborative, with limited frameworks treating humans as active peers in negotiation.

Significance and Claims

The paper claims to provide a foundational resource for researchers aiming to bridge the gap between LLM reasoning and real-world robotic tasks. Its significance lies in:

• Clarifying the Distinction: It rigorously distinguishes MRS from MAS, emphasizing that physical embodiment introduces constraints (dynamics, sensor noise, communication loss) that invalidate many assumptions made in virtual multi-agent research.
• Guiding Future Research: By outlining specific challenges (mathematical reasoning, hallucination, latency) and opportunities (fine-tuning, RAG, task-specific models, realistic simulators), the paper offers a roadmap for advancing the field.
• Promoting Standardization: It advocates for the development of unified benchmarks and metrics that go beyond simple task success rates to include deployment-relevant factors like communication overhead and robustness to noise.
• Facilitating Deployment: The survey aims to guide the transition from theoretical studies to practical deployments by highlighting the necessary steps to address the "Sim-to-Real" gap and improve the reliability of LLM-driven multi-robot systems in unstructured environments.

The authors maintain a modest tone, acknowledging that while LLMs offer transformative potential for MRS, current systems are still in early stages of development, with significant hurdles remaining before widespread, robust real-world deployment can be achieved. They emphasize that their work is a starting point for a rapidly evolving field, supported by a continuously updated open-source repository of relevant papers.