Large Language Models for Multi-Robot Systems: A Survey

In plain words

A team of robots has to talk to each other, split up the work, and keep going when one of them fails. A large language model is a text system that can read an instruction, break a goal into steps and explain its choices in ordinary language. This survey asks where such models can help a team of physical robots.

It sorts the work by level: deciding which robot does what, planning routes, producing low-level commands, and involving a human operator. It also looks at how the robots pass messages to each other, where these systems have been tested, and what stops them from working reliably in the field.

The main message is that the models help most when they produce a plan or a program that fast, ordinary controllers then carry out. The gaps that remain are latency, unreliable messages and small test teams. The survey is a map of published work, not a method to copy.

TL;DR

• First dedicated survey of large language models in multi-robot systems §8.
• Sorts the field into four decision levels: task split and planning, motion planning, low-level commands and operator involvement §4.
• Its synthesis: feedback loops and persistent executable artefacts beat one-shot, per-step querying §4.1.2 §4.3.
• Open problems: per-step latency, lossy communication, few validations beyond about ten robots, and no LLM-specific benchmark metrics §7.1 §6.3.
• It proposes no method and reports no experiments of its own; numbers below are from the cited works.

Problem

Multi-robot systems gain scalability and resilience from teams of simpler robots, but they need communication, coordination under uncertainty and collective decisions that adapt in real time §1 §2.1. LLMs could serve as a natural-language interface between robots and operators, and could decompose a mission into subtasks §1.

Multi-agent LLM results do not transfer directly. Those benchmarks treat actions as atomic and always successful, assume perfect observability, and assume lossless, turn-synchronous messaging with negligible latency §2.3. Earlier surveys cover single robots or abstract multi-agent systems, not multi-robot teams §2.3.

Method

The survey has no algorithm. Its structure is a taxonomy along two axes: the decision level at which the LLM acts, and the communication architecture.

Decision levels §4:

• High-level task allocation and planning: the LLM decides what each robot should do. Sub-cases are multi-robot multi-task planning §4.1.1 and complex task decomposition §4.1.2.
• Mid-level motion planning: navigation and path planning §4.2.
• Low-level action generation: the LLM produces actuator-level commands or posture control §4.3.
• Human intervention: the LLM interacts with operators across all levels §4.4.

The survey defines task planning, task allocation, motion planning, action generation, execution and reasoning as separate terms and applies them consistently §4.

Communication architectures §3: centralised (one coordinator with all information), decentralised (each robot runs its own LLM and talks to peers) and hierarchical or hybrid designs. Table 1 labels each system with its category tab. 1. Chen et al. compare four designs (fully decentralised DMAS, fully centralised CMAS, and two hybrids HMAS-1 and HMAS-2) §3.

Recurring patterns the survey draws across the cited systems:

• A two-stage split: the LLM parses instructions into a structure (subtask tree, dependency DAG, temporal-logic formula, skill graph), and a formal component (planner, linear-programming solver or scheduler) does the optimisation §4.1.2.
• Predefined skill libraries reduce hallucinated actions but limit generalisation. Capabilities described in prompts generalise more but are less reliable §4.1.2.
• Closed-loop correction (re-prompting on constraint violations, self-debugging, retrieval of past experience) generally outperforms one-shot planning §4.1.1 §4.1.2.
• The most successful low-level systems have the LLM emit a persistent artefact (a signed distance function, a Python control policy, or per-robot coordinates). A classical controller executes it, so the LLM is not called at every control step §4.3.
• Three motion-planning integration modes: a global planner queried at each step, an offline encoder whose embeddings shape the policy input, and a sparse conflict resolver triggered only on deadlocks §4.2.
• Human oversight modes: command-then-report, approval before execution, verification after plan generation, and on-demand help requests from the robot §4.4.

The survey gives no algorithm. The sketch below is the two-stage split it describes across the cited systems §4.1.2. It is an illustration, not a procedure from any one paper.

# Two-stage pattern shared by several high-level systems  [§4.1.2]
# Input:  mission (natural-language text), robots (capability descriptions)
# Output: per-robot assignments, checked by a formal component

def plan_mission(mission, robots):
    # Stage 1: the LLM turns free text into a structure  [§4.1.2]
    structure = llm_parse(mission, robots)     # subtasks, DAG, or temporal logic

    # Stage 2: a formal solver optimises the structure  [§4.1.2]
    assignments = solver(structure, robots)

    # Closed loop: feedback on failure triggers re-prompting  [§4.1.1]
    while not assignments_feasible(assignments, robots):
        feedback = collect_feedback(assignments)
        structure = llm_parse(mission, robots, feedback)
        assignments = solver(structure, robots)

    return assignments

Background concepts the survey defines: fine-tuning with low-rank adaptation, retrieval-augmented generation, multimodal LLMs, vision-language models and vision-language-action models §2.2 §6.1.

Hyperparameters

Not applicable. The survey proposes no method and gives no training or inference settings §1. Settings used by individual surveyed systems are not restated here; read the cited works.

Results

Author-reported. These are the survey's own counts and the figures it quotes from cited works. The survey runs no experiments of its own.

Paper vs code

No official code. The abstract links github.com/Zhourobotics/LLM-MRS-survey [abstract]. Checked its README: it is a curated, continuously updated list of papers, not implementation code, so it is not cloned or recorded as code. The survey text says the list keeps growing [abstract], so it may differ from the v5 text. No discrepancies between code and paper apply.

Pitfalls

• Double counting in Table 1. Wu et al. (hierarchical LLMs for multi-robot multi-target tracking) appears under both the high-level and the mid-level rows tab. 1 §4.1 §4.2. The 50 rows are 49 distinct papers.
• Unverified modality marks. Table 1 marks audio input (A) for Lakhnati et al. and for Lykov et al. (LLM-MARS). The text describes no audio processing for either §4.1 tab. 1. Treat as unverified.
• Latency claim is narrower than it reads. §7.1 says only Lim et al. and SayCoNav give concrete timings. The deployment table also reports a figure for GenSwarm (about 6 min, one-time generation) and a replan count for RoCo tab. 2. The claim holds for per-step latency only.
• Scale claim is a generalisation. §7.1 says few approaches are validated beyond about ten robots. The deployment table includes a simulation with up to 32 robots tab. 2. Quote the table's numbers rather than "about ten".
• BOLAA is not a robot benchmark. §6.3 presents BOLAA as useful for LLM-driven multi-robot systems. The later revised paragraph in §6.3 says it targets software agents on WebShop and HotPotQA.
• RoCoBench metrics conflict. The first description in §6.3 mentions precision and robustness metrics. The revised paragraph lists success rate, step counts and replan count, and says perception-noise robustness is not reported. Check before citing.
• Citation mismatch. The Habitat-MAS sentence in §6.2 cites the EMOS paper.
• Revision-marked passages. Several paragraphs (§2.3, §4.1, §4.2, §4.3, §4.4, §6.3, §7.1, §7.2 and §8) are set off with a revision macro in the LaTeX source. They hold most of the specific numeric claims. Check them against the arXiv version you cite.
• Table 2 is a sample. It covers twelve systems, not the 49 distinct papers in Table 1 tab. 2 caption.
• Figures are not ours to reuse. The acknowledgement says several figures are reproduced from other papers with permission §8. This card contains no figures, and the figures are not covered by this paper's license.
• Snapshot. The source is v5, updated 2026-05-03. Later work is not covered.

Lineage

• Inspired by: the survey of LLM-based multi-agent systems by Guo et al. (2024). Its arXiv id is not given in the source, so it is not linked §1 §2.3.
• Related surveys compared: Firoozi et al. (foundation models in robotics), Zeng et al. (LLMs in robotics), Kim et al. (LLM applications in communication, perception, planning and control), Wang et al. (LLMs for manipulation), Hunt et al. (language in robotics), Kawaharazuka et al. (foundation models in real-world robotics) §2.3.
• Superseded by: unknown. The survey describes itself as a continuously updated list [abstract].