Source-linked AI summary
Multi-Agent Coordination across Diverse Applications: A Survey
Lijun Sun, Yijun Yang, Qiqi Duan, Yuhui Shi, Chao Lyu, Yu-Cheng Chang, Chin-Teng Lin, Yang Shen
TL;DR
The survey addresses the lack of an interdisciplinary framework connecting coordination research across diverse MAS applications and the questions of who and how to coordinate. It synthesizes general coordination problems and applications through a unified framework, then discusses open challenges and promising directions including hybrid, human-MAS, and LLM-based coordination.
Problem
Existing surveys rarely unify interdisciplinary coordination research around who to coordinate with and how to coordinate across diverse applications.
Method
The survey organizes coordination across applications through a unified framework and reviews general MAS problems, six applications, and future research topics.
Results
The survey provides unified insight into MAS coordination by analyzing four questions: what coordination is, why it matters, who to coordinate with, and how to coordinate.
Takeaways & Limitations
Hybrid hierarchical and decentralized coordination, human-MAS coordination, and LLM-based MAS are identified as promising future directions.
Takeaways & Limitations
Guaranteed safety in resolving MAS conflicts of interest faces the curse of dimensionality, and complex inter-agent dependencies in coordinated multi-agent autonomous driving remain insufficiently studied.
Abstract
from arXiv · showhide
Multi-agent coordination studies the underlying mechanism enabling the trending spread of diverse multi-agent systems (MAS) and has received increasing attention, driven by the expansion of emerging applications and rapid AI advances. This survey outlines the current state of coordination research across applications through a unified understanding that answers four fundamental coordination questions: (1) what is coordination; (2) why coordination; (3) who to coordinate with; and (4) how to coordinate. Our purpose is to explore existing ideas and expertise in coordination and their connections across diverse applications, while identifying and highlighting emerging and promising research directions. First, general coordination problems that are essential to varied applications are identified and analyzed. Second, a number of MAS applications are surveyed, ranging from widely studied domains, e.g., search and rescue, warehouse automation and logistics, and transportation systems, to emerging fields including humanoid and anthropomorphic robots, satellite systems, and large language models (LLMs). Finally, open challenges about the scalability, heterogeneity, and learning mechanisms of MAS are analyzed and discussed. In particular, we identify the hybridization of hierarchical and decentralized coordination, human-MAS coordination, and LLM-based MAS as promising future directions.
1 INTRODUCTION
This survey develops a unified understanding of multi-agent coordination across diverse applications, motivated by expanding MAS research and the need to connect coordination knowledge across domains. It organizes the survey around four questions and reviews general problems, applications, and open research directions.
- Motivation: MAS research has expanded across diverse applications and disciplines, while new multi-agent tasks continue to emerge with technological advancement.The survey frames cross-domain knowledge transfer as necessary for broader coordination understanding.
- Emerging applications: Emerging MAS settings include LLM-based systems and coordinated autonomous driving, alongside established application areas.The introduction highlights collective intelligence, natural-language reasoning and planning, and increasing automation in road vehicles.
- Research gap: Existing surveys commonly classify coordination by techniques, tasks, or application domains, but rarely unify the questions of who to coordinate with and how to coordinate.The survey identifies this interdisciplinary unification as an important gap.
- Survey approach: The survey introduces a unified perspective covering what coordination is, why it matters, who to coordinate with, and how to coordinate.The framework connects coordinated tasks, agent relationships, coordination methods, and system-level objectives across applications.
- Survey scope: General coordination problems and six MAS applications are reviewed from complementary methodology and task perspectives.The survey also discusses challenges and promising directions before presenting its conclusions.
2 A FRAMEWORK FOR MULTI-AGENT COORDINATION
The framework defines coordination as agents interacting and making decisions for overall system-level performance, including resolving conflicting interests. It organizes coordination as an iterative process of evaluating performance, choosing coordination partners, and managing dependencies.
- Multi-agent systems: A multi-agent system consists of multiple independent interactive decision makers called agents, which may include people, robots, language models, or satellites.The definition treats agents broadly across physical, computational, and human entities.
- Coordination: Coordination is defined as agents interacting and making decisions for overall system-level performance, including resolving conflicting interests.The definition identifies system-level performance and conflict resolution as central coordination concerns.
- Unified framework: The unified coordination process iteratively evaluates system-level performance, chooses who to coordinate with, and determines how to coordinate.These three components form the survey’s framework for sequential decision-making.
- Who to coordinate with: Choosing who to coordinate with forms agent clusters from inter-agent dependencies that may be spatial, logical, hierarchical, decentralized, or task-based.Examples include coordination graphs, neighborhoods, conflict-resolution groups, and task-centered pursuit clusters.
- Who to coordinate with: Higher-order and transitive dependencies can expand local coordination clusters and support more sparse, accurately weighted coordination graph topologies.An agent influenced by another agent may need to coordinate with a third agent through transitive dependencies.
- How to coordinate: How to coordinate concerns managing dependencies through rule-based, game-theoretic, learning-based, and evolution-based methods.Application-specific taxonomies can combine these methodological categories, such as centralized resource sharing for heterogeneous swarms based on MARL.
- System-level evaluation: Coordination performance is evaluated at the system level, including trade-offs between individual interests and emergent collective, swarm, or societal intelligence.The framework is inspired by biological multi-agent systems and resembles brainstorming through repeated intra-cluster and inter-cluster interaction.
3 GENERAL MAS
The general MAS framework organizes coordination around inter-agent dependencies, learning, communication and cooperation, and conflict resolution. It uses evaluation and clustering to determine coordination partners, then updates to determine coordination behavior, while balancing coordination quality against scalability and communication constraints.
- Coordination structure: A coordination graph represents agents as nodes and their interactions or dependencies as edges, with clusters emerging among highly related agents.Clusters may reflect common interests, shared targets, or mutual benefits.
- Coordination structure: Evaluation and clustering determine which agents coordinate, while updating determines how coordination occurs across general MAS tasks.The surveyed tasks include coordinated learning, communication and cooperation, and conflict resolution.
- Coordinated learning: Coordinated learning ranges from independent learning to centralized training with decentralized execution, using global state or explicit communication to improve learning efficiency.Independent learning is disconnected and faces non-stationarity, whereas coordinated learning can use centralized critics and shared parameters.
- Coordinated learning: Centralized training and decentralized execution supports parallel decentralized policies with centralized critics that may access additional information during training.The paradigm can be combined with parameter sharing and includes actor-critic and value-decomposition methods.
- Communication and cooperation: Communication and cooperation coordinate shared goals through message exchange or environmental changes, using choices about partners, timing, content, encoding and communication form.Communication may be broadcast, multi-hop, point-to-point, synchronous, asynchronous or hierarchical.
- Communication and cooperation: Selective communication uses topology design, adaptive control or message weighting to determine communication partners, but attention-based weighting does not reduce communication overhead.Communication resources are constrained, motivating selective communication as an optimization objective.
- Communication and cooperation: Event-triggered communication can improve efficiency, flexibility and scalability relative to periodic schemes, potentially compromising coordination performance.The trigger may be predefined or learned.
- Conflict-of-interest resolution: Conflict resolution manages shared-resource conflicts and collisions to produce smooth coordinated behavior, but rule-based solutions are difficult to optimize for all cases.The discussion covers physical systems, centralized MAPF solvers and distributed learning-based solvers.
4 MAS APPLICATIONS
The survey presents MAS applications through a unified perspective spanning established domains such as search and rescue, logistics, and transportation, as well as humanoid and anthropomorphic robots. Across these applications, coordination supports search, task allocation, routing, safety, and system-level efficiency.
- Application scope: The survey uses a unified perspective to illustrate MAS applications across widely studied domains and emerging areas.It includes search and rescue, logistics, transportation, humanoid and anthropomorphic robots, and other agent forms.
- Search and rescue: Search-and-rescue robots coordinate to sweep environments and may form line or V-shaped formations while searching for targets.These systems address environments including rugged terrain, cities, battlefields, and water operations.
- Warehouse automation and logistics: Warehouse agents coordinate picking, sorting, transporting, communication, resource usage, and routing to reduce bottlenecks and idle time.Real-time task reassignment and route adjustment support changing demand or inventory, improving throughput, efficiency, and scalability.
- Transportation systems: Transportation MAS address system-level traffic flow, energy efficiency, safety, mobility accessibility, and congestion through coordinated agents.Applications include traffic-signal control, vehicle platooning, merging, intersection navigation, lane changing, and overtaking.
- Humanoid and anthropomorphic robots: Humanoid and anthropomorphic robots coordinate multiple arms through grasp selection, motion synchronization, and collision avoidance.Multi-arm systems are advantageous for heavy, large, or long objects, while decentralized solutions improve scalability and adaptability under varied team sizes or malfunctioning components.
5 FUTURE AND OPEN TOPICS
The survey identifies scalability, heterogeneity, and learning as major open topics in multi-agent coordination. It highlights hybrid hierarchical-decentralized mechanisms and human-MAS coordination as directions for improving large-scale adaptability, system coherence, and collaboration.
- Scalability: Scalability remains a central challenge because increasing agent numbers complicates conflict evaluation, clustering, updating, computation, and learning.Scalable coordination should preserve or proportionally improve system-level performance as the number of agents grows.
- Scalability and hybrid coordination: Hierarchical coordination assigns greater responsibilities to selected agents, supporting scalability and efficient management of large-scale MAS.The survey discusses regional leaders as an example in autonomous-driving networks and notes support for diverse local and global interaction densities.
- Scalability and hybrid coordination: Hybrid hierarchical-decentralized coordination combines global coordination and scalability with local adaptability, resilience, and swift decision-making.The survey presents this hybridization as suitable for complex tasks in large and dynamic environments.
- Heterogeneity: Heterogeneous MAS require specialized agents because agents may differ physically, behaviorally, logically, and in their coordination protocols.The survey identifies coordination among heterogeneous agents as insufficiently studied, especially when complex inter-agent dependencies and system-level concerns are involved.
- LLM-based and heterogeneous MAS: LLM-based heterogeneous MAS have shown experimental results on language-based AI tasks, but heterogeneity makes coordination more difficult than in homogeneous MAS.The survey links this difficulty to agents coordinating differently and to the role of homogeneity in forming self-organized patterns.
- Human-MAS coordination: Human-MAS teaming combines human intent and intelligence with machine capabilities through proactive coordinated decision-making.Interaction modalities include graphical interfaces, gestures, touch, voice, eye movement, and neural signals; trust modeling supports adaptive and safe collaboration.
- Human-MAS coordination: Neural interfaces can provide real-time cognitive feedback for adaptive human-MAS decision-making in dynamic environments.Reported examples translate error-related human responses into corrective commands and monitor cognitive states during unexpected robot behavior.
6 CONCLUSION
The survey unifies interdisciplinary MAS coordination around four questions and reviews general coordination problems alongside diverse applications. It identifies hybrid hierarchical-decentralized coordination, human-MAS coordination, and LLM-based MAS as future directions addressing scalability, heterogeneity, and learning mechanisms.
- Survey contribution: The survey provides a unified insight into MAS coordination by asking what coordination is, why it matters, who coordinates, and how coordination occurs.These four questions organize its interdisciplinary analysis.
- Survey scope: It surveys three general MAS coordination problems and six applications spanning fundamental tasks, established domains, and emerging areas.The application coverage is intended to connect coordination insights across diverse settings.
- Future directions: The survey explores hybrid hierarchical-decentralized coordination, human-MAS coordination, and LLM-based MAS as three future directions.These directions are discussed in relation to scalability, heterogeneity, and learning mechanisms.
- Outlook: The authors anticipate that multi-agent coordination will drive a new stage of general AI.This statement is presented as the survey’s broader outlook.