Collaboration in Multi-Robot Systems: Taxonomy and Survey over Frameworks for Collaboration

1. Problem Statement

Despite the prevalence of multi-robot systems (MRS) in applications like disaster response, logistics, and environmental monitoring, the terminology used to describe collective behaviors remains inconsistent across research communities. Specifically, the terms cooperation, coordination, and collaboration are frequently used interchangeably, obscuring critical distinctions regarding:

• Inter-robot dependencies: How much do robots rely on one another?
• Information exchange: Is there explicit communication or just shared intent?
• Joint capabilities: Do the robots achieve something impossible for an individual robot?

This lack of a unified definition complicates the comparison of methods, hinders the transfer of approaches across problem settings, and prevents the clear identification of systems that truly possess "collaborative" capabilities (i.e., capabilities that exceed the sum of individual parts).

2. Methodology

The authors employ a conceptual taxonomy followed by a comprehensive literature survey.

• Taxonomy Construction: The paper first rigorously defines and distinguishes three core concepts (Cooperation, Coordination, Collaboration) using a hierarchical framework. These definitions are visualized via a Venn diagram and illustrated through a unified warehouse transport example involving robots of varying sizes and packages of varying weights.
• Literature Survey: Building on the taxonomy, the authors review existing frameworks that support Collaboration (the intersection of all three concepts). The survey categorizes frameworks based on:
  1. Organizational Structures: Centralized, Decentralized, and Hierarchical.
  2. Methodological Inspirations: Ecology-inspired, Human-Swarm Interaction (HSI), Game Theoretic, and Learning-Based approaches.
• Comparative Analysis: The paper synthesizes findings into a classification table (Table 1) to map how different frameworks satisfy the specific criteria for collaboration.

3. Key Contributions

A. Formal Definitions and Taxonomy

The paper establishes a strict hierarchy to differentiate collective behaviors:

1. Cooperation: The foundational layer. Defined as a non-adversarial intention to achieve a shared goal. Robots may act independently without explicit coupling, provided they do not hinder others.
2. Coordination: Adds information exchange and rule-following to ensure actions are compatible. Robots plan interdependent tasks to improve efficiency, but the task remains achievable by a single robot (no new capabilities are created).
3. Capability Complementarity: The ability to combine mobility, sensing, or actuation to execute tasks unattainable by any individual robot alone. This expands the joint action set.
4. Collaboration (The Core Contribution): Defined as the intersection of Cooperation, Coordination, and Capability Complementarity.
   • Formula: Collaboration = Non-adversarial Intent + Information Exchange + Joint Action (Capability Complementarity).
   • Implication: True collaboration requires robots to jointly perform tasks that are physically or logically impossible for a single agent, necessitating both shared intent and explicit coupling.

B. Survey of Organizational Architectures

The paper reviews how different architectures support collaboration:

• Centralized: A master controller assigns roles. High global optimality but suffers from single points of failure and scalability issues. (e.g., LLM-based planners).
• Decentralized: Robots autonomously decide when to collaborate via negotiation or local rules. High robustness and scalability but requires complex consensus mechanisms. (e.g., Coalition formation via Game Theory).
• Hierarchical: A compromise using team leaders or supervisors. Leaders make strategic decisions while subordinates execute. Common in Human-Swarm Interaction (HSI).

C. Survey of Methodological Frameworks

The authors categorize collaboration enablers into four domains:

1. Ecology-Inspired: Uses principles like mutualism (e.g., symbiotic relationships) to design control strategies where robots expand each other's safe operating regions or sensing capabilities.
2. Human-Swarm Interaction (HSI): Focuses on the coupling of human strategic reasoning with swarm autonomy. Collaboration emerges when human oversight and distributed execution create capabilities neither could achieve alone (e.g., VR/AR interfaces for shared situational awareness).
3. Game-Theoretic: Uses cooperative game theory (coalitions, Shapley values) to model binding agreements where agents pool resources to achieve outcomes unattainable individually.
4. Learning-Based: Leverages Multi-Agent Reinforcement Learning (MARL) and Large Language Models (LLMs).
   • CTDE (Centralized Training, Decentralized Execution): Allows agents to learn policies that rely on capability complementarity.
   • LLMs: Enable semantic reasoning and dialog-based task planning, allowing robots to negotiate roles for complex joint manipulation.

4. Results and Findings

• Dominance of Decentralization: The survey reveals that decentralized architectures dominate the literature, particularly in ecology-inspired and game-theoretic frameworks, due to scalability and robustness.
• Hybrid Paradigms: Learning-based systems often employ a hybrid approach (Centralized Training/Decentralized Execution) to balance global optimization with local robustness.
• Misclassification in Literature: A significant finding is that many works labeled as "collaborative" in existing literature actually only satisfy cooperation or coordination. They often optimize for efficiency or consensus but fail to demonstrate capability complementarity (i.e., the task could theoretically be done by one robot if given enough time or resources).
• The "Collaboration Gap": True collaboration is rare because it requires a specific task formulation where joint action is necessary, not just beneficial.

5. Significance and Future Directions

• Conceptual Clarity: The paper provides a necessary standard for the field, allowing researchers to precisely classify systems and avoid conflating simple coordination with true collaboration.
• Research Gaps Identified:
  • Theoretical Foundations: Lack of formal models that explicitly capture "collaborative intent" distinct from cooperative control.
  • Safety & Stability: Guarantees for stability and safety are well-established for coordination but under-studied for tightly coupled collaborative systems.
  • Learning Representation: Current MARL often treats collaboration as a byproduct of shared rewards. Future work must explicitly represent when and with whom to collaborate in the decision space.
  • Human-AI Integration: As LLMs and agentic AI advance, frameworks are needed to manage the trade-offs between human oversight and autonomous collaborative decision-making.

Conclusion:
This paper serves as a foundational reference for the multi-robot community, arguing that collaboration is a distinct, high-level capability defined by the necessity of joint action to achieve the impossible. By distinguishing it from mere coordination, the authors provide a roadmap for developing next-generation robotic systems capable of operating in complex, heterogeneous, and dynamic environments where individual robots are insufficient.