ORCH：将人类组织理论引入具身智能体集群协同

My Translation Journey

Alright, let's break this down. My goal here is a professional, nuanced translation of this abstract, keeping the target audience – you, a field expert – in mind. It's crucial to convey the scientific rigor and the specific implications within multi-agent systems, collective intelligence, and organizational frameworks.

Firstly, I'm noting the core concept: collective intelligence isn't just about individual agent capabilities; it's how they're organized. The field often uses static structures, even when the job at hand demands something more dynamic. This immediately highlights the problem we're addressing: inflexible structures in artificial multi-agent systems and an opportunity for something new.

Next, I'm thinking about human organizational theory. We're not just taking inspiration; we're operationalizing it. That word, "operationalized," is key. I'm translating it as "转化为可操作的方法" to emphasize that this isn't just a theoretical exercise. It’s about building something useful.

Then comes ORCH. I'm choosing a translation, "组织角色与协同层级，Organizing Roles and Coordination Hierarchies", with the English expansion to make the link explicit for the audience and ensuring it's clearly understood. This method combines the concepts of pooled interdependence (集合式相互依存) for concurrent work and sequential interdependence (序贯式相互依存) for things that must happen in a particular order.

Now, for the validation: Firefighting missions. This isn't theoretical; we're running tests, and I'm stressing the scale, the heterogeneity of the agents, and the use of large language models to ensure the validation is clear. The key is in demonstrating that the approach is robust and efficient.

And the results? Significant improvements across the board: better mission outcomes, execution efficiency, exploration, and computational resource usage. Human-designed ORCH organizations consistently outperfom existing methods. Further, by using language models, we can automate this process. Finally, performance isn’t strictly tied to model size and scale. This is important to note and shows an interesting facet of this approach.

We're showing that hierarchical organization matters, especially in complex, long-duration missions. The organization enables these teams to maintain concurrent activities within specialized groups while coordinating between different task stages.

集体智能不仅取决于单个成员的能力，还取决于这些成员的组织方式。然而，当前的人工多智能体系统通常采用固定的组织架构，即使它们所执行的物理任务在协同需求上存在本质差异。在本研究中，我们证明可以将人类组织理论的原则转化为可操作的方法，用以组织大规模、异构的具身人工智能体集群。我们提出了 ORCH（组织角色与协同层级，Organizing Roles and Coordination Hierarchies），该方法通过将可并发推进工作中的集合式相互依存（pooled interdependence）与受前置关系制约工作中的序贯式相互依存（sequential interdependence）相结合，构建出针对特定任务的层级化组织。在涵盖侦察、搜救、运输、资源管理、火势遏制与扑灭的 25 项野火响应任务中，我们基于八个大型语言模型对规模多达 50 个异构智能体的团队进行了评估。基于这些原则构建的组织在任务成果、执行效率、探索能力以及计算资源利用率等各方面，均持续优于四种代表性的具身多智能体方法。相较于先前的四种框架，由人类设计的 ORCH 组织使最终得分平均提升了 63.97%，执行效率平均提升了 74.29%。而由语言模型自动生成的组织也使这两项指标分别提升了 43.63% 和 52.53%。这些优势在各类任务及不同的底层语言模型中均得以保持。值得注意的是，集体性能并非单纯由模型规模单调决定。对长周期任务的分析表明，层级化组织使团队能够在协调任务阶段间有序转换的同时，保持各专业小组内部活动的并发推进。