GenSwarm — 摘要与引言
逻辑标注

"Automatic generation and deployment of control policies for multi-robot systems is an appealing paradigm, as it promises substantial savings in terms of human effort and other resources. However, this paradigm is nontrivial to realize as a multi-robot group as a whole cannot be programmed directly; rather, a desired collective behavior can be achieved only by programming each individual robot, which relies on its locally available information. Previous methods for automatic development of multi-robot systems are primarily based on optimization techniques. For instance, an objective function is first crafted to mathematically describe a desired task and then optimized to generate policies through methods such as evolutionary computation or systematic search. Despite their promise, these optimization methods face the common limitation of requiring manual crafting of objective functions."

"Recent advances in large language models (LLMs) and vision language models (VLMs) offer new paradigms for developing robotic systems. In one paradigm, a language model can be deployed onboard a robot to directly make decisions online. Due to the generality of language models, this paradigm could be used to address open-ended tasks. However, it faces challenges in terms of reproducibility, interpretability, and hallucination. In another paradigm, a language model is used to generate executable code policies that are subsequently uploaded for execution onboard robots. A representative method that falls into this paradigm is Code-as-Policy (CaP). Due to the white-box nature of executable code, this paradigm offers high reproducibility and interpretability. Moreover, since executable code usually requires fewer resources than LLMs, this paradigm also enables real-time control on low-cost robot platforms. This is especially relevant for large-scale multi-robot systems, where collective behaviors emerge from robots with exceedingly limited onboard resources. Therefore, this code-policy paradigm is adopted in our work."

"Despite the promise of the code-policy paradigm, the development of control policies for multi-robot systems faces additional challenges compared to single-robot systems. First, the design of policies must consider a robot's interactions with its peers. In some situations, the robot may compete with its peers, for example, for limited resources, whereas in others it may cooperate with its peers to achieve a common goal. Second, the deployment and maintenance of policies require scalable software and hardware systems, which is particularly relevant for multi-robot systems that may have a large number of robots. Third, to maximize the utility of a multi-robot system, it needs to support a wide range of tasks. In addition, some studies proposed frameworks for automated software development such as MetaGPT, ChatDev. Although broadly relevant, these frameworks are not specifically designed for multi-robot systems. Recently, a number of studies explored the use of LLMs for multi-robot systems, but their applicability to general-purpose and real-world multi-robot systems still faces significant hurdles. Of particular relevance is LLM2Swarm, which takes user instructions as input and outputs control policies for individual robots. Although LLM2Swarm is intended to be task-agnostic, its generality is yet to be experimentally verified. Moreover, LLM2Swarm depends on manually-written demonstration examples, restricting its zero-shot capabilities. Other methods such as SmartLLM focus on high-level symbolic planning and do not generate executable low-level control policies. Furthermore, many methods are tailored for specific tasks–such as formation control, cooperative navigation, dancing, or manipulation–and thus lack the generality to address multiple multi-robot tasks. Moreover, the validation in most of the aforementioned methods is performed in simulation, leaving the significant challenge of automated policy deployment on physical multi-robot systems largely unexplored."

"Here, we propose GenSwarm, an end-to-end system that can automatically generate and deploy multi-robot policies on real-world platforms from natural language instructions for versatile multi-robot tasks. GenSwarm enables users to program a group of robots using simple natural language instructions. The user instructions are automatically processed via a pipeline of components, including constraint analysis, policy design, policy generation, policy deployment in simulation environments, policy deployment on real-world robots, and policy improvement based on feedback. These components are respectively empowered by LLM agents. GenSwarm can automatically deploy the generated code policies as well as the required runtime environments on real-world robots, thus achieving true end-to-end functionality. The automatic deployment is realized by a scalable multi-robot platform that features novel software and hardware architectures. GenSwarm enables zero-shot policy generation without the need for context learning based on demonstrative examples. When altered or unseen tasks arise, the system can re-generate and re-deploy policies in response to user requests, thereby offering high adaptability for dynamic tasks. Furthermore, due to the use of code policies, the approach is suitable for real-time execution on robots with limited onboard resources."

"Extensive experiments demonstrate the high success rate of GenSwarm across various multi-robot tasks. GenSwarm consistently outperforms the state-of-the-art methods including MetaGPT, CaP, and LLM2Swarm, achieving significant improvements of 37%, 34%, and 34% in average success rate. GenSwarm provides a promising new paradigm for developing multi-robot systems. Its significance lies in overcoming two limitations of existing work. First, developing multi-robot systems is time-consuming and labor-intensive, and this problem worsens as the number of robots increases. Second, current multi-robot systems lack generality and flexibility. They are often limited to specific tasks or cannot adapt to changing goals and new situations in a timely manner. GenSwarm overcomes these limitations and has the potential to transform the development paradigm of multi-robot systems."