SwarmLLM-Bench

SwarmLLM-Bench: LLM-Enhanced Multi-Agent Cooperative Control Simulation Platform SwarmLLM-Bench：大语言模型驱动的多智能体协同控制仿真平台

https://github.com/user-attachments/assets/f81fa67c-1477-4ca0-8559-a341273e5de0

Python Version Framework License Paper

Highlights | 核心亮点 [English] SwarmLLM-Bench is an open-source simulation platform bridging the gap between Large Language Models (LLMs) and physically constrained multi-agent systems. Unlike traditional text-only agents, our platform integrates LLMs into the OODA (Observe-Orient-Decide-Act) loop for real-time UAV swarm coordination. Key innovations include: FSM + LLM Hybrid Harness Architecture: Ensures 100% safety with deterministic FSM fallback while leveraging LLM for emergent tactical reasoning. 4-Layer Safety Harness: Strict validation of Role, State, Confidence, and Safety Rules to prevent LLM hallucinations in physical environments. OODA-Driven Emergence: Quantifiable enhancement in decision density (27x OODA loops) and survival rate (97%) in adversarial scenarios. High-Fidelity Physics: Built on Gymnasium, simulating realistic aerodynamics, battery drain, and communication delays. [中文] SwarmLLM-Bench 是一个旨在弥合大语言模型（LLM）与物理约束多智能体系统之间鸿沟的开源仿真平台。与传统的纯文本智能体不同，本平台将 LLM 深度集成到 OODA（观察-调整-决策-行动）循环中，用于实时无人机蜂群协同。核心创新包括： FSM + LLM 混合 Harness 架构：在利用 LLM 进行涌现式战术推理的同时，通过确定性 FSM 兜底，确保物理执行的绝对安全。 四层安全护栏（Harness）：对角色、状态、置信度和安全规则进行严格校验，彻底杜绝 LLM 在物理环境中的幻觉与危险指令。 OODA 驱动的战术涌现：在对抗场景中，决策密度（OODA循环频次）提升 27 倍，无人机平均存活率高达 97%。 高保真物理仿真：基于 Gymnasium 构建，真实模拟空气动力学、电量消耗及通信延迟等物理约束。 ️ Architecture | 系统架构 [English] The platform adopts a 3-layer decoupled architecture: DeepAgent Layer (Cognitive): Handles high-level intent parsing and long-horizon tactical planning via LLM. HybridUAV Layer (Physical): Executes real-time flight control under strict physical constraints. LLMClient Layer (Bridge): Responsible for bidirectional mapping (Symbolization & Grounding) and enforcing the 4-layer safety harness. [中文] 本平台采用三层解耦架构设计： DeepAgent 层（认知大脑）：通过 LLM 负责高层意图理解与长程战术规划。 HybridUAV 层（物理实体）：在严格的物理约束下执行实时飞行控制。 LLMClient 层（语义-物理桥梁）：负责双向映射（符号化与物理落地），并强制执行四层安全护栏机制。 (建议在此处插入你的架构流程图)

Quick Start | 快速上手 [English] Get up and running in 3 steps: bash

编辑

1. Clone the repository

git clone https://github.com/[YourUsername]/SwarmLLM-Bench.git cd SwarmLLM-Bench

2. Install dependencies

pip install -r requirements.txt

3. Run the simulation

python main.py –config configs/mvp.yaml [中文] 只需 3 步即可运行仿真： bash

编辑

1. 克隆仓库

git clone https://github.com/[YourUsername]/SwarmLLM-Bench.git cd SwarmLLM-Bench

2. 安装依赖

pip install -r requirements.txt

3. 启动仿真

python main.py –config configs/mvp.yaml Benchmarks | 性能基准 [English] Comparison between Pure FSM Baseline (P1) and LLM-Driven Swarm (P3): 表格 Metric P1 (Pure FSM) P3 (LLM-Driven) Improvement Mission Success Rate 0% 88% +88% UAV Survival Rate 12.5% 97% ️ +84.5% OODA Loops / Run 26 706 27x [中文] 纯规则基线（P1）与 LLM 驱动蜂群（P3）的核心指标对比： 表格 评估指标 P1 (纯 FSM) P3 (LLM 驱动) 提升幅度 任务成功率 0% 88% +88% 无人机存活率 12.5% 97% ️ +84.5% OODA循环频次 26 706 27倍 Citation | 引用 [English] If you find SwarmLLM-Bench useful in your research, please consider citing: bibtex

编辑

@article{swarmllm2024, title={SwarmLLM-Bench: LLM-Enhanced Multi-Agent Cooperative Control Simulation Platform}, author={[Your Name]}, journal={arXiv preprint}, year={2024} } [中文] 如果 SwarmLLM-Bench 对您的研究有所帮助，请考虑引用： bibtex

编辑

@article{swarmllm2024, title={SwarmLLM-Bench：大语言模型驱动的多智能体协同控制仿真平台}, author={[张杰]},

https://github.com/user-attachments/assets/370754e2-b71a-4d9b-8743-79b3f2c3024e

journal={arXiv preprint}, year={2024} } Contributing | 参与贡献 [English] Contributions are welcome! Please read our Contributing Guidelines before submitting a Pull Request. [中文] 欢迎提交 Issue 和 Pull Request！参与贡献前请先阅读 贡献指南。 License | 开源协议 This project is licensed under the MIT License. See the LICENSE file for details. 本项目基于 MIT 协议开源，详情请参阅 LICENSE 文件。