Safety-Oriented Multi-agent System

A Trusted Human-Multi-Agent Reinforcement Learning Interaction Framework

---

📖 Introduction

This repository implements a Multi-Agent System (MAS) framework for human-machine collaborative crisis response, combining vision-language models (VL) and reinforcement learning (RL) to enhance safety and reliability. The framework features:

• Real-Time Task Execution: Modular task chains with built-in safety rules and human oversight.
• Simulation Training: Experience replay library for risk prediction and optimization.
• Dynamic Trust Mechanism: Balances task utility and safety constraints through RL.

Key Contributions:

1. Dual-mode architecture (online execution + offline simulation).
2. First fine-tuned safe LLM and training dataset for emergency scenarios.
3. 15% improvement in helpfulness and 40% reduction in risk response rate compared to baseline.

---

🚀 Quick Start

Installation

git clone https://github.com/erwinmsmith/SOMAS.git
cd SOMAS
pip install -r requirements.txt

Usage

1. Real-Time Task Execution:

main.py --online

2. Simulation Training:

main.py --offline

---

🧠 Framework Architecture

Core Components

1. Online Execution System

   • Planning-Execution Pipeline: Modular task chains drive tool operations.
   • Safety Guardrails: Predefined rules and GPT-4-based risk assessment.

2. Offline Simulation System

   • Task Generation: Synthetic tasks from manual records and prior knowledge.
   • Experience Replay: Optimizes RL policies for dynamic environments.

---

📊 Experimental Results

Key Metrics

| Domain | Model | Safety (↑) | Helpfulness (↑) | Risk Response Rate (↓) | |-------------|--------------------|------------|-----------------|-------------------------| | Safety-CV | Qwen2-7B-VL | 4.5 | 4.7 | 40% |

Highlights

• VL models reduced operational risks by 30% via image semantic parsing.
• Dynamic safety validation improved helpfulness by 15% over ToolEmu.

---

Others

If you need a detailed data for Safty(train or test), contact me duanzhenke@sscapewh.com/duanzhenke03@gmail.com