erwinmsmith/SOMAS

A Trusted Human-Multi-Agent Reinforcement Learning Interaction Framework

What it solves

SOMAS addresses the challenge of ensuring safety and reliability in human-machine collaborative crisis response. It prevents risky AI actions in emergency scenarios by balancing task utility with strict safety constraints.

How it works

The framework uses a dual-mode architecture combining vision-language models and reinforcement learning:

  • Online Execution System: Manages real-time tasks through a planning-execution pipeline with modular task chains and safety guardrails, including GPT-4-based risk assessment.
  • Offline Simulation System: Uses synthetic task generation and an experience replay library to train and optimize reinforcement learning policies for risk prediction.

Who it’s for

Emergency response teams and developers building multi-agent systems for high-stakes, safety-critical environments where human oversight is required.

Highlights

  • Dual-Mode Architecture: Combines real-time execution with offline simulation training.
  • Safe LLM: Includes a fine-tuned LLM and training dataset specifically for emergency scenarios.
  • Proven Performance: Reduces risk response rates by 40% and increases helpfulness by 15% over baselines.

Related

  • Project
    TsinghuaC3I/MARTI

    An open-source framework for training LLM-based multi-agent systems using reinforcement learning, featuring tree search augmentation for complex reasoning and code generation.

  • Project
    liquidos-ai/AutoAgents

    A production-grade multi-agent framework in Rust that provides a type-safe environment for building intelligent systems with structured tool calling and pluggable LLM backends.

  • Project
    RasaHQ/rasa

    An open-source machine learning framework for building contextual AI assistants and agents that automate text and voice-based conversations.

  • Project
    hello-diana/MASCOT

    A multi-agent framework for building socio-collaborative AI companions that prevents persona collapse and social sycophancy through bi-level optimization of individual speakers and a coordinating director.

  • Project
    kyegomez/swarms

    A production-ready multi-agent orchestration framework that provides pre-built architectures for coordinating multiple AI agents to solve complex tasks.