RagLeap Physical: A Simulation-First Safety Bridge for AI Agents

What should happen when an AI agent proposes an action that affects a physical-system simulation?

At RagLeap, we're exploring a controlled workflow:

AI proposes → Human approves → Safety limits are checked → Simulator verifies

RagLeap Physical is an open-source, pre-alpha Python project built around this idea.

🔗 GitHub: https://github.com/ragleap/ragleap-core/tree/main/packages/ragleap-physical

What it does

RagLeap Physical provides a simulation-first bridge between AI agents and physical-system simulators.

The project includes:

• Safety envelopes
• Exact-proposal approval
• Dry runs
• Runtime monitoring
• Stop functionality
• Audit logs
• AI-facing tools
• MCP server
• Optional MuJoCo adapters
• Workspace limits
• Runnable simulation demos

The focus is not unrestricted AI control. It is controlled, observable interaction.

Release journey

v0.0.1 — Core bridge

The first release introduced the safety bridge, exact-proposal approval, runtime monitoring, audit logs, MCP tools, MuJoCo adapters, workspace limits, and simulation examples.

v0.0.2 — Operator console

The second release added a human operator console.

A named person can review pending proposals, approve the exact digest, apply the action, and stop.

The console is loopback-only, uses one token per person, and is not an AI tool.

v0.0.3 — Two-link workspace demo

The latest release adds:

examples/two_link_workspace_demo.py

The demo shows a two-link arm with one accepted workspace move and one refused move.

It also includes package metadata improvements and a fix for request handling in the operator console and MCP server.

Install

pip install ragleap-physical==0.0.3

Optional extras:

pip install "ragleap-physical[mujoco]"
pip install "ragleap-physical[test]"
pip install "ragleap-physical[interop]"

Testing

The v0.0.3 release reports:

• 255 tests passed with MuJoCo and the MCP SDK installed in Colab
• GitHub CI on Python 3.10, 3.11, and 3.12

The Windows refusal-fix behavior is not yet verified.

Important limitations

This is pre-alpha and simulation-only.

It does not control real hardware. No real-robot or real-machine validation, safety certification, or regulatory compliance is claimed.

MuJoCo results are repeatable within one installed environment only. The wear-study and machine-tool numbers are invented.

Contribute

We welcome feedback and contributions from developers working on AI agents, simulation, robotics, MCP, and safety engineering.

Explore the package:

https://github.com/ragleap/ragleap-core/tree/main/packages/ragleap-physical

Install:

pip install ragleap-physical==0.0.3

AI proposes. Humans approve. The system verifies.