--- title: ghostloop demo emoji: 🤖 colorFrom: green colorTo: indigo sdk: gradio sdk_version: 5.12.0 python_version: "3.12" app_file: app.py pinned: true license: mit short_description: Drive any robot through a fail-closed safety pipeline. tags: - robotics - embodied-ai - agent - mcp - safety - mujoco - ros2 --- # ghostloop · live demo The agent loop, embodied. Tool-using runtime + fail-closed safety pipeline + sim-first execution + post-hoc analysis layer for embodied AI / robotics. Sister project to [GhostLM](https://github.com/joemunene-by/GhostLM). This Space lets you pick a robot profile (Franka arm / Spot quadruped / Tello drone / Stretch mobile arm / humanoid / TurtleBot) and dispatch Intents through the safety pipeline that ships in the library. - **GitHub:** https://github.com/joemunene-by/ghostloop - **PyPI:** `pip install ghostloop` - **arXiv:** _[link to be added once preprint is up]_ ## What you can try here 1. Switch profiles to see how the same Runtime + safety pipeline shape covers totally different morphologies. 2. Send `move_to {"x": 5, "y": 0, "z": 0}` on the Franka profile to watch the GeofenceGate reject it with a structured reason. 3. Send `takeoff {"altitude": 1}` on the Tello profile to see the HITL gate escalate. 4. Read the trace event JSON for any call — that's the same shape the library emits for replay, diff, query, energy ledger, and judge scoring. ## Beyond the demo The full library does much more than this Space exposes: - 6 backends (Mock / MuJoCo / PyBullet / Gymnasium / ROS 2 / Randomized). - 12 policy gates including STL temporal properties. - Counterfactual trace replay, causal failure attribution, LLM-as-judge. - VLA-on-MuJoCo benchmark harness vs OpenVLA / π0 / RT-2 / Octo numbers. - Safe-RL training loop with Lagrangian multiplier + HER. - Production fleet dashboard with auth + rate limit + alarms + Prometheus. - MCP server for Claude Desktop / Cursor / Continue / Cline / Zed / Gemini CLI. `pip install ghostloop` and clone the repo for the full kit.