A desktop where your robot sim and your LLM live together

What it does Lyapunov is an LLM-first, 3D-native robotics desktop framework: scene building, robot preparation, controlled simulation, and camera data collection all share one workspace alongside chat, file editing, Git inspection, and a terminal. You can drive scenes, robots, and crowds either through workbench controls or by asking the agent in natural language. It ships with MuJoCo as its default physics engine, with Isaac Sim and Newton as optional add-ons.

The interesting bit It builds on DeepSeek Harness’s “everything-is-a-plugin” architecture and Cordis’s composition mechanism, so tools like Blender, extra engines, or policy runtimes plug in rather than get bolted on. The robot workflow is deliberately explicit — prepare, initialize, ready, execute — which keeps the LLM from freelancing with your hardware.

Key highlights

• Scene import for GLB, Gaussian representations, and native robot documents; USD/SimReady assets arrive with layers, materials, and physics preserved.
• Camera tooling includes RGB-D capture, calibration, annotations, recording, and dataset export, subject to engine support.
• Bundled crowds: 12 Rocketbox characters with route presets and per-group start/stop control.
• Guest mode runs a fully local workspace with no product servers, billing, or default model — bring your own provider.
• Runs on Linux x64, macOS Apple Silicon, and Windows/WSL2; Apache-2.0.

Caveats

• The signed-in path uses the vendor’s Peiri model and a Credits billing account; the README doesn’t detail what guest mode costs you in agent capability beyond “no default model.”
• The macOS arm64 package is not notarized, and desktop automation input is unreliable on the tested Wayland session — X11 is recommended.
• The Linux archive is roughly a gigabyte, and the README’s workbench screenshot is still a placeholder.

Verdict Worth a look if you build robot simulations or collect camera datasets and want an agent in the loop without leaving the IDE-like workspace. If you wanted a lightweight headless sim library, this is very much not that — it’s a full commercial-backed desktop product.