FREE · OPEN SOURCE · PRE-RELEASE
Software for hardware.
The script is the model. The model is the robot. The agent writes the script.
Design the concept with an agent
One chat turn. The AI has one way to change the model: it edits the script.
The script is the model
- tilt θ
- 0.0°
- rate ω
- 0°/s
- policy
- none
Every num() is a slider. Drag it. The model rebuilds with no AI in the loop.Exported to MuJoCo. No controller yet, so gravity wins every time.
The task is data, in the same script
| reward | alive | +1.0 | |
upright abs(θ) | −3.0 | ||
settle abs(ω) | −0.4 | ||
stay put abs(x) | −0.012 | ||
| effort | −0.0004 | ||
| terminate | θ beyond 45° | ||
| disturb | horizontal push, 0.4 to 3 N, twice per episode | ||
Observations, rewards, terminations and pushes are declared next to the geometry they act on.
Train on any GPU box
- iteration
- 0 / 120
- envs
- 16
- mean reward
- 0
- episodes
- 0
The bundled PPO trainer runs offboard. One policy file comes back, and the viewport plays it.
Evaluate and adapt the results
Geometry, physics and control live in one file, so a design change is one diff and one retrain.
What Cadex is
Cadex is an AI-native CAD application, free and open source. You describe a part in chat and the agent writes a parametric Python script. The script is the model: every declared dimension becomes a slider, every rebuild is exact solid geometry from the OCCT kernel, and it all lives in a Blender-class viewport with chat, sliders, model tree and script beside it.
Assemblies export to MuJoCo with exact inertias, so a mechanism falls, collides and is actuated the moment it is built. The control task is declared in the same script as the geometry: observations, rewards, terminations, disturbances. The bundled trainer runs on any GPU box and returns one policy file that plays back in the viewport. Change your mind, and the same loop runs again. A headless CLI drives the same engine for parameter sweeps, exports, renders and training runs, and spends no tokens after the first turn.