CAD to MuJoCo and Isaac Sim

Whether you're doing reinforcement learning in MuJoCo or GPU-scale sim in Isaac, the hard step is turning CAD into a physics-ready robot. Here's the pipeline and the shortcut.

Guide · ~5 min read

Modern robot learning runs on two simulators more than any others: MuJoCo (fast, accurate contact dynamics, popular for RL) and NVIDIA Isaac Sim (GPU-parallel, USD-based, popular for scaled training and sim-to-real). Both need your robot in their own format — and both start from the same place: a correct URDF.

The formats

• URDF — the ROS-native robot description. The common intermediate everything converts from.
• MJCF — MuJoCo's XML. MuJoCo can load URDF directly, but MJCF exposes actuators, tendons, contact parameters and solver settings you'll want for RL.
• USD — OpenUSD, the format Isaac Sim is built on. Robots are imported/authored as USD stages.

The pipeline

1. CAD → URDF. Recover links, joints, axes, inertia and collision from your CAD. This is the step that determines whether the sim is any good.
2. URDF → MJCF for MuJoCo, or URDF → USD for Isaac. MuJoCo ships a URDF loader and a compiler; Isaac Sim ships a URDF importer that writes USD.
3. Tune the sim. Add actuators, set friction and contact parameters, and define the observation/action spaces for RL.

Why step 1 dominates sim-to-real gap

If the URDF's inertia is wrong or its joint axes are off, no amount of solver tuning saves you — the simulated dynamics simply don't match the real robot. Accurate mass properties and correct axes are the cheapest large win for closing the sim-to-real gap. (See fixing URDF inertia and collision.)

The fast path

Jointly converts your CAD — STEP, mesh, SolidWorks or Onshape — and exports URDF, MJCF and USD directly, with correct inertia and convex collision and meshes packaged in a ZIP. Instead of chaining exporters and importers and debugging each one, you get all three formats from one conversion, ready to load into MuJoCo or Isaac.

Checklist before training

• Robot settles under gravity with no actuation (stable inertia + collision).
• Every joint moves on the correct axis with sensible limits.
• Masses and CoM match the real hardware within a few percent.
• Collision geometry is convex/simplified, not the raw visual mesh.