urdf-to-mjcf

urdf-to-mjcf

Wrap scripts/convert_urdf_to_mjcf.py, which prefers K-Scale Labs urdf2mjcf and falls back to MuJoCo's built-in URDF loader when that package is unavailable. Then post-tune the output because raw conversion is rarely usable as-is.

Simulation boundary

Treat this skill as a Level 0/1/1.5 generator, not a calibrated digital-twin generator:

• Level 0: loadable. XML compiles and opens in MuJoCo.
• Level 1: controllable-ish. Joints have plausible starter actuators and can be driven in the viewer.
• Level 1.5: geometry-informed. CAD/URDF provides credible link frames, collision/visual meshes, mass, and inertia.
• Level 2: calibrated. Requires real hardware or vendor data for torque-speed curves, controller gains, friction, backlash, armature, limits, and contact parameters. Do not invent this from geometry.

The wrapper prints simulation_level, actuator_source, and calibration_source. If those say generic_defaults, template_estimate, or manual_override, the model is a starting point for simulation, not a validated physical model.

The script gets the kinematic tree right but does not handle:

1. Mimic joints get silently dropped. MuJoCo's URDF loader has no way to express URDF <mimic> semantics. All mimic'd joints become independent free joints; the underlying coupling is lost.
2. Inertias are often 100x oversized. Stock URDFs from xacro/SDF pipelines frequently use placeholder inertia values that produce absurd dynamics in MuJoCo (e.g. fingers that won't move at any reasonable kp).
3. No calibrated actuators. URDF has joints but no MuJoCo motor or controller model. The wrapper can add starter position actuators, but those are estimates unless supplied from hardware data.
4. No contact excludes. Adjacent links that touch in the rest pose (palm↔first-knuckle pairs are the typical case) get treated as active contacts and the joint physically can't reach its commanded angle.
5. Converter schema quirks. Known urdf2mjcf outputs can put scale on <geom> instead of <mesh>, add a fixed-arm <freejoint>, omit required inertial pos, and keep non-portable file:// paths.
6. No autolimits, no implicitfast integrator. Defaults that matter for stable dexterous-hand sim aren't set.

Strong preference: check mujoco_menagerie first. If your robot or hand already lives in https://github.com/google-deepmind/mujoco_menagerie, adopt that MJCF verbatim — the MuJoCo team has already tuned it, and re-running the converter on a URDF will produce a strictly worse result.

These post-tune fixes are documented in references/gotchas.md and the step-by-step is in references/workflow.md.

When to use this skill

• A new robot/hand with a URDF you need to simulate in MuJoCo and no upstream MJCF (menagerie or otherwise) exists.
• Your existing URDF is the source of truth (e.g. you maintain it for RViz/pinocchio) and you need an MJCF for sim.

When NOT to use this skill

• An upstream MJCF exists in mujoco_menagerie or the manufacturer's repo — adopt it instead. Re-tuning is wasted work.
• You only need URDF (e.g. for RViz/pinocchio) — skip MJCF entirely.
• The robot is going to be combined with another via attach-end-effector and only that combined model needs to simulate — convert the components individually first.

Workflow

1. Check mujoco_menagerie first. If the robot is there, copy its MJCF into robots/<robot>/<robot>.xml and stop. Don't run the converter.
2. Make xacro/URDF paths portable. Replace ROS $(find pkg) and machine-local file:///home/... mesh paths with relative paths before conversion.
3. Verify the URDF loads in MuJoCo standalone. If MuJoCo can't parse it, the converter won't help — fix the URDF first.
4. Run the converter wrapper (see Commands). It applies deterministic fixes for scale placement, file URIs, missing inertial pos, fixed-base freejoints, compiler/option defaults, and basic joint dynamics.
5. Classify the metadata source. Decide whether this is Level 0, Level 1, or Level 1.5. Only call it Level 2 if actuator and dynamics parameters came from real hardware/vendor data and were validated against motion.
6. Post-tune the MJCF. Walk through the fix categories in references/gotchas.md: mimics→<equality>, inertias, actuators, contact excludes, mesh-origin problems, and tuned dynamics.
7. Verify in the viewer. Each actuator slider should drive the intended joint(s) to the commanded angle without numerical blowup or contact-blocked motion.

Commands

Run from the bundle root with the bundle-local venv. The wrapped script ships with this skill. Pass an explicit output path:

cd robot-assets-skills/
./.venv/bin/python .claude/skills/urdf-to-mjcf/scripts/convert_urdf_to_mjcf.py \
    robots/<robot>/<robot>.urdf \
    robots/<robot>/<robot>.xml

For an SO101-style arm using STS3215-class servos as a starter template, use the explicit profile. It is still an estimate:

./.venv/bin/python .claude/skills/urdf-to-mjcf/scripts/convert_urdf_to_mjcf.py \
    robots/<robot>/<robot>.urdf \
    robots/<robot>/<robot>.xml \
    --control-profile so101-sts3215

For a new CAD arm with unknown hardware, stay with the generic profile or override parameters explicitly:

./.venv/bin/python .claude/skills/urdf-to-mjcf/scripts/convert_urdf_to_mjcf.py \
    robots/<robot>/<robot>.urdf \
    robots/<robot>/<robot>.xml \
    --position-kp 15 \
    --position-force 3.35 \
    --joint-damping 0.6 \
    --joint-armature 0.028 \
    --joint-frictionloss 0.052

If using the vendored converter from pathonai_diy_pipeline without installing it into .venv, the wrapper checks this default path:

/home/aidy/Projects/pathonai_diy_pipeline/scripts/urdf2mjcf

Override it when needed:

URDF2MJCF_VENDOR_PATH=/path/to/scripts/urdf2mjcf \
./.venv/bin/python .claude/skills/urdf-to-mjcf/scripts/convert_urdf_to_mjcf.py \
    robots/<robot>/<robot>.urdf \
    robots/<robot>/<robot>.xml

Verify the converted MJCF compiles:

./.venv/bin/python -c "
import mujoco
m = mujoco.MjModel.from_xml_path('robots/<robot>/<robot>.xml')
print('nq=', m.nq, 'nu=', m.nu, 'nbody=', m.nbody)
"

If the wrapper falls back to MuJoCo's built-in loader, raw conversion would produce nu = 0; the postprocess step adds starter position actuators where possible. If K-Scale urdf2mjcf runs, generated motors are converted to position actuators by default unless --keep-motors is passed.

View interactively to drive the sliders (set $DISPLAY first if headless — see AGENTS.md):

./.venv/bin/python -m mujoco.viewer \
    --mjcf robots/<robot>/<robot>.xml

If .venv/ is missing or broken, bootstrap per AGENTS.md's "Python environment" section, then install this skill's requirements.

References

• Step-by-step procedure: references/workflow.md
• The five post-tune fix categories with concrete recipes: references/gotchas.md