Is manual URDF authoring ever the right choice?

Yes — for a small, one-off robot (a handful of links) that won't change, hand-writing the URDF can be faster than setting up any tooling, and gives full control over every value.

When does automation clearly win?

Larger link counts (errors compound with every manual joint), CAD that changes often (re-deriving axes/inertia by hand each revision doesn't scale), and mate-less inputs like a bare STEP file where there's nothing for a mates-based exporter to read.

Does automated conversion remove the human step entirely?

No — joint limits, travel and which joints are motorized can't be read from a single static CAD pose by any method, manual or automated. A human confirms those either way; automation just removes the axis/inertia/collision hand-authoring.

Automated vs manual CAD-to-URDF conversion

The real question isn't "which is better" — it's which parts of the work are worth automating for your specific robot and workflow.

Compare · ~3 min read

What's actually manual work, either way

Every conversion path — hand-written, sw2urdf-style exporters, or a fully automated tool — needs a human to confirm three things a single static CAD pose cannot reveal: joint travel/limits, which joints are motorized, and mimic relationships. Automation doesn't remove this step; it removes the tedious, error-prone part underneath it.

What automation removes

Per-joint axis placement. Manually, you place a reference coordinate system and axis at every joint. Automated inference (from B-Rep geometry or CAD mates) reads this directly. See why B-Rep geometry makes this exact.
Inertia calculation. Hand-computing a 3×3 tensor per link, or trusting whatever your CAD tool's mass properties happen to export, versus an exact polyhedral-mass algorithm run automatically.
Collision mesh simplification. Manually decimating or primitive-fitting every link's collision geometry, versus automatic convex decomposition.

A practical breakdown

Situation	Lean manual	Lean automated
Link count	Under ~5, simple	Growing, or already double digits
CAD input	Native assembly with clean mates	Bare STEP/mesh with no mates
Iteration	One-off, CAD won't change	CAD revises often — re-deriving by hand each time doesn't scale
Output needs	URDF only	Also need MJCF/USD/SDF for other simulators

Jointly is built for the right-hand column: STEP, mesh, SolidWorks or Onshape in, a simulation-ready URDF/SDF/MJCF/USD package out, with the human step limited to confirming limits and motorization — not re-deriving axes and inertia from scratch.

Skip the manual work

Jointly does everything on this page automatically: drop in your CAD (STEP, mesh, SolidWorks or Onshape), and it infers joints, axes, inertia and collision, then exports a simulation-ready URDF, SDF, MJCF or USD. The first conversions are free.

Try Jointly free →