What does this skill do, and when should you use it?
This is the URDF skill from the earthtojake/text-to-cad repository, which bundles 13 skills total. It treats URDF work as constrained kinematic modeling rather than XML writing: a frame/joint/geometry/unit/assumption ledger must be established before editing, and derived numbers like inertia tensors must be computed, never freehanded. Every file is validated with cadgen urdf validate and visually reviewed joint-by-joint in the local CAD Viewer or via snapshots. It fits developers targeting RViz, Gazebo, MoveIt or real robot drivers.
- Authors and edits .urdf XML directly (no code-generation pipeline), embedding a design-ledger comment block at the top of the file
- Prepares mesh assets: one mesh per link, exported in that link's frame
- Computes derived numbers (inertia tensors, centers of mass, unit conversions) via closed-form formulas or throwaway helper scripts
- Runs cadgen urdf validate, collecting all findings in one pass across XML structure, tree topology, joint semantics, geometry, mesh references, materials and inertial physics
- Renders PNG stills with cadgen urdf snapshot, posable via --joint-values
- Launches the CAD Viewer (cadgen viewer --detach) to review mesh scale, placement, and sweep every movable joint
- A robotics engineer building a robot model for Gazebo/Ignition simulation
- A ROS developer needing correct link/joint frames for robot_state_publisher and RViz
- A team preparing a URDF baseline before MoveIt integration (semantic groups belong to the SRDF skill)
- A mechanical engineer exporting link meshes from CAD who needs to verify mesh scale and placement
- Anyone debugging frame conventions or inertial data in an inherited .urdf file
- Users who just want text-to-3D model generation (STEP/STL) — that is the sibling CAD skill's job, not this one
- Users needing MoveIt planning groups or IK semantics — SKILL.md explicitly redirects those to the SRDF skill
- Offline-restricted environments without uv or network access to download the cadgen runtime
How do you install this skill?
- Telemetry is on by default (usage stats and crash reports tagged with a random install ID); it never sends file contents and can be disabled, but run `uvx cadgen telemetry off` or set CADGEN_TELEMETRY=0 immediately after install if you care.
- The CAD Viewer binds to 127.0.0.1 but serves unauthenticated: any local process can read files under the opened directory and trigger builds; never pass a non-loopback --host.
- The first run downloads cadgen==0.7.20 and a headless-browser snapshot from PyPI via uvx, which may be unreachable or slow from mainland-China networks; Windows 11 Smart App Control (on by default on fresh installs) blocks the unsigned OCP native module and breaks every cadgen command.
- The skill itself states that structural validation cannot prove spatial correctness (joint origins, axis signs, mesh scale, inertials); outputs still require per-joint visual review against the ledger, and inertials from assumed density are placeholders.
- The publisher is unverified, the skill has no independent version/changelog, and this static review executed nothing, so real-world usability carries uncertainty.
- Shell / CLI
- Network access
- Local filesystem
uv (uvx)Python 3.13 via uvcadgen==0.7.20headless browser for snapshotsGit LFS for mesh repos
This skill ships as part of the text-to-cad collection (13 skills), installed with the whole collection. Recommended: ask your agent to install it, or pick your host:
Claude Code
claude plugin marketplace add earthtojake/text-to-cad#latest
claude plugin install text-to-cad@earthtojakeCodex
codex plugin marketplace add earthtojake/text-to-cad --ref latest
codex plugin add text-to-cad@earthtojakeCursor
git clone --depth 1 --branch latest https://github.com/earthtojake/text-to-cad ~/.cursor/plugins/local/text-to-cadGrok Build
grok plugin install earthtojake/text-to-cad@latest --trust
grok plugin enable text-to-cadGemini
gemini extensions install https://github.com/earthtojake/text-to-cad --ref latest --consent --auto-updateOther agents (Skills CLI)
npx skills add earthtojake/text-to-cad#latestPrerequisites: uv must be installed; on Windows 11 with Smart App Control enabled, the unsigned OCP native module will be blocked (README suggests turning it off or using WSL).
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Write a URDF file for this four-wheeled robot with links, joints, limits and inertials, and make sure it passes validation
- Check my_robot.urdf for joint axis and frame issues, fix them, re-validate, and show me a rendered snapshot
- Add mesh references for these links to the URDF, verify scale and placement, then pose every joint in the CAD Viewer
- Compute the inertia tensors and centers of mass for each link in this URDF — don't hand-write numbers — then run a strict validation
The skill is triggered automatically when a request involves creating, editing, inspecting, validating or debugging .urdf files. The workflow: identify the target file and its consumers (RViz, Gazebo, MoveIt, etc.) → read or create the design ledger → prepare meshes → author URDF XML directly → compute inertials and derived values → validate with cadgen urdf validate → verify with external tools (check_urdf) and a joint-by-joint viewer sweep → report remaining assumptions. Key commands:
cadgen urdf validate path/to/robot.urdf --strict
cadgen urdf validate path/to/robot.urdf --
cadgen urdf validate path/to/robot.urdf --packages robot_description=/path/to/pkg
cadgen urdf snapshot path/to/robot.urdf review.png--strict treats warnings as failures; -- emits structured findings (severity, code, XML path, hint); --packages resolves package:// mesh URIs and can be repeated. Pose snapshots with --joint-values (JSON, in degrees); meshes from Git LFS repos must be checked out first (git lfs checkout) or the snapshot fails.
What are this skill's strengths and limitations?
- Rigorous methodology: the design ledger, frame-semantics references, and compute-don't-guess rules target URDF's most common failure points directly
- The validator collects all findings in one pass — structure, topology, joint semantics, geometry, meshes, materials, inertials — with a -- output for tooling
- Snapshots use the same rendering runtime as the Viewer, so what you see is consistent; unloadable link meshes fail the snapshot rather than being silently omitted
- MIT licensed, with active maintenance (CI tests, docs site)
- Heavy dependency chain: uv, a pinned cadgen==0.7.20, Python 3.13 via uv, plus a headless browser for snapshots
- Validation is a guardrail, not spatial proof — a joint placed in the wrong spot can still pass every structural check, so manual viewer sweeps are mandatory
- Windows 11 Smart App Control blocks the unsigned OCP native module, requiring disabling it or running under WSL
- Telemetry is on by default (anonymous, never file contents or prompts); you must opt out explicitly with uvx cadgen telemetry off
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| URDF Robot Description Skill this page | 57 · Use with care | ★ 19k | 1d ago | MIT |
| SRDF Skill: MoveIt Planning Semantics | 62 · Recommended | ★ 19k | 1d ago | MIT |
| text-to-cad SDF Skill | 55 · Use with care | ★ 19k | 1d ago | MIT |
| I4H Workflow Environment Creator ✓ NVIDIA · Official | 50 · Use with care | ★ 3.5k | 3d ago | Apache-2.0 |
| Isaac for Healthcare Dataset Replay ✓ NVIDIA · Official | 48 · Use with care | ★ 3.5k | 3d ago | Apache-2.0 |
The source explicitly delineates sibling skills in the same repo: the SRDF skill handles MoveIt planning groups and IK/path-planning semantics, the CAD skill handles STEP/STL/3MF/DXF/GLB outputs, and the SDF skill handles simulator models and worlds. This skill covers only the URDF description itself.
How did FollowSkills review this skill?
Data flow is transparently documented: telemetry is on by default but extensively disclosed with multiple opt-out mechanisms (cadgen telemetry off, DO_NOT_TRACK, CADGEN_TELEMETRY=0) and explicit claims that no file names, paths, contents or prompts are sent; the viewer binds only to 127.0.0.1 and SECURITY.md names loopback as the trust boundary. Deductions: telemetry defaults to on rather than off; the viewer serves unauthenticated so any local process can read opened directories; runtime pulls pinned cadgen==0.7.20 plus a headless-browser snapshot from PyPI via uvx, an external supply chain with no hash-locked verification; publisher identity is unverified; no explicit user-confirmation gate or rollback mechanism.
Documentation is highly self-consistent: workflows, commands, and failure modes (unhydrated Git LFS pointers, Smart App Control blocking OCP, viewer launch failure, snapshots failing on unloaded link meshes) are all explicit, and validator behavior (--strict, --, nonzero exit) is fully described. Deductions: static review cannot execute key paths; the first run requires network downloads from PyPI and a browser snapshot, so the happy path is not offline-capable; references are complete but there is no reproducible test evidence specific to the URDF skill paths; cadgen doctor was not run to confirm installation consistency.
Trigger conditions are precise (frontmatter description enumerates .urdf creation, editing, inspection, validation, debugging), target consumers (RViz, Gazebo, MoveIt, real drivers) and non-fit boundaries (SRDF/CAD skill boundaries, floating/planar joints) are declared, and the verification limits (validation cannot prove spatial correctness) are candidly disclosed. Deductions: no Chinese-language support declared; the core runtime depends on first-time uvx/PyPI downloads that may be unreachable or slow from mainland-China networks; host-tool availability (cad_show) is conditional and environment-fit evidence is thin.
Information architecture is excellent: SKILL.md plus seven references with progressive disclosure, a golden skeleton example, design-ledger templates and assumption-recording rules; MIT license and provenance link are clear; cadgen==0.7.20 is version-pinned with cadgen doctor for drift detection; README provides update/reinstall paths. Deductions: the skill itself has no independent version number or changelog; publisher is unverified and maintenance responsibility is inferred only from the repository link; key limitations (telemetry, Smart App Control) live in the shared README rather than the skill's own docs.
The claimed output is a directly usable .urdf file (single source of truth) backed by validator, check_urdf, and viewer sweep, with clear marginal value over hand-authoring (inertia formulas, mesh-scale rules, frame-semantics checklists). Deductions: static review did not verify that representative outputs are directly usable; validator and viewer sweep require installation and a first network download; the docs themselves admit validation does not prove spatial correctness, so complete delivery still requires human visual review across multiple steps.
The repository shows a CI test-workflow badge and committed test-model directories (models/tests/). Deductions: the provided fixtures target viewer/rendering, not the URDF skill's key paths; this static review could not reproduce execution; external tools (check_urdf) are only conditional 'when available'; evidence comes from a single repository with no independent third-party corroboration, so the verifiability ceiling is not met.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →