Described in words, never drawnPass
Indoor Delivery Rover
One paragraph of plain English became a spec, a twin, and a measured PASS.
Drive error
0.398%
Turn error
3.653°
Built with Meta's Muse Spark · Meta Global AI Developer Hackathon
Fix your robot before you spend a dollar.
RobotForge gives robot models a driving test before anyone trusts them. Describe a robot in words, a photo or a parts list — its simulated twin has to pass before you build or buy anything.
How it works
How the driving test works.
- 01
Describe it
Words, a photo, or a parts list.
- 02
Muse Spark writes the spec
The description becomes a spec sheet — numbers only.
- 03
The app builds the twin
The spec becomes a model (URDF/SDF), built in the Gazebo simulator.
- 04
The gate runs the test
The robot drives straight and turns. The gate compares odometry — where it thinks it went — with ground truth: where it actually went.
The verdict
Pass Distance error 5% or less, heading error 5° or less.
Marginal Between 5% and 10%.
Fail Over 10%, or over 10°. The gate names the spec numbers it suspects.
A physically impossible spec — ground clearance bigger than the wheel radius — is refused before the test runs.
Proof, from real runs
TurtleBot3 Burger — PASS. Distance off by 2.293%, heading by 1.421°.
Pathfinder — PASS.
Andino — FAIL. Heading off by 9.104°: its caster wheel drags in a turn.
A robot that has not run the test is marked NOT GATED.
Under the hood: Muse Spark → spec → deterministic URDF/SDF generation → Gazebo → odometry vs ground truth.
Proof
Two recorded runs.
One earned a PASS. One was rigged to fail — and its own sensors couldn't tell.
Fault injectionFail
Wheels 10% smaller than the paperwork
TurtleBot3 Burger with its physical wheels shrunk to 0.9× while odometry kept believing the spec. The robot's own sensors couldn't see it. The gate could.
Drive error
7.617%
Odometry vs truth
11.168%
Bars run 0–15: green passes, yellow is marginal, past it fails. Recorded in Gazebo Sim, headless, Oct 26, 2026.
All gate reports →AI inside
Muse Spark reads, drafts and proposes. It never decides.
Reads
Robot photos, spec sheets and part labels become numbers, each marked ESTIMATED or DOCUMENT-READ.
Drafts
Plain words or a parts list become a full engineering spec.
Proposes fixes
When a check fails, it suggests a change. The checks accept or reject it, up to three rounds.
Not AI, on purpose
The engineering checks, the rover estimates, the URDF and twin, and the Gazebo verdict. Same answer every time.
RobodexRobotForge’s catalogue of real robots, each with its spec, sources and test results.
Real robots, checked the same way.
See a broken robot get fixed in one click.
Watch a broken spec get caught, fixed and re-tested — the whole loop in one run.
