GitHub
Under the hood

What a PASS proves.

The test itself is explained once, on the home page: how the driving test works. This page keeps the fine print.

What a PASS proves

  • The spec's numbers agree with each other: geometry, mass budget, voltages.
  • The twin's odometry matched ground truth within the locked thresholds.

What it doesn't prove

  • That a physical robot will behave the same. The gate tests the spec's simulated twin, not hardware.
  • That a photo-drafted spec is accurate. Photo dimensions are marked ESTIMATED until measured.
  • Anything beyond drive and turn. Other behaviours aren't gated yet.

Every recorded gate report →

Run it yourself

Python, zero required dependencies for the core
pip install -e .
robotforge describe "a four-wheel rover, 30 cm long, with a front ultrasonic sensor" --mock --out rover.json
robotforge validate rover.json
robotforge urdf rover.json --out build/rover.urdf
robotforge run-gate rover.json

AI inside: where Muse Spark works

Each place a Meta model works, which model it is, and whether that step is live right now.

What Muse Spark doesModelWhereNow
Plain words → engineering specMuse Spark Open (Baseten, Fireworks failover)Forgechecking…
Parts list → assembled draftMuse Spark Open (Baseten, Fireworks failover)Forgechecking…
Robot photo or spec-sheet photo → draft (ESTIMATED / DOCUMENT-READ)Muse Spark vision (muse-spark-1.3, Meta Model API)Forge, My robotschecking…
Failed check → proposed fix, up to 3 roundsMuse Spark Open (Baseten, Fireworks failover)Demo, My robotschecking…
Part photo → identified part, label values readMuse Spark vision (muse-spark-1.3, Meta Model API)My partschecking…
Confirm each part in a chatMuse Spark Open (Baseten, Fireworks failover)My partschecking…
Forge Guide follow-up questionsMuse Spark Open (Baseten, Fireworks failover)Every app pagechecking…

Not AI, on purpose

The model proposes. These decide, the same way every time:

  • Engineering checks (geometry, mass, electrical, drivetrain)
  • Accepting or rejecting every model-proposed fix
  • Rover slope, step and runtime estimates
  • URDF and Gazebo twin generation
  • The Gazebo motion gate verdict, measured against ground truth

Plain English

The words on this site, without the jargon. Dotted words anywhere on the site show the same meaning on hover or tap.

Spec
A short structured description of a robot: size, mass, motors, battery, wheels. Every check reads it.
Engineering checks
Fixed rules that test whether a spec’s numbers fit together, like wheels big enough to lift the chassis. Same answer every time; no AI.
Digital twin
A simulated copy of the robot, built from its spec, that can drive in a physics simulator before any hardware exists.
Gate (motion test)
A pass/fail test in the Gazebo simulator: the twin drives 1 m, turns 90° and stops, and the result is measured against where it really went.
Gazebo
Open-source robot physics simulator. RobotForge runs the gate in it, headless.
URDF
The standard file format ROS uses to describe a robot’s parts and joints. RobotForge exports one from every spec.
Harness (electrical check)
The robot’s wiring. The harness check adds up current on each power rail and flags wires that are too thin or rails that are overloaded.
Odometry
Where the robot thinks it went, from counting wheel turns. The gate compares it with ground truth to catch wrong wheel sizes.
Ground truth
Where the robot actually went in the simulator, read straight from the physics engine.
Forge loop
When a check fails, Muse Spark proposes a change, the checks re-run, up to three rounds. The model proposes; the checks decide.
Robodex
RobotForge’s catalogue of real robots, each with its spec, sources and test results.
Spotter
Identifies a robot from a photo and looks it up in the Robodex.
Take-apart
An exploded 3D view of a robot’s parts, with a printable kit for the parts you can 3D-print.
ESTIMATED
A number guessed from a photo. Treat it as a starting point until you measure the real part.
DOCUMENT-READ
A number copied from a printed spec sheet or label, not guessed.
MEASURED
A number from a real measurement, such as a 3D scan or a simulator reading.
LIVE
Muse Spark answered just now from your input.
RECORDED
A saved result from an earlier real run, replayed. Not generated from your input.
MOCK
An offline helper using fixed textbook rules, used when the live model is unavailable. Not AI.