Evidence
Measured, not believed
Every verdict is judged against simulator ground truth, where the robot actually went, never against its own odometry alone. These are the gate's own reports from real runs in Gazebo Sim, headless, on Oct 26, 2026.
Locked thresholds
Printed on every report and never quietly loosened.
How a run is graded
The twin is commanded to drive 1 m and turn 90°. Drive and turn are graded against the simulator's true pose. Odometry is graded against that same truth, so a robot whose sensors are wrong about itself can't pass. When something is off, the gate names the spec field to look at first.
Gate reports
Two clean passes and two failures the gate caught. Each report shows every check, the raw measurements, the ground-truth path, and any suspects the gate named.
Indoor Delivery Rover
Described in words, never drawn
This robot existed only as one paragraph of plain English at 4:52 PM. Muse Spark turned it into a validated spec, the forge built its twin, and the gate measured it: PASS, with odometry matching ground truth to 0.002% over the drive. The full chain, one flow.
Drive vs ground truthPass
0.398%
0pass ≤5%fail >10%15
Turn vs ground truthPass
3.653°
0pass ≤5°fail >10°15
Odometry distance vs truthPass
0.002%
0pass ≤5%fail >10%15
Odometry heading vs truthPass
1.088°
0pass ≤5°fail >10°15
| Drive: truth / odometry | 1.004 m / 1.004 m |
|---|
| Turn: truth / odometry | 93.7° / 92.6° |
|---|
White dot: start. Amber dot: where the robot actually ended up (ground truth).
No suspects. Odometry and ground truth agree inside tolerance on every check.
TurtleBot3 Burger
As specified
A differential-drive robot built to the public ROBOTIS spec. Nothing injected, nothing hidden.
Drive vs ground truthPass
0.968%
0pass ≤5%fail >10%15
Turn vs ground truthPass
2.028°
0pass ≤5°fail >10°15
Odometry distance vs truthPass
0.042%
0pass ≤5%fail >10%15
Odometry heading vs truthPass
4.548°
0pass ≤5°fail >10°15
| Drive: truth / odometry | 1.010 m / 1.009 m |
|---|
| Turn: truth / odometry | 88.0° / 92.5° |
|---|
White dot: start. Amber dot: where the robot actually ended up (ground truth).
No suspects. Odometry and ground truth agree inside tolerance on every check.
TurtleBot3 Burger
Wheels 10% smaller than the paperwork
Fault injection: the physical wheels were scaled to 0.9x while odometry kept believing the spec. The robot's own sensors could not see it. The gate could.
Drive vs ground truthMarginal
7.617%
0pass ≤5%fail >10%15
Turn vs ground truthMarginal
6.766°
0pass ≤5°fail >10°15
Odometry distance vs truthFail
11.168%
0pass ≤5%fail >10%15
Odometry heading vs truthFail
11.041°
0pass ≤5°fail >10°15
| Drive: truth / odometry | 0.924 m / 1.027 m |
|---|
| Turn: truth / odometry | 83.2° / 94.3° |
|---|
| Injected fault | physical wheels × 0.9 (odometry keeps the spec size) |
|---|
White dot: start. Amber dot: where the robot actually ended up (ground truth).
Suspect: wheel.diameter_m
odometry distance disagrees with ground truth while heading agrees — a scale error: the wheel radius odometry uses does not match the physical wheels
Suspect: wheel_separation (robot width / wheel track)
odometry heading disagrees with ground truth — the track width odometry integrates over does not match the physical robot (or the wheels scrub/slip in turns)
Pathfinder skid-steer
The turn that wasn't
Odometry reported a near-perfect drive. In the turn, its wheels scrubbed: it believed 94.6 degrees, the ground truth was 30.6. Caught only because the gate measures truth, not belief.
Drive vs ground truthPass
1.550%
0pass ≤5%fail >10%15
Turn vs ground truthFail
59.368°
0pass ≤5°fail >10°15
Odometry distance vs truthPass
0.000%
0pass ≤5%fail >10%15
Odometry heading vs truthFail
64.003°
0pass ≤5°fail >10°15
| Drive: truth / odometry | 1.016 m / 1.015 m |
|---|
| Turn: truth / odometry | 30.6° / 94.6° |
|---|
White dot: start. Amber dot: where the robot actually ended up (ground truth).
Suspect: wheel_separation (robot width / wheel track)
odometry heading disagrees with ground truth — the track width odometry integrates over does not match the physical robot (or the wheels scrub/slip in turns)
Robodex verdicts
The current verdict for each robot and its latest recorded run. Failures stay on the record exactly as measured.
Static stability checks
A different kind of evidence, kept separate from the gate reports above: for the legged robots, a computation from the spec — the support polygon its feet make, and how far the robot can tip before its centre of mass passes over the edge. Nothing here was simulated or driven, and a static check never changes a gate verdict: every robot below is still Not gated. Method, inputs and assumptions are printed in each report.
Harness checks
A third kind of evidence, kept separate from gate runs and static checks: electrical computations over each robot's documented wiring graph — voltage drop, ampacity, logic levels, rail budgets, connector matching. The headline is never a blanket pass; it is counts. A check that could not run says which data it is missing, and that gap list is printed in full on each robot's page: it is the shopping list of datasheets still to hunt.