DEMO · SIMULATION / SYNTHETIC DATA

Physical-Validity Gate

A generated scene that looks fine but isn't. Universtato catches the three physics flaws — floating, interpenetrating, tipping — before the data is used for training.
Objects
7
Violations
3
Verdict
INVALID
Determinism
all identical
All 5 checks
4.6 ms
ground (y = 0) crate_A crate_B floater clip_L clip_R stack_base stack_top 1.9 m gap floating overlap tips +x teal = valid red = floating / overlap orange = tips
Side elevation (x horizontal, y up). Teal = valid crates · red = floating crate (drop arrow) and interpenetrating pair · orange = overhanging crate that tips.
Supportcheck_stability (verdict 'falls')
"Any floating / unsupported objects?"
↳ YES — floater is floating: no support beneath it (verdict 'falls')
0.049 ms20,522 ops/sdet ✓
Penetrationcollision (AABB overlap)
"Any interpenetrating objects?"
↳ YES — clip_L and clip_R interpenetrate (share 0.3 m of the same volume)
0.030 ms33,521 ops/sdet ✓
Stabilitycheck_stability (verdict 'tips')
"Any unstable stacks that will tip?"
↳ YES — stack_top tips +x (center of mass 0.30 m past its support edge)
0.045 ms22,396 ops/sdet ✓
Physicspredict (gravity rollout)
"Drop-test the floating object — does it settle or fall?"
↳ FALLS — drop-test on floater: y 2.4 m -> 0.5 m over 3.0 s (falls and settles on the ground)
4.238 ms236 ops/sdet ✓
Verdictvalidity summary
"Is the scene physically valid for training?"
↳ INVALID — 3 violations: 1 floating, 1 penetration, 1 unstable
0.244 ms4,098 ops/sdet ✓
What this shows. A machine-generated 3D scene that looks plausible but hides three physics violations — exactly what a world-model or synthetic-data pipeline must catch before the scene poisons a training set. Universtato gates it deterministically: a floating crate (check_stability verdict falls), an interpenetrating pair (AABB overlap), and an overhanging stack that tips (check_stability verdict tips), confirmed by a gravity predict drop-test. The engine computes each answer exactly — it does not eyeball the render. Maps to research items #27–28 (physical-validity filtering for generated 3D / synthetic data). Latency and determinism measured on this machine.