See the engine think
Live spatial reasoning, computed by the engine — exact, fast, identical every run.
The Spatial Flip Test →
Same geometry to a bare LLM and to Universtato. One guesses; one computes — left/right, will-it-tip, occlusion, reachability, persistent memory — in microseconds.
Warehouse Digital-Twin Agent →
Five agents share one persistent warehouse twin. Each calls a spatial tool — route around a blocked aisle, true line-of-sight, stack stability, live pallet location — instead of guessing.
Load-Plan Verifier →
An LLM proposes a container load; Universtato verifies it exactly — fit, collisions, stack stability, cube utilization, unload order. Two planted flaws, both caught in milliseconds.
Bin-Picking Sanity Check →
A cluttered parts bin. The robot asks which part is clear to pick, whether a grasp topples the stack, what's occluded, and the collision-free approach — exact answers, not guesses.
Fits & Reaches →
A robot work cell: collision-free reach to the part, fixture interference, slot fit with tolerance, and cell-layout collisions — the part is 25 mm oversize, and the engine catches it.
BIM Coordinator's Kit →
A building floor: duct-vs-pipe clash detection, equipment fit through doorways, ADA turning clearance, and camera line-of-sight coverage with blind spots — computed from the model.
Will-It-Fit AR →
A scanned living room. Does the new couch fit the gap, clear existing furniture, and pass through the doorway? It fits the room but is too wide for the door — caught before purchase.
AV Scenario Ground-Truth →
A road scene where a parked truck occludes a pedestrian from the ego sensor — the canonical AV danger. Plus exact relations, a collision-free lane change, and object permanence.
NPC That Can't See Through Walls →
A game level: the guard's line-of-sight is blocked by cover (no seeing through walls), it knows left from right, paths around the wall, and places props that rest stably.
Physical-Validity Gate →
A machine-generated scene with hidden physics flaws — a floating object, an interpenetration, an unstable stack. Universtato flags all three before the data is used for training.