5 agents · 1 shared world · 1 ground truth. The pedestrian behind the parked truck is the canonical AV danger — the engine flags it occluded while the camera would miss it.
Agents served
5
Shared world
1 twin
Occluded pedestrian
caught
Determinism
all identical
All 5 queries
5.5 ms
Top-down road (x → down-road, z ↑ across lanes). Yellow = ego sensor · green dashed = clear line of sight · red dashed = sight line blocked by the parked truck · blue = ego's verified lane-change path.
Perceptioncheck_visibility
"Is the pedestrian visible to the ego sensor?"
↳ NO — pedestrian is in the sensor's FOV but OCCLUDED by parked_truck (do NOT clear the path)
"Is the intended lane-change path collision-free?"
↳ clear — collision-free lane-change path, 23.8 m (91 waypoints) routed around lead_car/left_car
5.332 ms188 ops/sdet ✓
Trackerpersistent shared world
"What is the pedestrian's last-known position while occluded?"
↳ last-known position (24.0, 1.0) — the engine still holds the pedestrian while the sensor cannot (object permanence)
0.002 ms500,838 ops/sdet ✓
Scene-GTrelational_graph (near/between)
"What is the nearest hazard to ego, and what blocks its sight line?"
↳ nearest hazard is left_car at 12.5 m; ground truth: parked_truck lies between ego_sensor and pedestrian
0.163 ms6,130 ops/sdet ✓
What this shows. An autonomous-vehicle scenario where five agents share one Universtato world holding the deterministic ground truth. The safety-critical answer: the pedestrian is inside the ego sensor's field of view yet occluded by the parked truck via check_visibility — a VLM that sees the pedestrian in frame would wrongly assume the ego does too. perspective_transform gives ahead/left/right from ego's own frame, navigation_plan verifies the lane-change path is collision-free, the persistent world holds the pedestrian's last-known position while occluded (object permanence), and relational_graph reports the nearest hazard plus the between relation that corroborates the blockage. Each answer is computed exactly — not eyeballed. Maps to research items #29 (scenario ground-truth) and #30 (occlusion-aware perception). Latency/determinism measured on this machine.