Seoul World Model — Physical AI Pipeline

Real-World Data Capture

LiDAR sensor vehicles scan Seoul's entire road network

1.2M+ high-resolution 360° panoramic frames captured

±cm precision lane geometry, tunnel interiors, overpasses

NAVER Cloud's spatial data infrastructure stores & processes

NVIDIA Cosmos 3 Fusion

Spatial data fed into Cosmos 3 World Foundation Model

Physics engine trained on hundreds of millions of hours of real-world motion video

Material properties: friction, deformation, light refraction — all learned, not coded

Text prompt or sensor input → photorealistic physics simulation output

Edge-Case Simulation

Virtual Seoul: unlimited scenario generation without real-world risk

Night-time heavy rain + sudden pedestrian jaywalking

Tunnel lighting transitions — rapid illumination shifts

Tire slip on wet urban surfaces — physics-accurate simulation

Real-World Robot Control

Sim-trained policies transferred to NAVER's ARC robot network

ROOKIE ground logistics robot: tuned city-level navigation

Reduced real-world testing cycles — fewer physical trials needed

Simulation-to-reality gap closed by physics-accurate training data

Example Edge Cases — Simulated Safely Before Real-World Deployment

🌧️ Night Rain + Jaywalker

Heavy rainfall at 2 AM, dim streetlights, sudden pedestrian crossing mid-block — robot must brake before the human is visible to forward camera.

🚇 Tunnel Light Transition

Vehicle exits tunnel into direct afternoon sunlight — rapid camera overexposure window during which the robot must maintain lane without visual input.

🚧 Wet Surface Slip Control

Intersection with oil-contaminated wet asphalt — physics simulation calculates exact slip coefficient for ROOKIE's wheel contact patches at approach speed.

🏗️ Construction Zone

Unmarked lane closure, temporary barriers replacing road markings — robot navigates by 3D spatial model of the environment rather than painted lane lines.