Seoul World Model Generates Realistic Robotics Training Data via Spatial RAG

bilawalsidhu · x · 2026-08-22

Bilawal Sidhu explains how the Seoul World Model generates realistic training data for robotics. Similar to Google's Street View grounding, it uses a "spatial RAG" mechanism that retrieves the nearest real-world reference image to condition generation. This allows creating high-fidelity training data for specific locations (like downtown Seoul) under various lighting conditions, aligning with NVIDIA's Omniverse approach.

Original post →

More from Embodied

Embodied channel →