The Cho Chun Shik Graduate School of Mobility will host a seminar featuring recent work from NAVER LABS on “SFiD-Init,” a diffusion-based approach for generating realistic multi-agent traffic scene initializations from vectorized HD maps. The talk will cover how the method captures lane topology, agent interactions, and geometric relations to produce map-consistent and physically valid scenes, while introducing safety-shaped regularization to reduce collisions and off-road placements and fidelity-preserving objectives to maintain realistic scene structure. Experimental results on nuPlan, WOMD, and an unseen map will also be presented. Please note that this seminar will be conducted in Korean.

• Time: March 31 (Mon), 4:00 PM
• Location: Room L409, KAIST Munji Campus
• Speakers: Dae Jung Kim, Jinhan Lee (NAVER LABS)