Vectorized Traffic World Models (Flow-ERD)
Learning generative models of interactive environments for realistic, diverse, and controllable closed-loop simulation.
Flow-ERD
Seulbin Hwang
Building traffic world models for AD 🚗 & Robot 🤖
Researcher at NAVER LABS 🇰🇷
Researcher at NAVER LABS building world models for autonomous driving and mobile robots, with a particular focus on the realism of dynamic agents. (
WOSAC 1st!) Current work also focuses on real-time world models that remain effective over long horizons, enabling closed-loop training and evaluation of ego planners.
Takes a problem-driven approach to research, proactively identifying and solving challenges that deliver tangible value to users of autonomous vehicles and robots. Readily learns and adopts new methods and technologies as the problem demands.
Research mission
Realistic traffic worlds for planner training and evaluation.
Building closed-loop traffic world models that capture realistic dynamic-agent behavior and run in real time over long horizons, enabling planner training and evaluation for autonomous vehicles and mobile robots.
Multi-View Camera Traffic World Models
Extending Flow-ERD with multi-view image rendering, using approaches such as NVIDIA OmniDreams to bridge vectorized traffic simulation and camera-based world models.
Traffic World Models Across Robot Platforms
Developing traffic world models that support multiple platforms, from autonomous vehicles to outdoor mobile robots, for closed-loop planner training and evaluation.
News
Recent updates
- 🏆 Our recent paper Flow-ERD ranked 1st on the official Waymo WOSAC 2025 Sim Agents Challenge leaderboard.
- Flow-ERD preprint and project page released.
- 📝🦾 OVBEVSeg is accepted to ECCV 2026.
Experience
