Portrait of Seulbin Hwang

Seulbin Hwang

Generative World Models
for Autonomous Driving & Robotics

My current research focuses on generative multi-agent traffic simulation, flow matching, and closed-loop learning, toward world models for autonomous driving and robotics.

Researcher at NAVER LABS

Research Google Scholar LinkedIn GitHub Hugging Face X Email Publications

News

Recent updates

  1. 🏆 Our recent paper Flow-ERD ranked 1st on the official WOSAC 2025 Sim Agents Challenge leaderboard.
  2. Flow-ERD preprint and project page released.
  3. OVBEVSeg preprint released.

Experience

Education

Research directions

Building systems that understand and simulate interactive worlds.

01

Generative World Modeling & Simulation

Learning generative models of interactive environments for realistic, diverse, and controllable closed-loop simulation.

Flow-ERD
02

Open-World 3D Perception

Recognizing previously unseen objects while preserving geometric consistency and efficient inference for autonomous systems.

OVBEVSeg
03

Safe Decision-Making & Planning

Developing policies that account for interaction, uncertainty, and risk in autonomous navigation and driving.

Selected work

Selected publications

Research across simulation, perception, and planning.

Complete list on Scholar
Flow-ERD traffic simulation showing two plausible worlds

2026 preprint · Co-first author

Flow-ERD: Agent-type Aware Flow Matching with Entropy-Regularized Distillation for Diverse Traffic Simulation

Continuous, type-compatible traffic generation with closed-loop fine-tuning that preserves multimodality.

Result: #1 on the WOSAC 2025 test benchmark, RMM 0.7878.

Paper Project
Comparison of conventional and 3D-aware Gaussian splatting for open-vocabulary BEV segmentation

2026 preprint · Co-author

Open-Vocabulary BEV Segmentation with 3D-Aware Geometric Constraints

Open-world bird's-eye-view perception that recognizes unseen categories while preserving 3D geometry and real-time efficiency.

Result: +15.3 mIoU on unseen categories, 2.5x faster inference, and 0.22x memory.

Paper Project
Soft Actor-Critic lane-change planner controlling acceleration and steering in a traffic simulator

IEEE T-ITS · 2022 · First author

Autonomous Vehicle Cut-In Algorithm for Lane-Merging Scenarios via Policy-Based Reinforcement Learning Nested Within Finite-State Machine

A structured driving policy that combines interpretable state transitions with learned decision-making for interactive lane merging.

Venue: IEEE Transactions on Intelligent Transportation Systems.

Paper
Mobile robot navigating a real indoor hallway during risk-sensitive navigation experiments

ICRA · 2021 · Co-author

Risk-Conditioned Distributional Soft Actor-Critic for Risk-Sensitive Navigation

A distributional reinforcement-learning policy that adapts its risk preference at runtime without retraining.

Evaluation: Simulated and real-world mobile-robot navigation.

Paper DOI