Driven by vision. Defined by persistence.
B.S. in Automotive EngineeringKookmin University, Senior
A custom-built RL-oriented self-driving simulator in Unity for synthetic data generation and training — with Blender-modeled environments, sensor pipelines, and diffusion-based domain adaptation to bridge the Sim-to-Real gap.
Realistic vehicle simulation using Unity's WheelCollider system with custom C# controllers. Tire friction curves, suspension parameters, and drivetrain characteristics are tuned to match real-world ERP-42 and JetRacer dynamics for meaningful Sim-to-Real transfer.
Blender-modeled tracks, roads, and urban environments rendered in Unity. Camera and LiDAR sensor pipelines generate synthetic training data — camera images via Unity's render pipeline, LiDAR point clouds via GPU-accelerated raycasting. Data format matches real sensor output for seamless pipeline integration.
Developing diffusion models to reduce the visual domain gap between simulated and real-world driving data. The goal is photorealistic synthetic data that enables RL policies and perception models trained in simulation to transfer directly to real vehicles with minimal fine-tuning.
The simulator exposes a Gym-compatible interface for reinforcement learning — observation space (camera, LiDAR, vehicle state), action space (steering, throttle), and reward shaping for lane following, collision avoidance, and smooth driving behavior.