Projects and Research Interests

• Goal

  • VLA Policy Design : Test-Time Action Verification, Best-of-K Selection

  • Data-Efficient Self-Improving Learning for Physical AI

  • VLA Success Rate Maximization under Real-Time Inference Constraints

• Technical Approaches

  • Vision-Language-Action (VLA) Models

  • Test-Time Verification (Reward Modeling)

  • Reinforcement Learning (DPO)

• Goal

  • Autonomous Valet Parking (AVP) System Design based on the CARLA Simulator

  • Human-in-the-Loop-based Parking Slot Selection and Autonomous Parking Pipeline Construction

Technical Approaches

• Perception

  • CVT (Cross-View Transformer)-based BEV (Bird's Eye View) Perception 

  • Human-in-the-Loop (HITL)-based Parking Slot Selection 

• Planning

  • Hybrid A* Algorithm for Initial Path Search

  • OCP (Optimal Control Problem)-based Optimized Path Control

  • Training a Path Generation Model using Trajectories Generated by Hybrid A* / OCP as Training Data

• Control

  • SAC (Soft Actor-Critic)-based Reinforcement Learning Control

  • Dynamic Obstacle (Vehicle, Pedestrian) Response