Student projects
Exploring whether video generation AI can augment autonomous driving datasets — using SORA, LUMA, and Kling to synthesize realistic driving scenarios.
Lane tracing in the CARLA simulator — progressing from a carrot-chasing algorithm to NVIDIA DAVE-2 and finally Google DeepLabV3+ with pure pursuit control.
Building an autonomous vehicle simulation environment in Unity using ML-agent, combining reinforcement learning and imitation learning for parking behavior.
9 projects spanning Unity & Unreal Engine simulation, RC car & drone object tracking, stereo camera 3D localization, object re-identification, and web/app development.
Autonomous parking AI in Unity — the agent learns to find an empty parking space using obstacle avoidance, imitation learning, and reinforcement learning.
Investigated whether AI-generated driving videos (using SORA, LUMA, and Kling) can supplement real and simulation-based autonomous driving datasets. Evaluated generated videos using VBench quality metrics and YOLO object detection scores.
- Sourced open-source datasets (BDD100k, SYNTHIA) and identified video generation AI options
- Defined 10 edge-case driving scenarios with OpenSCENARIO DSL prompts
- Generated ~60 videos using Text-to-Video and Image-to-Video pipelines
- Set up VBench and YOLO evaluation on the HASS Lab GPU server
- Automated testing pipeline and analyzed results across models
Explored autonomous driving in the CARLA simulator through three progressive stages — starting with a simple geometric algorithm, then applying end-to-end deep learning, and finally combining semantic segmentation with classical control for robust lane following.
- Configured CARLA (Unreal Engine on Windows, Python API on Ubuntu WSL)
- Implemented Carrot-Chasing algorithm using Global Route Planner waypoints
- Trained NVIDIA DAVE-2 end-to-end steering model on 8,000+ collected images
- Integrated Google DeepLabV3+ lane segmentation with pure pursuit steering
- Iterated on camera placement, FOV, and frame interpolation to improve model input
Built a Unity-based autonomous driving simulation using ML-agent. Learned Unity fundamentals from scratch, then trained a kart-racing agent with reinforcement learning before tackling full autonomous vehicle simulation with parking behavior.
- Learned Unity C# scripting basics through a 2D game project
- Trained a kart-racing ML-agent using PPO reinforcement learning
- Set up a highway simulation environment with realistic assets
- Implemented autonomous driving and parking using RL + imitation learning
- Debugged version compatibility issues (ML-agent, Unity, Python)
Worked across 9 domains — from game engine simulation to physical hardware (RC cars and drones), stereo camera 3D tracking, and object re-identification. All code and hardware implementations were done independently, with research methodologies referenced from IEEE papers.
Car driving, custom virtual environments, a balance game, RL-based autonomous car, drone simulation, and pedestrian behavior tree.
Six projects including a custom environment, shooting game, Fall Guys clone, maze game, tank game, and RPG.
Reconstructed the DGIST Technopolis road network in Vissim and ran vehicle flow simulations — pedestrians, parking, 2D/3D views.
Built an RC car from scratch. The car detects a chosen person on camera and follows only that person using object detection + ByteTracker.
Extended the RC car project to a Tello Edu drone. Implemented PID control and mode switching (keyboard / tracking / attacking).
Extracted real-world 3D coordinates of a moving person using two cameras. Extended an IEEE paper's method to work even when cameras are moving.
GPS-free camera position tracking, depth sensing, 3D object detection, and skeleton tracking — all in Python.
Matching and tracking the same object across different camera views and frames despite appearance changes (lighting, viewpoint, occlusion).
Built the original HASS Lab undergraduate web page. Also developed a washing machine booking app similar to HypenPay.
- If you have a research topic, project, or study in mind, you can freely pursue it. Any topic is fine — the professor does his best to advise.
- If you don't have a specific topic yet, the professor recommends study topics and actively helps you along the way.
- One of the professor's greatest strengths is how well he explains things and gives sharp, constructive criticism. His advice is genuinely helpful no matter what you're studying.
- He is the kindest and most approachable person I have encountered in my time at DGIST.
- The professor never puts pressure on undergraduates regardless of their pace, and always offers support.
- He genuinely understands that everyone comes from a different background and grows at a different pace — and he acts accordingly.
- The students in the lab seriously tried to find the professor's shortcomings. They really couldn't find anything.
— Written by Seokmin Lee
Developed a parking AI for autonomous vehicles in Unity. The agent learns to navigate to and park in an empty space while avoiding obstacles — trained using a combination of user-demonstrated imitation learning and PPO reinforcement learning.
- Designed a Unity parking environment with realistic vehicle physics
- Recorded user parking demonstrations for imitation learning
- Trained agent with ML-agent (PPO + behavioral cloning)
- Implemented obstacle avoidance alongside parking behavior
- Evaluated results with TensorBoard (cumulative reward, policy loss)