Skip to content
View kimsunguk0's full-sized avatar

Block or report kimsunguk0

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
kimsunguk0/README.md

Sung-Uk Kim

Autonomous Driving Researcher — End-to-End Driving & Vision-Language-Action Models

Junior Researcher at Beyless (Autonomous Intelligence Team) · B.S. Computer Engineering, Gachon University 📧 ksu903@naver.com

I work on end-to-end autonomous driving with vision-language-action (VLA) models: making large driving models small and fast enough to run on a vehicle, and understanding why a compressed model behaves differently from its teacher. Before that I worked across perception on real vehicles — camera and LiDAR detection, camera–LiDAR calibration, multi-sensor time sync and SLAM.


Highlights

  • Ishinca-4B — distilled NVIDIA Alpamayo-1.5-10B into an open 4B VLA. Teacher-only supervision with image-token relation distillation and a 28-layer flow-matching action expert; drives on par with the 10B teacher in AlpaSim closed loop (150 scenes), and DART variants prune 25–50% of vision tokens without losing closed-loop score. Weights, inference code and evaluation protocol released. → Ishinca · 🤗 suukim/Ishinca-4B
  • 1st place, E2E Driving track — 2026 Autonomous Driving AI Challenge (Minister of Science and ICT Award). Led a team of four and built the winning camera-only end-to-end planner myself: weighted L2 0.1171 m, 26.7M parameters, ~31 ms on an RTX 4090. → E_Drive
  • Ported Alpamayo to TensorRT-Edge-LLM on Jetson Thor — fixed a multi-image KV-cache race condition and brought up a 10 Hz sensor-to-planner live loop. → tensorrt-edge-llm-for-alpamayo
  • Minister of Science and ICT Award — 2024 Autonomous Driving AI Challenge (vehicle 3D object detection).

Research interests

  • Efficient VLA perception: reducing and selecting vision tokens based on what the action model actually reads.
  • Spatial understanding in VLA: bringing ego-motion and geometry (SLAM-style cues) into vision encoders for driving and robotics.
  • Distillation of driving foundation models: representation-level distillation and closed-loop evaluation.

Projects

Project What it is Stack
Ishinca stars Open 4B student of NVIDIA Alpamayo-1.5-10B (Cosmos-Reason2-2B VLM + flow-matching action expert) with DART vision-token pruning; closed-loop parity with the teacher. Weights on Hugging Face. PyTorch · Transformers · AlpaSim
E_Drive Camera-only E2E planner, winner of the 2026 Autonomous Driving AI Challenge E2E track. Goal and ego state used only as BEV attention-query conditions; 15 candidate trajectories with goal-based selection. PyTorch · BEV attention · Docker
tensorrt-edge-llm-for-alpamayo stars Alpamayo VLA stack ported to NVIDIA TensorRT-Edge-LLM for Jetson Thor. C++ runtime fixes and ONNX export path. C++ · TensorRT · Jetson Thor
alpamayo-sensor-runtime Jetson Thor runtime for four GMSL2 cameras, IMU and GNSS with online clock-drift modeling for microsecond-level alignment; dataset recorder + live planner streaming. C++17 · V4L2 · Docker
GLIM_SLAM LiDAR-IMU-GNSS SLAM over a ~27 km bridge route, with GNSS factors in the GLIM odometry graph and VGICP localization. ROS2 · GLIM · GTSAM
MultiObjectLidarTracking LiDAR 3D detection (SFA3D) + multi-object tracking (SimpleTrack) ported to C++/CUDA/TensorRT for AGX Orin and Hailo. C++ · CUDA · TensorRT

Skills

Models — vision-language-action models, flow-matching action experts, BEV end-to-end planners, 3D detection and tracking Research methods — knowledge and relation distillation, GRPO, vision-token compression, CKA and probing, causal interventions Deployment — TensorRT, TensorRT-Edge-LLM, TRT-LLM, FP8 / NVFP4 quantization, Jetson Thor, AGX Orin Systems — PyTorch, C++, CUDA, ROS2, Docker

Experience

  • Beyless — Junior Researcher, Autonomous Intelligence Team (2026.07 – present); Intern, Autonomous Driving Development Team (2025.09 – 2026.06)
  • MORAI — Intern, Engineering Service Team (2025.01 – 2025.02)
  • GADIS, Gachon University autonomous driving team — Perception lead (2024.01 – 2025.11), K-City field tests

Awards

  • 2026 · Minister's Award (MSIT) — Autonomous Driving AI Challenge, E2E Driving 1st place
  • 2025 · President's Award, KEPCO KDN — International Student EV Autonomous Driving Competition
  • 2024, 2025 · Honorable Mention, MOLIT — Student Creative Mobility Competition
  • 2024 · Minister's Award (MSIT) — Autonomous Driving AI Challenge, 3D object detection
  • 2024 · Grand Prize, Gachon University P-Project (AI track)

Pinned Loading

  1. tensorrt-edge-llm-for-alpamayo tensorrt-edge-llm-for-alpamayo Public

    Apply alpamayo for Jetson Thor with TRT

    C++ 14 2

  2. alpamayo-sensor-runtime alpamayo-sensor-runtime Public

    Jetson Thor sensor runtime for 4K camera, IMU, and GNSS recording and live planner sample streaming.

    C++

  3. MultiObjectLidarTracking MultiObjectLidarTracking Public

    3D MOT for Embeddied PC: AGX Orin, Hailo etc...

    C++ 1

  4. Ishinca Ishinca Public

    Ishinca: Knowledge Distillation NVIDIA Alpamayo 1.5 (10B) into a 4B vision-language-action model for autonomous driving

    Python 1