Computer Vision · Robotics · Multimodal AI Engineer — Production & Embodied AI Systems
Morgan Hill, CA · 🟢 U.S. Green Card · Open to ML Engineer / Robotics AI / Embodied AI roles
7+ years shipping production robotic AI at KETI (Korea Electronics Technology Institute). Builds the full perception stack — detection, segmentation, 6D pose estimation, multi-object tracking — and integrates VLMs, LLMs, and VLA policies into grounded, multi-step robot execution. From research paper to Jetson deployment.
| Achievement | Numbers | |
|---|---|---|
| 🏭 | Industrial bin-picking — end-to-end perception pipeline | 70% → 92% success in cluttered scenes · 30+ FPS on Jetson AGX (TensorRT INT8) · 3+ years in production across multiple sites |
| 🦾 | MARS — 7-DOF mobile manipulation system (Ubiquitous Robots 2025) | 4-layer ROS2 architecture · hybrid gripper · RGB-D perception · LLM task planning + Skill Controller with hierarchical retry |
| 🤖 | VLA manipulation — Franka Panda on Isaac Lab + MuJoCo | 98.75% on LIBERO (exceeds OpenVLA-OFT paper baseline of 97.1%) · 100% Spatial / Object / Goal · 95% Long |
| 👁️ | Fine-aware ViT — precision grasp pose detection (IEICE 2025) | 20% precision improvement over CNN · trained on 200K+ annotations · deployed on Jetson at 30+ FPS |
| 🔁 | Multi-object tracking — 4 trackers from scratch | 11× fewer ID switches vs DeepSORT on MOT17-val · SORT · DeepSORT · ByteTrack · custom |
| ⚙️ | C++ ML kernels — from scratch, no Eigen | 14× GEMM speedup · naive → cache-blocked → tiled → AVX2+FMA SIMD |
| Repo | What it is | Stack | Key result |
|---|---|---|---|
| robot_sim_vla | VLA policy benchmarking on Franka Panda (Isaac Lab + MuJoCo) + classical 3D CV from scratch: hand-eye calibration, PnP, ICP | Python · Isaac Sim 4.5 · MuJoCo · OpenVLA-OFT · LeRobot | 98.75% LIBERO |
| vision_tracking | SORT · DeepSORT · ByteTrack · custom tracker from scratch — Kalman, Hungarian, ReID — + FastAPI + Next.js UI with side-by-side comparison | Python · TypeScript · YOLOv11 · OSNet · Docker · Vercel | 11× fewer ID switches |
| cpp-ml-system | ML/CV/Robotics in C++17 — GEMM variants, im2col CNN, NMS, Kalman/EKF, PID, point cloud; no external ML libs | C++17 · AVX2/FMA · CMake · GoogleTest | 14× GEMM speedup |
| carerobotagent | LangGraph multi-agent care robot — Whisper STT, intent routing, task planning, ROS2 execution, auto-replan, dual memory | LangGraph · LangChain · AI2-THOR · SQLite · ChromaDB · Streamlit | Full MARS stack |
| pyplanner | Pluggable LLM planning library — 7 methods (CoT, ReAct, Self-Refine, Hierarchical, LLM Router…) — AI2-THOR benchmark · live demo ↗ | Python · OpenAI · Anthropic · Ollama | 7 planning methods |
| groundingdino_tool | Semi-auto labeling: GroundingDINO → FastSAM → human verify → YOLOv8 fine-tune loop; targets long-tail | Python · Streamlit · GroundingDINO · FastSAM | Active learning |
| Single-Image-Dehazing | PhD research — Color Ellipsoid Prior for single-image dehazing | Python | IEEE TIP 2018 · 288 citations |
Perception & Computer Vision
PyTorch · OpenCV · RGB-D · 6D Pose Estimation · Multi-Object Tracking · Vision Transformers
TensorRT · ONNX · CUDA · INT8 Quantization · 30+ FPS edge inference
Robotics & Manipulation
ROS2 / ROS · MoveIt2 · Nav2 · Mobile Manipulation · Bin-Picking · Grasp Planning
Skill Controllers (hierarchical retry) · Hand-Eye Calibration · Point Clouds
Multimodal AI & LLM Agents
LangGraph · LangChain · VLA: OpenVLA-OFT · LeRobot Diffusion Policy
VLMs: CLIP · BLIP · LLaVA · GroundingDINO · Gemini Vision
RAG Pipelines · Tool-Use Agents · Episodic + Semantic Memory (SQLite + ChromaDB)
Simulation & Training
Isaac Lab / Isaac Sim 4.5 · MuJoCo · AI2-THOR · LIBERO
Dataset curation · Active learning · Distributed training · Evaluation harnesses
Languages & Infrastructure
Python (expert) · C++17 · TypeScript · C
Docker · FastAPI · Next.js · Streamlit · Vercel · RunPod serverless GPU
| Year | Title | Venue | Citations |
|---|---|---|---|
| 2025 | A Fine-Aware Vision Transformer for Precision Grasp Pose Detection | IEICE Transactions on Information and Systems | — |
| 2025 | Development of a Mobile Assistive Robot for Daily Living Support | Ubiquitous Robots | — |
| 2018 | Single Image Dehazing Using Color Ellipsoid Prior | IEEE Transactions on Image Processing (Q1) | 288 |
| 2014 | Segmenting Dark Channel Prior in Single Image Dehazing | IET Electronics Letters | 5 |
Senior Computer Vision, Robotics & AI Engineer
Korea Electronics Technology Institute (KETI) · Seongnam, South Korea · Mar 2019 – Present
- Led production bin-picking system: 70% → 92% success, 30+ FPS on Jetson AGX (TensorRT INT8), 3+ years deployed across multiple sites, hundreds of items/hour
- Architected MARS: 7-DOF arm + hybrid gripper on 4-layer ROS2 — WebRTC frontend → LLM Task Manager → Skill Controller (hierarchical retry) → Hardware
- Designed Fine-aware ViT: 20% precision improvement over CNN, 200K+ annotations, custom augmentation
- Integrated VLMs (CLIP, BLIP, LLaVA, GroundingDINO, Gemini Vision) for zero-shot recognition; RAG pipelines over internal knowledge bases
- Built LangGraph multi-agent orchestration with structured outputs grounded in robot skill affordances; episodic + semantic memory (SQLite + ChromaDB)
- Optimized inference for edge with TensorRT, ONNX, INT8 — sub-200ms perception-to-action latency
- Mentored 3 junior engineers; led national R&D projects with multi-million USD funding
| Degree | Institution | Years |
|---|---|---|
| Ph.D., Computer Vision | Kyung Hee University, South Korea | 2014 – 2019 |
| M.Eng., Computer Vision | Kyung Hee University, South Korea | 2011 – 2014 |
| B.Eng., Electrical & Electronics | HCMC University of Technology, Vietnam | 2005 – 2010 |
Ph.D. thesis: basis of IEEE TIP 2018 publication — 288 citations
📬 bmtrungvp@gmail.com · 📞 +1 (669) 326-2460 · 📍 Morgan Hill, CA · U.S. Green Card