Physical AI · Robotics software | Imitation-learning data · Simulation · Sim-to-Real | Python · C++ · ROS 2 · MuJoCo
After earning an M.S. in Advanced Device Technology, I spent two years and four months in optical-device packaging R&D. I then worked in web-service planning for three years and nine months and led a full-stack web development team for one year and four months.
For SWEEPICK, a four-person project in Dongguk University’s DAPIER Physical AI program, I was responsible for generating simulation training data and integrating the physical robot. I connected the mobile base and both arms through ROS 2 and developed and evaluated the ACT model with a teammate. The robot completed the task at 132 of 143 positions in simulation overall and at 27 of 31 test positions held out from training. On November 1, 2026, the physical robot completed its first full run without human intervention: picking up the object, carrying and placing it, and returning to its starting pose.
국문 소개
차세대소자공학 석사를 마치고 2년 4개월 동안 광소자 패키징을 연구했습니다. 이후 웹서비스를 3년 9개월 기획했고, 풀스택 웹 개발팀장으로 1년 4개월 일했습니다.
동국대 DAPIER Physical AI 과정의 4인 팀 프로젝트 SWEEPICK에서는 시뮬레이션 학습 데이터 제작과 실물 로봇 통합을 맡았습니다. 이동 베이스와 양팔을 ROS 2로 연동했고, ACT 모델은 팀원과 함께 개발하고 평가했습니다. 시뮬레이션에서는 143곳 중 132곳에서 작업을 완료했습니다. 학습에 쓰지 않은 Test 위치에서는 31곳 중 27곳에서 완료했습니다. 2026년 11월 1일에는 실물 로봇이 사람 개입 없이 집기부터 운반·놓기와 복귀까지 처음 마쳤습니다.
🎓 Physical AI program (DAPIER Cohort 1), Dongguk University · 656 hours · Jul 6–Nov 4, 2026 · completed
🧹 SWEEPICK (Aug–Nov 2026): Our four-person team built a low-cost mobile robot with two arms to clear objects from the floor. I also worked on domain randomization. We froze the epoch-30 checkpoint before evaluating it once per location across 80 grid points and 63 cell centers. On hardware, the grasp results were 7/10, 8/10 and 7/10 across three objects. Evaluation setup.
🚗 deepThinkCar: Added stop conditions and evaluated a candidate steering model. Steering MAE fell from 14.93° to 4.93° on the same split. I held off on using the model in the car because the evaluation was narrow.
⏱️ SyncGuard: Checked timestamp alignment before training. On synthetic data with jitter and dropped samples, valid alignment rose from 33.6% to 98.2% and p95 skew fell from 46.7 to 3.3 ms.
🧭 ROS 2 practice: Worked with URDF, TF, SLAM and Nav2 on Jazzy and diagnosed issues while migrating from Humble to Jazzy.
🏗️ VERA: interior/construction SCM platform built with coding agents in two weeks; LLM APIs with rule-based fallbacks and a Vitest/Playwright QA pipeline.
💼 Earlier work: full-stack web team lead (React/Next.js, Supabase, Playwright automation), freelance full-stack (NestJS, payments and webhooks, GitHub Actions, Docker/EC2), optical transceiver R&D.
| Topic | Details |
|---|---|
| 📍 | Korea · completed the Physical AI program on Nov 4, 2026 · looking for robotics software roles in imitation learning, simulation, Sim-to-Real and robot integration |
| 🔧 | Hands-on: Python, C++, ROS 2 (topics, TF, URDF/RViz), Nav2 basics, MuJoCo, Isaac Sim, RoboTwin, LeRobot / ACT, PyTorch, OpenCV, Linux |
| 🧪 | Recent work: dual-arm ACT data and evaluation, Sim-to-Real hand-off on SO-101 arms, ROS 2 base↔arm interface |
| 🤖 | I give Claude Code and Codex task specs and acceptance criteria, then verify the code with tests. I also build LLM features with rule-based fallbacks and agents for personal automation. |
| 📚 | Start here: portfolio · SWEEPICK · mini projects & practice |
Windows bug fixes in widely used Python projects. I reproduce each bug on Korean Windows (cp949) and add a regression test that fails before the fix.
| Project | Change | Status |
|---|---|---|
| jupytext | Use UTF-8 for the temporary files exchanged with marimo | |
| Panel | Read and write user files as UTF-8 | |
| commitizen | Explicit encodings in cz commit --edit and cz init |
|
| huggingface_hub | Read --filter-from files as UTF-8 |
|
| Locust | Read locustfiles as UTF-8 before sending them to workers | |
| python-semantic-release | Read the previous changelog as UTF-8 | |
| Salt | Fix network.connect always failing with proto=udp |
|
| xonsh | Read the RC file, xontrib sources and the stash log as UTF-8 | |
| Apprise | Read Discord, Slack, Telegram and Workflows templates as UTF-8 | |
| pipenv | Keep the downloaded requirements file for pipenv install -r <url> |
|
| OpenJarvis | Format digest dates without glibc-only strftime flags |
|
| mycli | Read and write --batch and --checkpoint files as UTF-8 |
|
| trimesh | Close the meshio temporary file before reading it on Windows | |
| Harlequin | Read DuckDB and SQLite init scripts as UTF-8 | |
| SABnzbd | Read the password file as UTF-8, with a code page fallback | |
| dirsearch | Read wordlists and URL/proxy list files as UTF-8 | |
| ComfyUI-Manager | Read a node pack's requirements.txt with the encoding fallback when installing |
|
| manga-image-translator | Read and write translation text files as UTF-8 | |
| semantic-router | Read and write router YAML/JSON files as UTF-8 | |
| thonny | Read pyvenv.cfg as UTF-8 so venvs under non-ASCII paths work on Windows |
|
| ursina | Read .obj files as UTF-8 so non-ASCII material names keep their colors |
|
| Flexget | Read template files as UTF-8 in the flexget templates command |
|
| ursina | Read ursina_settings.py as UTF-8 so settings with non-ASCII text load |
|
| syncplay | Read and write playlist files (.m3u8, .txt) as UTF-8 |


