CV
Scholar
Email
LinkedIn
[C1] Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport
Jinha Choi
Physics-informed Dexterous Manipulation · Humanoid Learning · Reinforcement Learning
Jinha Choi
I'm Jinha Choi, an undergraduate student at Yonsei University, South Korea, majoring in Computer Science and Nano Science and Engineering and expected to graduate in February 2027.
I am currently advised by Prof. Yonghyeon Lee at RoPhi Lab at Yonsei University, working on high-DoF dexterous manipulation incorporating physics- and control-informed priors toward generalizable robot intelligence for unstructured environments and real-world deployment. Previously, I was a research intern at HuGeLab at KAIST, advised by Prof. Beomjoon Kim, where I worked on real-world deployment of humanoid locomotion policies. I also worked at DILLab at Yonsei University, advised by Prof. Jongmin Lee, on offline reinforcement learning that distills a value-aware multi-step flow policy into an efficient one-step policy using state-wise entropic optimal transport. Earlier, at Yonsei RLLAB, advised by Prof. Youngwoon Lee, I built humanoid locomotion baselines using retargeted motion priors and studied their effects on motion quality and stability.
News
- Sep 2026 Started research internship at RoPhi Lab, Yonsei University, working on high-DoF dexterous manipulation with Prof. Yonghyeon Lee.
- Jun 2026 Started summer internship at KAIST HuGeLab, working on humanoid robot learning with Prof. Beomjoon Kim.
- Jun 2026 Submitted "Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport" to NeurIPS 2026.
- Jan 2026 Started research internship at DILLAB, Yonsei University, working on multimodal flow policies and offline RL with Prof. Jongmin Lee.
- Jul 2025 Started research internship at RLLAB, Yonsei University, working on humanoid loco-manipulation with Prof. Youngwoon Lee.
- May 2025 "When AI Co-Scientists Fail: SPOT" — a benchmark for automated verification of scientific research — available on arXiv (contributing co-author).
Publications
[P1] When AI Co-Scientists Fail: SPOT — a Benchmark for Automated Verification of Scientific Research
Experience
Demo coming soon
Grasping via Global Planning and Local Correction
Advised by Prof. Yonghyeon Lee
- Developing a hierarchical grasping system coupling real-time global planning and control based on robot kinematics and dynamics with learning-based local correction.
- Integrating side-view and wrist-mounted visual feedback across both control levels, targeting robust grasping across diverse objects and positions.
Demo coming soon
Franka FR3 Control and Perception
Advised by Prof. Yonghyeon Lee
- Implemented torque-level controllers on a Franka FR3, including gravity compensation, collision-aware control, QP-based inverse kinematics, and joint/task-space impedance control.
- Integrated a calibrated two-camera setup with point-cloud perception and visual servoing for perception-guided grasping on the physical robot.
Humanoid Deployment — Locomotion
Advised by Prof. Beomjoon Kim
- Trained, validated, and deployed locomotion policies on the Unitree G1 and a custom 108 kg humanoid via sim-to-sim and sim-to-real evaluation.
Franka FR3 Teleoperation Data Collection
Advised by Prof. Beomjoon Kim
- Integrated a Franka Panda with a 3D mouse, Touch haptic device, and Meta Quest for teleoperated demonstration collection.
Humanoid Fixed-Base Manipulation
Advised by Prof. Youngwoon Lee
- Investigated PPO exploration limits for humanoid fixed-base manipulation in massively parallel simulation.
- Compared PPO with SAPG and EPO and analyzed methods for improving exploration.
A1-Challenge Autonomous Driving Competition
Ministry of Trade & KIAPI
- Placed 4th; qualified in Nov. 2025.
- Built behavioral-cloning and DAgger racing policies in MORAI using targeted corner-case demonstrations.
- Explored reinforcement learning to improve robustness and driving performance.
Physical AI Robotics Hackathon — Arm Manipulation
Kernel Academy & Doosan Robotics
- Built a speech-guided pick-and-place pipeline with a Doosan E0509, RGB-D sensing, Whisper, SAM2, GPT-4o, and ROS2.
- Implemented calibrated 3D localization, task verification, and error recovery for autonomous execution.
Image Inpainting with LLMs and Diffusion Models
OUTTA AI Bootcamp
- Developed a language-conditioned image editing pipeline combining LLM-based attribute extraction, semantic segmentation, and diffusion-based inpainting.
Korea Defense Intelligence Command — Translator
Military Service
- Translated classified documents and interpreted classified meetings for KDIC and partner agencies, including the National Intelligence Service and the U.S. National Geospatial-Intelligence Agency.
VEX Robotics
Team Lead, 1010B, WVSS Robotics
- Led a 12-member team across three seasons, overseeing strategy, mechanical and electrical design, programming, autonomous control, and competitive driving.
- Earned first-place finishes at regional and British Columbia provincial competitions and advanced to the VEX Robotics World Championship.
Awards
Hyundai NGV Robotics Track H-Mobility Class Certificate
Mar – Jul 2026 · Hyundai NGV
Completed Hyundai’s Robotics Track in robotics and robot learning and received the H-Mobility Class Certificate.
2nd Place — Physical AI Robotics Hackathon
Mar 2026 · Kernel Academy & Doosan Robotics
Presented a manipulation pipeline and demonstrated perception-guided grasping at the Kernel Academy hackathon.
Top Award — Korea Defense Intelligence Command AI Security Competition
Awarded by Major General Sangho Moon (OF-7), Commander, KDIC
Proposed an on-device AI framework for military systems with partial internet exposure, designed to support operational convenience while strengthening protection against security vulnerabilities and malicious misuse.
Honor Roll — Yonsei University
Jul 2022
Recognized for academic performance in the top 3%.
1st Place — Yonsei–Nexon RC Creative Platform
Mar – Dec 2021
Led a five-member team to develop an automated revolving-door assistance system for people with mobility impairments; ranked 1st of 90 teams and received a $10,000 award.
Division Finalist — VEX Robotics World Championship
Nov 2019 · Kentucky, USA
Reached the Science Division final in a 100-team division within a field of 600 qualified teams.
1st Place — VEX Robotics Canada BC Provincial Championship
Jun 2019
Won the British Columbia Provincial Championship and qualified for the VEX Robotics World Championship.
Contact
Feel free to reach out for research discussions, collaboration, or general inquiries.
Email
jinhac174@gmail.com
Google Scholar
Jinha Choi
LinkedIn
Jinha Choi
Location
50 Yonsei-Ro, Seodaemun-Gu
Seoul, Republic of Korea
Seoul, Republic of Korea