Samsung Research
As a Robotics Research Engineer at Samsung Research, I worked on learning-based robot manipulation: vision-language-action (VLA) models, imitation learning, and reinforcement learning for manipulators and humanoids. I also developed the sound source localization (SSL) algorithm for the home robot Samsung Ballie.
VLA models for manipulation and humanoid control
- Developed a reinforcement learning post-training method for VLA models: trained residual RL policies on top of the pre-trained VLA and imitation-learning policies, targeting precise assembly tasks with 2 mm tolerance on real hardware. Also explored training a small policy with offline RL and fine-tuning it online (following residual assembly and WSRL).
- Benchmarked existing VLA models (π0, π0.5, GR00T N1.7, OpenVLA-OFT) across multiple robot platforms, including the RB-Y1 humanoid, and built a unified VLA deployment framework for systematic evaluation from coarse- to fine-grained tasks.
- Collaborated with academic labs, industry partners, and startups.
End-to-end policies for bi-manual assembly
- Designed an imitation-learning architecture that incorporates force/torque sensing for high-precision bi-manual assembly, building on Octo. Added vision backbones for stronger spatial perception and built tooling that automates data acquisition.
- Benchmarked behavior cloning models (ACT, Diffusion Policy, Octo) across diverse manipulation tasks.
Dexterous in-hand manipulation via reinforcement learning
- Trained in-hand object rotation policies from scratch in Isaac Lab with a multi-fingered robotic hand, building on HORA (In-Hand Object Rotation via Rapid Motor Adaptation, CoRL 2022). HORA trains a policy in simulation with privileged object information (size, mass, friction), then trains an adaptation module that estimates those properties from proprioception alone, so the policy can be deployed on the real hand without vision. I trained the policy in Isaac Lab and deployed it on the real multi-fingered hand.
Sound source localization for Samsung Ballie
- Developed a lightweight MobileNetV3-based SSL model robust to 0 dB SNR conditions.
- Deployed and tuned a classical SSL pipeline on the Ballie platform.