Samsung Research
As a Robotics AI Engineer at Samsung Research, I conducted research on imitation learning, vision-language-action models (VLA model) and reinforcement learning for robot manipulation and humanoids (e.g. nvidia promotion video (1-second appearance at 2:45), Samsung Tech Conference 2025 (7:15~7:35)). I also developed a sound source localization algorithm for Samsung Ballie (shown at 0:48).
Period: 2022.07 - present
Projects
My work at Samsung Research focuses on two main areas:
- Developing VLA models (e.g., OpenVLA, π0) and imitation learning models (e.g., ACT, Diffusion Policy) for robot manipulation and humanoid control.
- Applying reinforcement learning to fine-tune these models for precise tasks.
In the imitation learning project, I designed a model architecture that incorporates force/torque (FT) sensors to enable high-precision manipulation. To further enhance performance, I integrated various vision backbones to leverage spatial perception and developed algorithms that automate the data acquisition process. (My work was based on OCTO)
For VLA models, I evaluated various architectures across multiple robot platforms and built a unified VLA deployment framework, enabling systematic benchmarking from coarse-grained to fine-grained tasks.
In the reinforcement learning fine-tuning project, I worked on training the residuals of pre-trained imitation learning and VLA models, which enabled the application of learning-based algorithms for precision tasks. Furthermore, my research involved an approach of training a relatively small reinforcement learning policy with offline RL and then fine-tuning it with online RL. (My work was based on paper1 and paper2)
Additionally, I also have experience developing a light-weight Sound Source Localization (SSL) algorithm for the robot, Samsung Ballie.
A brief introduction to a humanoid manipulation demo I worked on can be seen at 5:35.
A short demonstration of the humanoid control I worked on appears at 2:45 for one second.
The SSL feature I worked on for Samsung Ballie appears at 0:48.