Research

Research areas

Reinforcement learning and vision-language-action models for robot behavior.

I am exploring hybrid policy improvement methods that combine on-policy and off-policy learning.

My current focus is on how these approaches can complement each other within a policy improvement framework.
I am reviewing related work and developing the research direction, with robot learning as a potential application.

I am interested in connecting visual observations and language-based task descriptions to executable robot actions.

In the AI Pet-Sitter project, I implemented EXAONE 4.5 reasoning over structured YOLOE observations and connected its outputs to robot action commands and voice warnings.
In Prebot, I am exploring SmolVLA and learning-based manipulation for café pick-and-place tasks, using manipulation as a setting for testing robot learning methods.

Failure-aware path planning and waypoint replanning for mobile robots.

My navigation research focuses on detecting failures and triggering replanning when a mobile robot cannot make progress.

I developed failure-aware LLM-DWA replanning that monitors navigation progress and recovery events to selectively request new waypoints.
The method was evaluated in static U-shaped obstacle and dynamic maze simulations, examining goal-reaching performance and the effects of replanning timing.

I am interested in how global planning, waypoint selection, and local control work together during navigation.

My LLM-DWA work combines LLM-generated waypoints with NavFn path planning and DWA-based local control.
I examine when a waypoint sequence should be revised as obstacles disrupt execution, and how replanning can support recovery without repeatedly interrupting ongoing navigation.