About
I’m a PhD candidate in Electrical Engineering at National Taiwan University (NTU), working on multi-agent RL, safe control, and real-world UAV experiments. Before my PhD, I worked on localization for self-driving vehicles and robotics software / embedded systems.
What I build (quick snapshot)
- UAV autonomy: safe learning + formation control + onboard perception
- AMR fleet autonomy: route planning, deadlock resolution, validation pipelines (ROS2)
- Robust systems: performance-focused C++ implementations and real deployment constraints
Featured Videos & Demos
Safe Bearing-Only UAV Formation with CBF-QP Obstacle Avoidance
Problem: How can a team of UAVs preserve a rigid formation while navigating cluttered environments and avoiding collisions with obstacles, without relying on relative-position measurements?
What I built: The end-to-end system architecture for a multi-UAV formation — a distributed bearing-only sensing controller for formation maintenance, combined with a CBF-QP (Control Barrier Function – Quadratic Programming) safety filter for real-time obstacle avoidance.
How I validated it: First verified in simulation, then transferred to real-world flight experiments. Successful task execution on hardware validates the integrated navigation, localization, distributed control, and sensing pipeline shown in this video.
Impact: Accepted to IFAC 2026 (Busan).
UAV Mapping with RGBD Camera (Depth → Map)
Problem: GPS-denied mapping using onboard depth sensing.
What I built: ROS pipeline from depth → point cloud → mapping + tuning for stability.
Why it matters: Enables reliable autonomy and inspection without external infrastructure.
AMR Navigation + Task Planning with Behavior Trees
Problem: In warehouse environments, task execution can become hard to debug and maintain as the number of states, exceptions, and recovery behaviors grows.
What I built: A behavior-tree-based task planner in C++ to make execution logic modular (clear node responsibilities) and easy to extend with recovery/fallback behaviors.
How I validated it: Used Groot for real-time visualization and monitoring of task status, and integrated navigation with move_base_flex to support flexible planners/controllers and safer motion execution.
Impact: More transparent task status, faster debugging, and a more maintainable autonomy stack for real deployments.