Portfolio

A multi-link inverted pendulum that swings itself up and balances by combining a learned policy with classical optimal control. A Q-learning policy handles the nonlinear swing-up; an LQR regulator takes over for refined stabilization, with a dynamic threshold-switching strategy deciding when to hand off between them.

Built with discretized state variables, epsilon-greedy exploration, and real-time angle wrapping and matrix computation. The approach earned a Bronze Medal at the 2nd Koh Young AI Competition (ICCAS 2024).

An autonomous navigation system for a mobile robot performing mapping, localization, and waypoint navigation with LiDAR, the SLAM Toolbox, and the Nav2 stack in Gazebo Harmonic.

Alongside navigation, a ROS 2 logging node records pose, timestamp, and goal status at every waypoint, producing a tamper-resistant audit trail of the robot’s behavior, a step toward verifiable and accountable autonomy.

A field study injecting point-cloud data into the π0 vision-language-action family on an SO-101 arm, shown here as a side-by-side of the 3D-augmented policy against the 2D baseline.

Using a single 12 GB GPU, 70 real-robot demonstrations, and roughly 0.15% of parameters trainable, the method cut the frame-edge failure rate from 40% to 10% under appearance and viewpoint shift. The work received the Best Poster Award at the IEEE ARM 2026 Workshop on Safety-Critical Software, Control and Learning for Robotics.

A LeRobot-to-ROS 2 bridge that runs LeRobot-trained policies on a Yaskawa MH5LF through ros2_control, since Yaskawa is not natively supported. Then with Gazebo and Isaac Sim environments set up to compare policy transfer across simulators.

On real hardware, the accompanying work included eye-in-hand calibration on a Yaskawa MH5LS arm: solving the AX = XB hand-eye transformation in MATLAB to sub-millimetre accuracy against a Zivid benchmark, then integrating it into a ROS 2 and MoveIt 2 pipeline for perception-guided control. (Real-robot footage not shown here.)