Projects

Frontier-based autonomous mapping of a GNSS-denied cave with a PX4/Gazebo quadrotor. Replaced GPS with external-vision EKF2 fusion, bridging RTAB-Map RGB-D odometry into PX4 and adding reactive lidar avoidance and a return-home behavior.

tech: ROS 2, PX4, Gazebo, RTAB-Map, RGB-D SLAM, Python, Lidar

• Built a PX4/Gazebo cave drone with RGB-D camera, 2D lidar, IMU, and a front spotlight.
• Configured EKF2 for GNSS-denied flight using external vision for position and heading.
• Bridged RTAB-Map odometry into PX4 VehicleOdometry with ENU/FLU to NED/FRD frame handling.
• Wrote a ROS 2 autonomy node for takeoff, cave entry, frontier exploration, local planning, avoidance, and return.

case-study github demo (todo)

Modified a Unitree Go2 with a 2-joint spine and trained an RL policy for tight-corner navigation in narrow hallways. Designed observation/action spaces using IMU, joint states, and relative obstacle-goal pose. Training done in Isaac Sim with iterative dynamics constraints.

tech: Reinforcement Learning, Isaac Sim, Python, Robot Dynamics, IMU

• Trained a locomotion policy for a Unitree Go2 quadruped under challenging terrain conditions.
• Designed observation and action spaces using IMU, joint states, and relative obstacle-goal pose.
• Iteratively tuned rewards and simulation constraints for robust tight-corner and hallway navigation.

case-study (on hold) github demo (todo)

Mobile robot navigation using a nonlinear cost function with motion-capture calibration, EKF filtering, and model predictive control.

IJRR manuscript submission is currently in progress for this project.

tech: ROS, C++, Motion Capture, EKF, MPC

• Programmed a mobile robot to move based on a nonlinear peer-aware cost function.
• Calibrated motion capture to estimate real-time positions of rigid bodies.
• Developed ROS nodes in C++ for real-time control and feedback collection.
• Used Extended Kalman Filter to reduce motion-capture noise.
• Applied Model Predictive Control for trajectory planning and robot motion control.
• Integrated Bayesian intent inference for belief-aware decision making.

case-study (on hold) github demo (todo)

Built a robot system that maps user language commands into pick/place actions using ROS2 orchestration and 3D perception.

tech: ROS2, Python, Computer Vision, NLP

• Programmed a robot to pick and place colored and sized blocks based on user instructions.
• Created a ROS2 orchestrator node to delegate tasks using ROS2 actions.
• Implemented stereo perception for object size, color, and depth estimation.
• Added retry and verification loops for failed manipulations.

case-study (on hold) github demo

Built and evaluated LQR-based controllers for motor position control and disturbance-robust cart-pole stabilization.

tech: SIMULINK, Arduino, ROS2, Gazebo, LQR, State Feedback, IIR Filtering

• Built an LQR position controller in SIMULINK and ported it to Arduino.
• Implemented state-feedback control with IIR velocity filtering to reject encoder noise.
• Developed and analyzed an LQR cart-pole controller in ROS2 and Gazebo under stochastic earthquake disturbances, tuning Q and R matrices to balance stability and control effort.
• Evaluated system performance using pole angle deviation, cart displacement, and control effort metrics.

case-study (on hold) github (todo) demo (todo)

Vision-based maze solving with VLM supervision and A* search for pathfinding.

tech: OpenCLIP, YOLO, CV2, A* Search

• Used OpenCLIP VLM as a supervisor agent to trigger the maze-solving pipeline.
• Leveraged YOLO and CV2 for computer vision to isolate the maze region.
• Implemented A* search to find maze solutions.
• Used multithreading to separate motion control and maze detection modules.

github (todo) demo

6-DOF robotic system from CAD and dynamic analysis to inverse kinematics and ROS-based network control. Placed 3rd out of 25 teams.

tech: SOLIDWORKS, MSC Adams, MATLAB, ROS

• Achieved 3rd position in the 2022 industrial open house out of 25 competing groups.
• Designed the robot using SOLIDWORKS and MSC Adams for dynamic analysis.
• Simulated inverse kinematic equations for joint-angle calculation in MATLAB.
• Developed ROS nodes for controlling the robot over the network.
• Tuned control parameters and sensors for real-robot operation.

github demo (todo)