projects

Centralized Collaborative SLAM
Merged an in-house visual SLAM module with an open-source centralized optimization (COVINS) for collaborative map merging with multiple RC Cars. Deploy on a Jetson Nano Super and equip the robot with a Zed mini camera. Tools: C++, ROS 1, OpenCV, Ceres, Eigen

Fuzzified Transformer for EEG Decoding (BSc Thesis)
Implemented a Neuro-fuzzy block Type 1&2 and integrated into a Transformer Neural Network for EEG SSVEP classification to enhance noise robustness. Tools: Python, Scipy, Pytorch

Unified autonomous navigation pipeline for Turtlebot 3
Integrated perception, localization, and planning into a single pipeline to get a unified navigation ROS node for Turtlebot 3, adaptable to changing environments. Designed a two-fold map generation that fuses a static map from ORB-SLAM 3 and the projection of RGB-D images with YOLOv5 to detect moving people. Tools: C++, Python, ROS 1, OpenCV, g2o, Pytorch, Gazebo

Training a Lunar Lander with Reinforcement Learning
We implemented two RL algorithms, Deep-Q network and tile coding, to train the agent in the Lunar Lander environment from Gym library and compare their performance. Tools: Python, Tensorflow, Gym, OpenCV

Designing a non-linear controller for an EV with brushless motors We designed a cascade controller for speed control of an EV by combining a PID controller for low-level control of the brushless motor actuator with a chosen nonlinear controller. We evaluated Sliding mode controller, Feedback linearization, and Adaptive control architectures. Tools: MATLAB

Extending Boston Dynamics Spot Arm to 7 DoF We modify the Spot Arm to include an additional DoF for better dexterity in manipulation tasks. We 3D model the arm, import its mesh, and create a URDF file for loading in Gazebo. We also analyzed the direct and inverse kinematics of the arm, and designed two controllers to handle the kinematic and dynamic behaviour. Tools: ROS 1, C++, Python, Fusion 360, RBDL, Gazebo

Automated shrimp feeding system powered by computer vision Integrated tank-mounted cameras with real-time image processing to monitor food density to automate the shrimp feeding with a low-cost computer platform. I was in charge of the communication protocols (I2C and Serial) between the main system (Raspberry PI-3) and an MCU (Arduino Pro Mini), and real-time food segmentation. Tools: Python, I2C, Raspberry-PI, OpenCV

Synthesis and comparison of two 8-bar linkages for exoskeletons We synthesized two 8-bar linkages (Peaucellier-Lipkin) using real-world data from Tracker to model human gait. We compared several synthesis results using different optimizers (gradient descent, mesh search and stochastic solvers) Tools: MATLAB, Tracker Toolbox, Fusion 360.