Hi, I'm Rithish

Trained deep learning classification models using ResNet18 (97% accuracy) and DenseNet121 (94% accuracy) for cancer cell classification from microscopy images. Implemented model explainability analysis using GradCAM, GradCAM++, EigenCAM, and ScoreCAM to interpret model decisions — ScoreCAM performed best.

Built a ROS2 package for Turtlebot Waffle (Jazzy Jalisco) to navigate an arena within 90 seconds. Implemented D* Lite for dynamic path planning and obstacle avoidance, and used 2D LiDAR point cloud data to generate an occupancy grid SLAM map in real time.

Gazebo Simulation — TurtleBot SLAM & D* Lite Navigation

Designed and validated a two-sided CUSUM fault detection framework for six-channel IMU fault monitoring, built upon a Random Walk EKF (RW-EKF) estimating accelerometer and gyroscope bias faults. Multi-objective optimisation via 200-point Latin Hypercube Sampling minimised detection delay while eliminating false alarms — from 105 to 0 with comparable mean detection (56.24s → 56.33s).

6-Channel CUSUM Fault Signal Estimates and Alarm Times

Designed and 3D printed a mobile manipulator robot. Integrated ultrasonic sensors and wheel encoders for odometry-based localisation, with finite state machines for autonomous task execution. Added a 2-DOF manipulator with end-effector control that held a pen to mark a target on a board — completing the full task in 1m 45s.

Programmed an e-puck robot for two autonomous tasks in a controlled arena. Obstacle avoidance: IR-sensor wall detection with state switching between explore and avoid, plus bias logic to escape tight corners. Object following: distance-based ToF tracking to approach or retreat safely, with a rotational search mode when the target is lost — fully autonomous using local sensing only.

E-Puck — Obstacle Avoidance & Object Following

Implemented procedural terrain generation using Perlin noise and evaluated A*, RRT, RRT*, and D* under partial environmental knowledge — analysing trade-offs between computational efficiency, path optimality, and adaptability. D* proved best on path-finding time.

Built a real-time digital twin for a 3D printer with 2-way Firebase data communication to detect process deviations as they occur. Implemented Random Forest (77%) and XGBoost (76%) classifiers for automated anomaly inspection, plus a Unity 3D app for state visualisation and bidirectional control. Accepted on Springer publication; ICAARS 2025.

Owned and led the design and build of a high-speed automated pick-and-place delta robot for food packaging, integrating RGBD camera object detection (93% accuracy). Managed cross-functional delivery across perception, electronics, and control sub-teams within a commercial MVP timeline via agile project management and systems integration.

Delta Robot and 3D Printed End Effector