Intelligent Ground Vehicle Competition (IGVC)

I achieved 2nd and 3rd place in the Cyber and AutoNAV challenges, respectively, at the annual international Intelligent Ground Vehicle Competition hosted at Oakland University, Michigan. In this competition, multidisciplinary teams are tasked with designing and developing autonomous vehicles that can navigate and complete specific tasks according to a set of rules. During this competition, I developed a fully functioning autonomous vehicle equipped with advanced object detection, path planning, and traversal capabilities. My role involved integrating these technologies to ensure the vehicle could accurately detect and avoid obstacles, plan efficient routes, and traverse complex environments autonomously. This achievement showcases my practical skills in robotics and my ability to apply theoretical knowledge to real-world challenges.

Reinforcement Learning for Game Optimization

In my recent project, I optimized gameplay in BreakoutDeterministic using Deep Q-Networks (DQN) and Double DQN, achieving mean rewards of 8.04 and 10.1, respectively. The agents used an ε-greedy policy for balanced exploration and exploitation, and replay memory for stable learning. While DQN relied on reward prediction errors, Double DQN utilized a second network for more reliable training. Detailed analysis and visualizations are available on my GitHub.

Generalized RAcing Intelligence Competition (GRAIC)

Implemented path planning and control algorithms for an autonomous racing car in the CARLA simulator. Implemented Hybrid A* search for optimal waypoint navigation in complex environments, particularly racetracks with a Proportional-Derivative (PD) controller for real-time autonomous with obstacle avoidance and adjustments of steering angle, speed, and braking, enhancing the car’s efficient navigation through racetracks.

Autonomous Vehicle (GEM e2)

In Progress: Learning on using GEMstack to control the GEM e2 Vehicle

Reinforment Learning On Unitree Go1

In Progress: Implementing a RL controller to replace and benchmark the Baseline MPC controller

Vehicle Control

Designed and executed a control system for steering a vehicle (GEM e2) along a specific track, which was segmented into various waypoints to assist with navigation. This control system was comprised of two main elements: a longitudinal controller for managing the vehicle's speed and a lateral controller for handling its direction.

Lane Detection Using Computer Vision

Implemented a lane detection system for an autonomous vehicle using the ROS and Gazebo platforms, leveraging the Python OpenCV library. This system processes a video feed by dividing it into individual frames and annotates each frame to distinctly outline the lanes. The process begins with a function that applies both gradient and color thresholds to identify crucial image features, resulting in a binary image that emphasizes road edges and lane markings. Following this, a perspective transformation is applied to achieve a top-down view of the lanes, providing a 2D geometric depiction. The final step involves dividing the image into horizontal segments and using a histogram to locate the lane center, identifying the area with the highest pixel density as the lane's midpoint.

Robotis Mini

In Progress: Implementing Basic Kinematics to control the Mini Robot in real world and ROS

Lidar Based SLAM Implemention

Developed an algorithm using ROS and Python to gather robot data from a rosbag, including LiDAR measurements and a trajectory estimated by an Extended Kalman Filter (EKF). I utilized the split-and-merge line fitting algorithm to identify and fit corners in the LiDAR data. This information, along with the robot's trajectory, facilitated the creation of a geometric map of the robot's surroundings. Additionally, a map was constructed using Simultaneous Localization and Mapping (SLAM), with the red trace representing the results from my line-fitting algorithm, while the gray map was generated using Gmapping.

Semi-Soft Robotic Hand

Created a Soft Robotic Hand controlled by five individual stepper motors, enhancing dexterity and flexibility, with Arduino for control and 3D modeling and printing for construction

Robot localization Using Particle Filtering

Implemented a Monte Carlo Localization (MCL) algorithm in Python for vehicle localization within an unknown environment, utilizing simulations on ROS and Gazebo. The MCL algorithm initializes with a set of randomly generated particles and, as the vehicle moves, adjusts these particles to predict the vehicle's new position. This iterative process allows the particles to gradually converge towards the vehicle's actual location. The goal was to determine the vehicle's position using sensor data and a preconstructed environmental map. A specialized module was developed to process raw LiDAR sensor point cloud data, calculating the vehicle's distance from surrounding walls in four directions. The particle filter algorithm maintains information on each particle's position (x-y coordinates), orientation, weight, and an environmental map.

Chronic Disease Detection System using Machine Learning

Developed an expert system that uses patient data to diagnose a wide range of chronic conditions

We used Logistic Regression and Random Forest to detect chronic kidney disease; for diabetes, we used Logistic Regression (LR) and K-Nearest Neighbor (KNN); for heart disease, we used Random Forest Regression and Decision Tree; and for pneumonia and COVID-19, we used Convolutional Neural Network (CNN) on chest x-rays. I was primarily in charge of combining all four models and developing the model to detect pneumonia and COVID - 19.

Drone

Engineered a custom surveillance drone featuring a modular 3D-printed body and high-performance 1200KV BLDC motors, controlled via a Pix hawk Flight Controller and an ESC for motor control

Lane Detection using Hough Transform and Histogram

Demonstrated a perception algorithm that is entirely based on vision or camera input. I concentrated on presenting a powerful end-to-end lane recognition approach for self-driving automobiles utilising modern computer vision techniques. I begin with a rudimentary strategy based on edge detection and polynomial regression, which is the foundational approach for recognising only straight lane lines. Using the Histogram and the Hough Transform

e-Yantra Robotics Competition (eYRC) by IIT Bombay

Participated in a competition held by IIT Bombay. Our task was to make a drone for the delivery of parcels.