Jay Prajapati
Portfolio

In this project, I developed a software module for a multi-robot system, focusing on coordination and communication between robots. The system was designed to optimize task execution by distributing tasks among multiple robots, ensuring efficient task completion with minimal redundancy. I used ROS to manage inter-robot communication and integrated sensors for real-time data exchange, enabling the robots to adjust their paths based on the actions of other robots in the system. This project demonstrated my ability to build complex systems that require real-time collaboration between autonomous robots.

• View Project

This project involved the development of dynamic path-planning algorithms for an autonomous robot to navigate unpredictable environments. Using ROS and Gazebo for simulation, I implemented global path planning using Dijkstra and A* algorithms, while employing local planners for real-time obstacle avoidance. The robot successfully navigated a test environment while dynamically adjusting its path based on sensor input. This project not only refined my understanding of planning algorithms but also helped me master ROS and simulation tools like Rviz and Gazebo.

• Project-1
• Project-2

This project aimed to assist visually impaired individuals by developing an AI-driven image captioning system. I trained a deep learning model using the Inception V3 CNN encoder for feature extraction, reaching a 92% accuracy rate in generating descriptive captions from real-time video feeds. To optimize real-time performance, I implemented Block Static Expansion and multi-headed attention vectors, which improved both accuracy and response time. Additionally, I developed Python scripts for seamless integration with mobile devices, enabling visually impaired users to capture and process video in real-time, with automatic caption generation and voice narration. This project highlighted my ability to apply AI solutions for real-world accessibility challenges.

• Code
• Video

This project revolved around real-time food item segmentation using deep learning techniques. I employed RNN-based ResNet50 and PointRend architectures to achieve an 80% accuracy rate in food item segmentation from live video feeds. Using transfer learning, I was able to improve model performance, reaching 67% test accuracy on a limited dataset of 4,935 training images and 2,135 test images, outperforming a custom model’s 43% accuracy. Fine-tuning pre-trained architectures allowed me to develop a reliable and efficient segmentation system, which was aimed at applications in food identification and inventory management in commercial kitchens.

• View Project

This project involved the design and development of an autonomous robot capable of picking and placing color-coded blocks in a designated area. I engineered a real-time color detection system using a Pi Camera, which was integrated with the A* path planning algorithm to ensure accurate retrieval and placement of the blocks. The robot was equipped with IMU sensors (MPU6050) and encoders for precise navigation, and I implemented PID control for stable maneuvering. Obstacle avoidance was achieved through the use of ultrasonic sensors, allowing the robot to navigate dynamically changing environments. The project achieved an 80% success rate in goal placement, demonstrating effective robot localization and task execution.

• View Project

In this project, I led a team to scan and 3D print miniature replicas of an 11ft statue to celebrate the 75th anniversary of my university. We used a 3D scanner to capture the intricate details of the original statue, and I led the scanning process to ensure accuracy and completeness of the data. Using Fusion 360, I performed reverse engineering to create a CAD model from the scanned data. I then used 3D printing technology to produce prototypes with different printing parameters, optimizing for quality, speed, and material usage. The final models were produced and distributed to over 1000 alumni attendees.

• View Project

This project focused on building a human tracking system using advanced object detection algorithms in C++. I integrated the YOLOv3 model for real-time detection, which achieved an 84% confidence level and 99 ms inference time on the MS-COCO dataset. I also incorporated a KCF tracker for precise tracking across 164,000 images, enabling reliable performance in dynamic environments. This project demonstrated significant improvement in accuracy and efficiency. To further enhance development, I integrated CI/CD pipelines with Travis CI and Coveralls, which boosted code quality and testing coverage. The module was used in a multi-robot system, where the detection and tracking data fed into autonomous path-planning algorithms for mobile robots.

• View Project

In this perception-focused project, I worked with the TurtleBot 3 WafflePi to autonomously navigate through a set of waypoints on a flat surface. My key contribution was calibrating the monocular camera using a checkerboard pattern to derive intrinsic and extrinsic parameters. From there, I calculated the homography of the floor, identified the vanishing point, and determined the horizon line, enabling the robot to detect and follow waypoints seamlessly. Additionally, I integrated YOLO-based stop sign detection, allowing the robot to recognize and respond to stop signs in varying orientations. This project sharpened my skills in perception, camera calibration, and robot path planning.

I co-invented a patented IoT-based wastewater spillage detection system designed to monitor and alert authorities of any spillage in real-time. The system integrates multiple sensors to detect abnormal spillage patterns, and the data is transmitted via IoT devices to a centralized system for monitoring and action. This project combined sensor networks, data transmission protocols, and environmental monitoring technologies to create a solution that can have a significant impact on environmental safety.

• View Project

I developed a RESTful API server to manage a collection of OpenRAVE robots using Debian , Docker, and Python. This project aimed to provide a scalable and efficient solution for handling multiple robot configurations stored in JSON format. I utilized Python's Flask framework to implement CRUD operations, allowing for seamless robot data management through endpoints for adding, updating, and deleting robot specifications. Additionally, I containerized the API server using Docker to ensure consistent deployment and compatibility across different environments. The server could efficiently parse JSON requests, validate robot parameters, and provide real-time data response, thereby improving the robotic fleet management workflow. This project deepened my understanding of REST API architecture, Docker containerization, and robotic data handling.

Conducted a comprehensive evaluation of 2D mapping algorithms, including HectorSLAM, Cartographer, and Gmapping. Mapped a 20,000 sqft hostel floor to validate the study, providing actionable insights for optimal mapping algorithm selection.

This project involved setting up a robotic control system for a collaborative robot (FRCobot) using ROS2, leveraging its real-time communication capabilities with DDS (Data Distribution Service) for reliable control. I created a ROS2 workspace and developed custom Python nodes for joint control, sensor processing, and safety features. Using URDF models, I defined the robot's kinematics and visualized it in RViz2, ensuring accurate transformations. I also configured MoveIt2 for motion planning, enabling pick-and-place and obstacle avoidance tasks, and integrated Gazebo for simulation to validate control algorithms before hardware deployment.

• View Project

I designed and developed an autonomous rover capable of navigating both indoor and outdoor environments using SLAM-based and GPS-based navigation, respectively. The rover's indoor navigation was achieved using the Gmapping SLAM algorithm, which combined LIDAR and wheel odometry data for precise localization and mapping of the environment. For outdoor navigation, I integrated a GPS module that allowed the rover to accurately navigate through predefined waypoints in open areas. The project also involved developing a ROS-based control system to seamlessly switch between SLAM and GPS modes, ensuring continuous navigation without manual intervention. I implemented an Extended Kalman Filter (EKF) to fuse sensor data, achieving high accuracy in localization and reducing drift over time. Despite constraints on funding and a tight timeline, I sourced aluminum materials and fabricated custom mounts to house the sensors and electronic components, maintaining a balance between durability and weight. The final prototype demonstrated robust autonomous navigation in a mixed environment, showcasing my proficiency in integrating SLAM, GPS, and ROS, as well as expertise in sensor fusion techniques for practical robotic applications.

• View Project