Projects

STREAMS: Self-Training Robotic End-to-end Adaptive Multimodal Shared autonomy

STREAMS is a deep reinforcement learning framework that combines environment data and user input to produce smooth, stable end-effector trajectories for assistive robots, achieving a 96% success rate in simulations and 83% in user studies without requiring pre-existing datasets. GitHub Link

Kinova RL: a ROS package for Shared Control Deep Reinforcement Learning

This ROS package implements a shared control Deep Reinforcement Learning (DRL) system for robotic manipulation using a Kinova robotic arm. The system integrates visual input from a camera, head motion data from an IMU sensor, and a trained DQN (Deep Q-Network) agent to control the robot’s actions. GitHub Link

MAGnet: Magnetic Actuation for Guided Neurorehabilitation

This project introduces a novel approach to actuating rehabilitation robots using magnetic technology as a haptic force generator, combined with an Extended Kalman Filter for precise tracking and disturbance compensation in upper-limb motor rehabilitation. The study demonstrates improved smoothness, comfort, and safety in motor training, paving the way for more efficient and user-friendly rehabilitation technologies.

Adaptive Robotic Control for Users with Severe Impairments using DRL

A system that optimizes mapping from low-DoF inputs to high-dimensional robotic actions, enabling intuitive control for users with severe impairments. It uses adaptive goal prediction and reinforcement learning to guide actions in real-time, seamlessly blending user input with autonomous assistance. GitHub Link

Text-to-Speech App

This is a Streamlit-based web application that converts text to speech. It supports text extraction from uploaded Word and PDF files, manual text input, and text summarization using an LLM model. GitHub Link

LLM-based Autonomous Sorter

A simulation of an autonomous robot that can sort objects based on natural language commands. The robot uses a zero-shot classification model to interpret commands and can sort objects based on their shape or color. GitHub Link

Dual-Mode Robotic Arm Control with GUI and IMU Integration

This project integrates a graphical user interface (GUI) with an inertial measurement unit (IMU) sensor to provide dual-mode control of a robotic arm. Users can choose to control the robotic arm by clicking on cursor buttons in the GUI or by using the IMU sensor to move the cursor for button selection. GitHub Link

Real-time Face Orientation Detection

Developed a real-time video processing application that detects human faces and determines the orientation of the head for robot control applications. GitHub Link

Autonomous vision-based reach-to-grasp DQN agent

An autonomous vision-based reach-to-grasp system using a Deep Q-Network (DQN) agent for robotic arm control. This project utilizes real-time visual input to enable efficient and adaptive grasping in dynamic environments. GitHub Link

CartPole solution using DQN

his project implements a Deep Q-Network (DQN) to solve the CartPole-v1 environment from OpenAI Gym. The CartPole problem involves balancing a pole on a moving cart. This implementation uses a DQN to learn a policy that keeps the pole balanced for as long as possible. GitHub Link

Panorama Creation via Harris Corner Detection and Image Stitching

This project involves the creation of panoramic images by detecting key interest points in multiple images using the Harris corner detector. After identifying these points, the project matches them between images and computes the homography to accurately align and stitch the images together, resulting in panoramic views. GitHub Link

EEG Data Analysis

Developed a Python-based toolkit for processing and analyzing complex EEG datasets (64 channels, 640 time points, 99 trials). Implemented advanced signal processing techniques including epoch extraction, ERP computation, peak time identification, topographical mapping, and Laplacian filtering for enhanced spatial resolution. GitHub Link

IMDB Sentiment Analysis

Implemented a comprehensive sentiment analysis pipeline on the IMDB dataset, employing various text preprocessing techniques (stop word removal, stemming, lemmatization). Developed and compared multiple classification models for sentiment prediction, and evaluated different clustering methods to uncover patterns in the dataset. GitHub Link

GAN and DCGAN Implementation for Synthetic Digit Generation

Implemented Generative Adversarial Networks (GAN) and Deep Convolutional GANs (DCGAN) to synthesize realistic handwritten digits. Trained models on the MNIST dataset, comparing the performance and quality of generated images between standard GANs and DCGANs. Demonstrated proficiency in deep learning architectures, image generation techniques, and evaluating generative model outputs. GitHub Link