Developed a dynamic, vectorized RaiSim environment and integrated deep reinforcement learning (PyTorch) to train a navigation policy on top of a low-level locomotion controller. Designed custom rewards to optimize point-to-goal navigation. Created diverse terrains and explored motion planning (RRT, A*) and dataset generation for imitation learning. Gained expertise in robot kinematics and frame transformations for legged robots.

Worked with a team of dentists at one of the leading hospitals in Pakistan to develop a deep-learning pipeline for teeth segmentation and labeling using U-NET and FRCNN. An Orthopantomograms (OPGs) dataset was self-annotated & processed for the project using PIL, numpy, JSON & OpenCV.

Designed and developed a contactless vitals measurement robot equipped with an Ackermann drive for my undergraduate final year project. The robot utilised a camera, machine learning, and signal processing to measure vitals and decrease medical staff contact time with contagious patients. The system reported around 70% accuracy for Blood Pressure, and 90% for Heart Rate, Oxygen Saturation, and Temperature. I presented the work as the first author at ICARA 2022 and received the Best Capstone Project award for this work.

Building a bottom-up simulation framework for mobile manipulation in PyBullet, to study and understand Task and Motion Planning (TAMP) for robotic systems like a mobile manipulator. A Clearpath Husky base with a Franka Panda arm executes multi-step language instructions by combining LLM-based task planning with end-effector camera perception and IK motion execution. The system successfully achieves reliable single-cube picks and two-cube stacking, while multi-cube stacking reveals challenges such as motion smoothness, base drift, and grasp retries. This ongoing project not only demonstrates the integration of symbolic task planning with continuous control but also serves as a platform to study richer perception, collision-aware motion planning, feedback-driven correction, and vision–language grounding to push toward general-purpose mobile manipulation.

This project is currently ongoing and further updates will be made as I progress.

Working on this project to understand and study how LLMs can be used for path planning. Current findings are that vanilla local policies achieve around 60% success on A*-solvable cases, while vanilla global one-shot planning reaches only about 25% and degrades more quickly as obstacle density increases. These initial results highlight the relative resilience of local prompting, but the work remains ongoing. I am systematically exploring improvements through richer local contexts (e.g., 5×5 windows, clearance features), prompt refinements (few-shots, ranking), and hybridization with classical priors (subgoal induction, A* cost-to-go). The broader goal is to examine how far prompt engineering and minimal feature additions can push LLM-based planning toward reliable performance, and to assess the feasibility of such methods for robotic path planning in more realistic settings.

This project is currently ongoing and further updates will be made as I progress.

Implemented RRT from scratch with custom collision detection conditions to move a 3 link arm in the presence of obstacles and walls. Simulated the environment and robot using PyGame and tuned RRT and PID parameters for optimal performance in varying configerations. Implemented the forward and inverse kinematics from scratch to compute path variables. Created multiple demos, to show different aspects such as wrist limitations, reachable space considerations, and RRT performance.

Tackled the challenge of navigating an unknown maze by implementing a combination of the Bug algorithm, collision avoidance, and CNN-based image classification to create a solution invariant to the maze structure, starting point, and end point. Optimized a CNN to detect signs with 99% accuracy by augmenting for invariance and extensively preprocessing the data. The motion plan integrated CNN-generated maneuvering commands, lidar-based wall detection for collision avoidance, the Bug algorithm as a fallback, and PID control to ensure accurate straight-line movement and sharp turns, achieving 100% navigation success.

Road Image Segmentation with Supervised & Unsupervised Methods

Supervised & Unsupervised ML, CARLA dataset

Deployed supervised and unsupervised methods like K-Means, DB Scan, Gaussian Mixture Models, and U-NET for segmenting road images from the CARLA dataset. Achieved a 95% Dice Score using U-NET for instance segmentation and a Silhouette Coefficient of 0.5 with K-Means and Gaussian Mixture Models.

Report

Pedestrian Behaviour Classification

Detection + Transfer Learning (VGG16)

Prepared a pedestrian image dataset by combining data from various sources and annotating pedestrian state as distracted or non-distracted. Detected pedestrians with 100% accuracy via FRCNN and identified distraction with 90% accuracy via transfer learning with VGG16.

Report

Led ideation to translate client needs into a scalable web solution, developed with Python (Flask), HTML, CSS, JavaScript, and deployed on AWS Lightsail. Engineered a Mapbox-based interactive globe enabling filterable data visualization to drive thematic storytelling. Working on the backend to support content automation and transition from manual Excel-based workflows to a database-driven system.

Worked with a Ph.D. student to enhance the circuit and mechanical design quality of a hip exoskeleton, resolving electronic and mechanical design issues by cleaning up circuits, fabricating new PCBs, 3D-printing worn-out clippers, and reinforcing components to reduce vibrations and eliminate circuit failures. Concurrently conducted data collection, amassing 60 hours of respiratory and EMG data from human subject trials, required by lab researchers to assess metabolic costs and muscle activity.

Spearheaded the establishment of a new robotics lab by creating an experiment bench with PhantomX manipulators and Intel RealSense D435i. Enabled students to test their robots for a robo soccer project by engineering a 6x8 feet robot tracking arena using Python, C, and camera vision.

Course Assisted: Physics 2211 - Matter and Interactions (Spring 2023, Fall 2023 & Spring 2024)

Conducted bi-weekly lab experiment and recitation sessions for an average class size of 60 students. Consistently rated 4–5/5 in student evaluations across all metrics, reflecting strong teaching impact.

Designed and instructed labs in Physics 101 and Basic Electronics, strengthening students’ grasp of fundamental physical principles and electronic circuits.

Courses Assisted: Calculus II (Spring 2019), Engineering Mathematics (Fall 2019), Electrical Machines (Spring & Fall 2020), Mechanics & Thermodynamics (Spring 2020), Probability & Statistics (Spring & Fall 2020), Electric Network Analysis (Spring 2021)

Managed teaching responsibilities for classes of 30–60 students, encompassing grading, assignment development, recitation and revision sessions, and individual tutoring support.

Coursework: Introduction to Robotics Research, Robotics, Machine Learning, Deep Learning, Deep Reinforcement Learning for Intelligent Control, Machine Vision, Linear Control Systems, Artificial Intelligence, Robotics Professional Prep Seminar 1–3, Robotics Capstone Project 1 & 2

Selected Coursework: Feedback and Control Theory, Computer Vision, Mobile Robotics, Instrumentation, Linear Algebra, Probability and Statistics, Calculus, Engineering Mathematics, Engineering Design