profile photo

Rayyan Rafikh

AI/ML Engineer, IpserLab
MS Robotics, Northeastern University

Email  /  Google Scholar  /  ORCID  /  LinkedIn  /  GitHub  /  CV


About Me

I am a robotics engineer and a graduate of Northeastern University, where I completed my master’s degree in Robotics. I currently work as an AI/ML Engineer at IpserLab.

My work lies at the intersection of perception, planning and learning for robots using classical and novel algorithms, optimization techniques, Machine Learning and Reinforcement Learning.

I aim to apply AI and build systems such as autonomous vehicles and drones, mobile and service robots, and swarm and multi-robot systems that help solve real-world problems and contribute to the UN Sustainable Development Goals, in areas including Space Exploration, Defense, Agriculture, Disaster Management, Climate Action and Governance.

My research has been published in the Journal of Robotics and Mechatronics and at international conferences, including one published by Springer, on multi-drone coordination for search and rescue, and on error compensation in delta robots.

Previously, I earned a bachelor’s degree in Mechatronics Engineering from Manipal Institute of Technology.

I have worked across robotics, automation and applied AI. At Black-i Robotics, I worked on a vision-guided robotic palletizing system. At Breaking Inc., I developed hardware for scaling up a bioreactor platform for microbial growth. At Sultan Qaboos University, I worked on an autonomous vessel for detecting and sampling oil spills, and at Samsung, I helped automate a product inspection line.

In my free time, I enjoy swimming, playing football, reading books and watching movies.


Publications
A Scoping Review on Unmanned Aerial Vehicles in Disaster Management: Challenges and Opportunities
Vishnu G. Nair, Jeane Marena D'Souza, Rayyan Muhammad Rafikh
Journal of Robotics and Control, 2024
Paper  

We explore the transformative role of drones in disaster management, focusing on their use in response and recovery operations for both natural and man-made incidents; and investigate advancements in AI and sensor technology, while addressing key challenges like privacy concerns, regulatory barriers, and ethical dilemmas.

Autonomous Navigation System for Multi-Quadrotor Coordination and Human Detection in Search and Rescue
Vishnu G. Nair, Jeane Marena D'Souza, Rayyan Muhammad Rafikh
Journal of Robotics and Mechatronics, 2023
Paper  

This work focuses on developing a method for simulating life detection during sudden disasters using deep learning models. It also involves coordinating multiple UAVs through region-partitioning techniques to accelerate search and rescue operations, tested in a simulated platform resembling real-life disaster environments.

Simulated Evaluation of Navigation System for Multi-quadrotor Coordination in Search and Rescue
Rayyan Muhammad Rafikh, Jeane Marena D'Souza
International Conference on Robotics, Control, Automation and Artificial Intelligence, 2022
Paper  

This proposes a simulation method for detecting life in sudden disasters using deep learning models, particularly Convolutional Neural Networks. It also involves multi-quadrotor coordination through region-partitioning techniques to enhance the efficiency of search and rescue operations. Additionally, the paper explores other non-conventional methods to further improve disaster response.

Position error estimation and compensation of 3-DOF delta robot under the effect of link tolerances
Darren Alton Dsouza, Rayyan Muhammad Rafikh, Ankur Jaiswal
International Conference on Recent Advances in Modelling and Simulations Techniques in Engineering and Science, 2022
Paper  

My work investigates position error in a 3-DOF delta robot due to link tolerance. It includes a mathematical formulation of the problem, followed by a comparative analysis of the forward and inverse kinematics to assess the impact on robot performance.


Projects
Decentralized Multi-Rover Exploration with Learned Goals on Rough Terrain, Present
Github

I am currently building a decentralized exploration system in which two rovers map unknown rough terrain together, using ROS2. Each rover runs its own SLAM and shares maps directly with the other, so the system has no central server. I wrote a C++ frontier-detection node that finds the edges between explored and unexplored space. A PPO reinforcement-learning policy then picks each rover’s next goal from those frontiers, splitting the work so the rovers cover different regions. I’m now connecting this learned planner to an MPC trajectory tracker so the rovers can follow their planned paths reliably over uneven ground. Next, I’ll benchmark the full system against classical frontier-exploration baselines to measure how much the learned goal selection improves exploration speed and coverage.

MPC & Passive Wheel Design for a Wheeled-Legged Quadruped, Dec 2025
Github

This project builds a whole-body model predictive controller for a wheeled-legged quadruped. The controller uses iterative LQR on a single rigid-body dynamics model and includes gait scheduling for walking and trotting, running at 50 Hz in MuJoCo simulation. To keep the body steadier, I add thrusters with PID control to stabilize roll and pitch. In simulation, this lowers leg torque and pitch deviation compared with a purely legged baseline. On the hardware side, I design 3D-printed one-way ratchet wheels in CAD that let the robot roll forward passively without extra motors. I test the wheels on a Unitree Go2 and in a Spot simulation.

Quadruped Gait Learning for Planning & Control of Minitaur using Deep RL, Dec 2024
Github

This project uses deep reinforcement learning to teach the Minitaur quadruped to walk. I implement TD3 and PPO controllers for the robot in PyBullet through a Gymnasium interface, with high-dimensional state inputs. The policies output continuous joint torques, and I design composite rewards that balance forward progress, stability and energy efficiency. I tune both the hyperparameters and the reward terms, then compare the two algorithms through their learning curves and resulting gaits. The comparison shows that TD3 converges better and produces more robust walking than PPO.

Large-Scale Underwater Image Mosaicking with Factor Graphs, Oct 2024
Github

This project builds large-scale mosaics from underwater images by treating mosaicking as a factor graph problem. I model the camera poses and the relationships between overlapping images as a factor graph, then refine all the poses together using GTSAM so the final mosaic stays consistent across the whole scene. I evaluate the results by measuring drift and how well overlapping regions line up. I also study how loop closures and the structure of the graph affect reconstruction quality.

Design and Evaluation of an Incremental Structure-from-Motion Pipeline,May 2024
Github

This project builds an incremental Structure-from-Motion pipeline that reconstructs 3D scenes from sets of images. I detect and match features with SIFT, filter out bad matches with RANSAC, and estimate the relative camera motion from the essential matrix. I then triangulate 3D points and refine the camera poses and structure together with bundle adjustment in GTSAM. To evaluate the pipeline, I benchmark it against COLMAP on three multi-view datasets and analyze reconstruction quality, drift and the cases where it fails.

Real-Time Object Recognition with Classical Vision and Deep Features,Feb 2024
Github

This project compares classical computer vision with deep learning features for real-time object recognition. I build a segmentation pipeline that isolates objects using thresholding, morphological operations and blob detection. I then describe each object with Hu moments and classify it with k-nearest neighbors, which reaches about 71% accuracy. Next, I replace the handcrafted features with embeddings from a pre-trained CNN and compare the two approaches using confusion matrices. The deep features raise accuracy to about 91%.

Comparative Implementation of SLAM on Mobile Robots using Ultrasonic and LiDAR,Dec 2023
Github

This project compares SLAM on a mobile robot using two different sensors: a rotating ultrasonic sonar and a LiDAR. I build a differential-drive robot with a Raspberry Pi, an Arduino and electric drives, running ROS with code in Python and C++. To estimate the robot’s position, I fuse data from GPS, an IMU and wheel encoders. I then run closed-loop SLAM in RViz with each sensor and compare the resulting point clouds. The sonar and LiDAR point clouds match with about 80% accuracy.

Last updated: Nov 2026