Tasfia Noor Chowdhury · Robotics & AI researcher
Mechatronics engineer and M.Sc. researcher at the University of Dhaka, working on robot learning from demonstration and medical image analysis.
A good model is one people can trust, not just one that scores well.
I want robots and AI systems that learn from people quickly and can be trusted around them. My PhD goal is to bring three strands of my work together.
Robots that pick up a skill from one or a few human demonstrations and still generalise safely. This builds on our one-demo-to-fifty work in MuJoCo.
Using motion and muscle signals (IMU, EMG) to understand what a person intends, so a robot or assistive device can adapt to them.
Uncertainty-aware deep learning for medical imaging, starting from my M.Sc. thesis on knee MRI at the University of Dhaka.
Five projects, from wearable sensing for exoskeletons and robot arms that learn from one demonstration to hospital data at scale.
Wearable IMU, EMG and goniometer signals predict whether someone is about to walk, take a ramp, use stairs or stop. The prediction then switches the controller of a simulated knee exoskeleton in MuJoCo.
A webcam hand-tracking interface teaches a simulated Franka Panda arm in MuJoCo. One recorded demo becomes 50 physics-validated training episodes.
A TurtleBot3 learns local navigation from teleoperation in a Webots warehouse (ROS 2 Jazzy), wrapped in a safety layer that reasons about gaps between moving obstacles.
Pixel-wise water segmentation for emergency response, trained on a new dataset that combines benchmark, drone, in-field and social-media flood images.
A Spark pipeline on AWS over New York inpatient records, finding what drives length of stay and predicting it at admission with tree ensembles, process mining and simulation.
Report what the data supports, including the negative results. Five seeds, paired tests, and no claims a comparison can't back up.
Robots that learn from how people already move, and models that tell a clinician how confident they are.
Five years teaching Physics, Mathematics and IT. If I can't explain something simply, I don't understand it yet.
Imitation learning, behaviour cloning, MLP policies, demonstration augmentation
ROS 2 (Jazzy, Nav2), MuJoCo, Webots, Gazebo, TurtleBot3, UR3 cobot
PyTorch, TensorFlow, U-Net, ResNet, DeepLab v3, diffusion models
Segmentation, hand-landmark tracking, MRI analysis
Scikit-learn, XGBoost, LightGBM, PCA, cross-validation, Wilcoxon tests
Apache Spark, AWS clusters, Pandas, NumPy, Power BI
Siemens TIA Portal, PLC ladder logic, Factory I/O, Arduino, sensors
Python, C, C++, MATLAB/Simulink, HTML/CSS/JS, SolidWorks, AutoCAD
On GitHub I share research code, robotics notes and small tools I build. I also write a weekly newsletter on LinkedIn.
An interactive master guide to industrial robotics and automation: hardware, trajectory maths with a calculator, CIM programming and cobots.
/ research codeThe code behind my first-author arXiv paper: Spark, PCA and tree ensembles on 2.3M records.
Predicting hospital length of stay from 2.3M inpatient records with ML, process mining and simulation (arXiv:2501.18535)
Jupyter Notebook Robotics-PlaybookA brief discussion of heavily discussed robot basics
HTML Zakat-Calculator-with-different-PuritiesZakat Pro: calculates Zakat on gold, silver and liquid assets of different purities, with a "Golden Glass" interface.
HTML City-Tech-ITBusiness portfolio website for City Tech IT
HTMLMy weekly LinkedIn newsletter. Latest on robotics: Giving Robots a Sense of “Up” (Mar 2026)
Presenting our paper on turning one hand-gesture demonstration into fifty robot skills in MuJoCo.
Engineer by training, researcher by choice, teacher by habit. Based in Dhaka, Bangladesh.
Leadership, energy and analytics training from 2025. Click a certificate to see it in full.
I'm looking for funded PhD positions in robot learning, human-robot interaction and medical AI. Supervisors and collaborators are welcome to get in touch.