I work on robot learning: training manipulation policies in simulation and making them work on real hardware. I recently completed an MSc in Robotics & AI at Newcastle University (Distinction), where my dissertation transferred reinforcement-learning policies from simulation to a real UR3e arm and distilled them into vision-language-action models.
Before that I spent 13+ years as an engineer in Syria, China, Ghana and Türkiye: running a 20-machine CNC department, commissioning robotic production lines for solar components, and building machine-learning forecasting pipelines for humanitarian operations across 1,328 sites.
- Research interests: sim-to-real RL, diffusion-based domain randomisation, vision-language-action models, embodied AI
- Engineering: robot kinematics and control, edge computer vision, PLCs and factory automation
- Data & optimisation: forecasting, robust multi-objective optimisation under uncertainty
- Open to robotics and machine-learning engineering roles in the UK
| Project | What it does | Built with |
|---|---|---|
| PGDA | Physically-Grounded Diffusion Augmentation (PGDA): Coupling diffusion-generated material textures to physical parameters for sim-to-real robotic manipulation. Code for the paper 'Used but Not Needed: Testing Whether Manipulation Policies Rely on Visual Material Cues in Domain Randomisation' submitted to Robotics and Autonomous Systems (RAS). | Python diffusion-models domain-randomisation material-perception |
| kurnool-thermal-dev | ||
| pv-thermal-defect-mapper | Maps hot cells, failed bypass diodes and offline strings in drone thermal images of a solar farm to the exact module and cell, using the flight plan and pose refinement from the plant itself, then plans a glare-free re-inspection flight. Tested on a thermal digital twin and on a real flight (PV Hawk). | Python computer-vision defect-detection digital-twin |
| pv-drone-inspection-planner | Drone mission planner for thermal inspection of solar farms: finds the valid irradiance window, keeps the sun's reflection out of the camera, splits batteries, exports ArduPilot/QGroundControl missions. Flown and verified in ArduPilot SITL. | Python ardupilot drone inspection |
| ros2-ur-ik-service | ROS 2 inverse and forward kinematics services for Universal Robots UR3e, UR5e and UR10e: closed-form, all 8 solutions, closest to the current joint state. Humble and Jazzy, Docker-ready. | Python docker inverse-kinematics python |
| pv-power-forecasting | Solar PV power forecasting from open weather data: pvlib physics + gradient boosting with conformal prediction intervals, 30 min to day-ahead, tested on a year of NREL measurements. | Python conformal-prediction forecasting gradient-boosting |
| robust-moo-water-energy | Scenario-based robust multi-objective planning of solar water-energy systems under climate, fuel-price and grid uncertainty (Pyomo + HiGHS): CVaR, worst-case, AUGMECON Pareto fronts, TOPSIS. | Python climate-adaptation cvar humanitarian |
| ur3e-reach-rl | UR3e reaching with reinforcement learning: MuJoCo + Gymnasium environment built from the real DH model, SAC and SAC+HER baselines with trained policies. | Python gymnasium hindsight-experience-replay manipulation |
This list refreshes itself every day through GitHub Actions.
Physically-Grounded Diffusion Augmentation for robust robot manipulation. Standard domain randomisation varies appearance and physics independently, so a policy cannot learn that an object which looks like rubber also behaves like rubber. PGDA binds Stable Diffusion material textures to matching mass, friction and inertia ranges, then trains a SAC+HER pushing policy with a ResNet-18 visual encoder for a UR3e arm.
| Experiment | Result |
|---|---|
| Mismatched texture (ablation) | Success drops by 43.4 pp: the policy genuinely relies on material cues |
| Real UR3e, material label supplied | 96.7 % success vs 63–70 % for standard domain randomisation |
| Distilled into an Octo VLA model (simulation) | 59 % vs 41 % for the baseline |
Paper in preparation for IEEE Robotics and Automation Letters.
- Alfandi, M. & Seçkiner, S. U.(2026). A scenario-based robust multi-objective optimization framework for climate-resilient water-energy systems. Journal of Cleaner Production. Article
- Alfandi, M. & Seçkiner, S. U. (2025). Renewable energy and greenhouse gas reduction for water supply in conflict-affected Syria. Renewable and Sustainable Energy Reviews, 211, 115353. doi:10.1016/j.rser.2025.115353
- Alfandi, M. & Seçkiner, S. U. (2022). Robotic process automation: a literature review on quantitative benefits. Proc. Int. Conf. on Industrial Engineering and Operations Management, Istanbul, pp. 280–295.
Machine learning
Robotics
Tools
Design & manufacturing
| Years | Role | Where |
|---|---|---|
| 2025–2026 | MSc Robotics & Artificial Intelligence (Distinction) | Newcastle University, UK |
| 2021–2024 | Postgraduate Researcher, Industrial Engineering | Gaziantep University, Türkiye |
| 2020–2024 | Technical Programme Manager / Data Engineer | Relief International & IHR, Türkiye |
| 2018–2020 | Technical Manager, Automation & Renewable Energy | Falioun Intelligente Dynamik, Türkiye |
| 2015–2017 | Head of CNC Department | United Fabrication, Ghana |
| 2014–2015 | Founder & Factory Manager | Guangzhou, China |
| 2010–2014 | Engineering Team Lead | Directorate of Engineering and Services, Syria |
The banner is a UR3e doing pick-and-place, with every frame solved by the analytical inverse
kinematics in ur-kinematics
(see scripts/make_banner.py).