Learning from demonstration
Using a small number of human demonstrations to bootstrap reinforcement learning, so a robot needs far less trial and error to master a manipulation skill.
Machine Learning Scientist · Robot Learning
Karim, to most people.
I work on deep reinforcement learning for robots: teaching them manipulation skills from a handful of human demonstrations, and from rewards they learn to recognise in camera images instead of rewards written by hand.
I'm a machine learning scientist at Zebra Technologies, working across deep learning, reinforcement learning and robotics.
I did my PhD in Computer Science at the Lincoln Centre for Autonomous Systems (L-CAS), University of Lincoln, UK. My doctoral research was on deep reinforcement learning for robotic manipulation and grasping. It looked at how to combine human demonstrations with reinforcement learning, and how to learn reward functions from visual observations rather than engineering them for every task.
My route into robotics started with hardware. I studied mechatronics engineering at Palestine Polytechnic University and embedded systems at the University of Essex, and my first paper described a self-balancing two-wheel electric vehicle.
I'm always happy to talk about robot learning, share code from my projects, or explore a collaboration. Get in touch.
Using a small number of human demonstrations to bootstrap reinforcement learning, so a robot needs far less trial and error to master a manipulation skill.
Learning reward signals and task-success classifiers directly from camera images, instead of hand-engineering a reward function for every new task.
Scaling policy learning to multi-step tasks, where a robot has to reach several sub-goals in the right order to finish the job.
Robotics and Autonomous Systems, vol. 198, 105311
@article{mohtasib2026policy,
title = {Robot policy learning from demonstrations and visual rewards
for sequential manipulation tasks},
author = {Mohtasib, Abdalkarim and Cuay{\'a}huitl, Heriberto},
journal = {Robotics and Autonomous Systems},
volume = {198},
pages = {105311},
year = {2026},
doi = {10.1016/j.robot.2025.105311}
}
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 7414–7420
@inproceedings{mohtasib2022awet,
title = {Robot Policy Learning from Demonstration Using Advantage
Weighting and Early Termination},
author = {Mohtasib, Abdalkarim and Neumann, Gerhard
and Cuay{\'a}huitl, Heriberto},
booktitle = {2022 IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS)},
pages = {7414--7420},
year = {2022},
doi = {10.1109/IROS47612.2022.9981056}
}
Towards Autonomous Robotic Systems (TAROS), Lecture Notes in Computer Science, pp. 3–13
@inproceedings{mohtasib2021dense,
title = {A Study on Dense and Sparse (Visual) Rewards in Robot Policy
Learning},
author = {Mohtasib, Abdalkarim and Neumann, Gerhard
and Cuay{\'a}huitl, Heriberto},
booktitle = {Towards Autonomous Robotic Systems (TAROS)},
series = {Lecture Notes in Computer Science},
publisher = {Springer},
pages = {3--13},
year = {2021},
doi = {10.1007/978-3-030-89177-0_1}
}
International Joint Conference on Neural Networks (IJCNN), pp. 1–8
@inproceedings{mohtasib2021classifiers,
title = {Neural Task Success Classifiers for Robotic Manipulation from
Few Real Demonstrations},
author = {Mohtasib, Abdalkarim and Amir Ghalamzan, E. and Bellotto, Nicola
and Cuay{\'a}huitl, Heriberto},
booktitle = {2021 International Joint Conference on Neural Networks (IJCNN)},
pages = {1--8},
year = {2021},
doi = {10.1109/IJCNN52387.2021.9534141}
}
9th International Symposium on Mechatronics and its Applications (ISMA), pp. 1–8
@inproceedings{mohtasib2013steve,
title = {Self-balancing two-wheel electric vehicle ({STEVE})},
author = {Mohtasib, A. M. and Shawar, M. H.},
booktitle = {2013 9th International Symposium on Mechatronics and its
Applications (ISMA)},
pages = {1--8},
year = {2013},
doi = {10.1109/ISMA.2013.6547384}
}
The most up-to-date list, with citation counts, is on Google Scholar.
Implementation of AWET (advantage weighting and early termination), built on the Stable-Baselines3 versions of DDPG, TD3 and SAC. Code for the IROS 2022 paper [2].
RewardLearning ↗Eight neural classifiers, from fully convolutional nets to Transformers and domain-adaptation models, for learning the rewards of five robotic manipulation tasks. Datasets included.
Custom-Gym ↗Custom Gym environments: reacher and pusher tasks, each with dense-reward and sparse-reward variants.
Static_Features_Classifier ↗An attention-based RNN classifier for the static features of point-cloud clusters.
Zebra Technologies
University of Lincoln, UK · Lincoln Centre for Autonomous Systems (L-CAS)
University of Essex, Colchester, UK
Palestine Polytechnic University, Hebron, Palestine