Research: deep reinforcement learning and autonomous systems
Autonomous drones
His research develops deep reinforcement learning agents for drones and asks how they should act when part of their sensing degrades or fails in flight.
See the researchAutonomous cars
Vehicles exchange bird's-eye views over vehicle-to-vehicle (V2V) links, so the affected car can still detect the road users around it. Reconstructing that view without inventing obstacles is the question behind a graduate project he supervises.
See the researchSensor fusion and computer vision
A camera degrades in fog well before LiDAR does. Fusing the two keeps detections confident as conditions worsen, the kind of multimodal perception on which his drone research depends.
See the researchGraph neural networks
Each layer of a graph neural network lets every node aggregate information from its neighbours. In a drone swarm, each drone hears only those within range, yet what one of them finds reaches the rest within a few hops. His graduate students have applied the same approach to misinformation and sentiment analysis on social media.
See the researchIllustration: a drone learns by trial and error to fly across a city, then flies the route on its LiDAR (orange: the path it plans) and keeps flying when part of the scan fails. Episode 1; sensor nominal; speed 0.0 m/s.
Illustration: the truck's LiDAR working; its view rebuilt from 0 of 3 nearby cars' scans, 0% of its surroundings; no pedestrian detected.
Illustration: clear night. Road users detected: camera 8 of 8, LiDAR 7 of 8, fused 8 of 8.
Illustration: a graph network, run here in your browser over the swarm's radio links, passes what each of 16 drones senses to its neighbours. Layer 0 of 3; 0 of 16 drones know of the find; 0% of the area searched.
His research aims to give unmanned aerial vehicles autonomy that is robust, reliable and trustworthy in complex, GPS-denied and degraded environments, moving them from lab prototypes to field-ready systems for infrastructure inspection, search and rescue, and environmental monitoring.
The guiding question
How can autonomous agents sense, decide and coordinate under uncertainty, when sensors fail, conditions change and maps are incomplete, while remaining safe and reliable?
Three connected directions form one autonomy stack: sensing, deciding under uncertainty, and acting as a team. Uncertainty is measured and passed along at every level.
Sense
Multimodal fusion that keeps a drone aware of its surroundings when a sensor fails, using generative models to repair missing signals in real time and reporting how confident each estimate is.
Decide
Reinforcement learning controllers that act on that uncertainty: slowing down in clutter, keeping wider clearance in poor visibility, and backed by lightweight safety layers.
Act together
Teams of drones that coordinate to build accurate maps rather than simply cover ground, and that reorganize when an agent fails or communication drops.
Each layer passes on how certain it is to the next.
This work develops deep reinforcement learning agents that navigate unmanned aerial vehicles (UAVs) through cluttered, dynamic environments using onboard camera input alone. Attention mechanisms direct the agent to the regions of the scene most relevant to avoiding obstacles.
This work examines how an autonomous system should perceive and act when a sensor degrades or fails, and how the missing information can be recovered without introducing spurious detections. For connected vehicles, cooperative perception over vehicle-to-vehicle (V2V) communication allows a vehicle whose LiDAR has failed to reconstruct its view of the road from data shared by nearby vehicles.
Current work, including the graduate project on cooperative perception that he supervises.
The same deep learning methods extend beyond aerial robotics to social network analysis, anomaly detection in Internet of Things data, and medical image analysis.
ConferenceIEEE SmartNets
Size-aware learning and contrast enhancement for multiple sclerosis lesion segmentation in brain MRIAref, Y., Awadallah, O., Charaf, N., AlMahamid, F., Sadhu, A. and Grolinger, K.
JournalScientific Reports, 15(1)
Agile DQN: Adaptive deep recurrent attention reinforcement learning for autonomous UAV obstacle avoidanceAlMahamid, F. and Grolinger, K.
ConferenceIEEE International Conference on New Trends in Computing Sciences
A multi-step comparative framework for anomaly detection in IoT data streamsAl-Qudah, M. and AlMahamid, F.
JournalDrones, 8(5), 173
VizNav: A modular off-policy deep reinforcement learning framework for vision-based autonomous UAV navigation in 3D dynamic environmentsAlMahamid, F. and Grolinger, K.
ThesisPh.D. thesis, Western University
Deep reinforcement learning for autonomous unmanned aerial vehicle navigationJournalEngineering Applications of Artificial Intelligence, 115, 105321
Autonomous unmanned aerial vehicle navigation using reinforcement learning: A systematic reviewAlMahamid, F. and Grolinger, K.
ConferenceIEEE CCECE
Agglomerative hierarchical clustering with dynamic time warping for household load curve clusteringAlMahamid, F. and Grolinger, K.
ConferenceIEEE CCECE
Virtual sensor middleware: Managing IoT data for the fog-cloud platformAlMahamid, F., Lutfiyya, H. and Grolinger, K.
ConferenceIEEE CCECE
Reinforcement learning algorithms: An overview and classificationAlMahamid, F. and Grolinger, K.
ThesisM.Sc. thesis, Western University
Virtual sensor middleware: A middleware for managing IoT data for the fog-cloud platformAlso listed on Google Scholar and ORCID.