From research papers to perception systems running on real robots and edge devices.
Ongoing Vision for deployed robots: face recognition and risk-area detection and avoidance, taken from data strategy and training to deployment on edge devices and robot fleets through ROS perception modules. Includes real-time multi-modal sensor fusion for terrain analysis.
Published Constraint-aware federated learning and system-aware LoRA scaling for language models running under device latency, memory and energy budgets. Two NeurIPS 2025 workshop papers.
Published Automated inspection of laser sensor coatings: ROI segmentation, GAN-based synthetic data and an unsupervised normalizing-flow detector, reaching over 92% accuracy, with a 3D-printed Raspberry Pi rig for the production line.
Completed Optimized YOLOv8 to detect plastic, paperboard and foam debris on production lines, deployed on NVIDIA Jetson Nano at 96% AP. Stable Diffusion editing was used to augment rare scenes.
Published Learning rare object categories for instance segmentation using freely available object segments. Published at ECCV 2022.