UCL Industrial Research Project with Flower Labs

Supervised weekly by Dr. Javier Fernández-Marques (Lead Research Scientist, Flower Labs) and Dr. Chris Xiaoxuan Lu (UCL Computer Science).

Designed and trained privacy-preserving diffusion policies (conditional DDPMs) for visuomotor robot control, generating smooth 6-DoF joint trajectories from raw RGB images and proprioception.

Implemented end-to-end pipelines for decentralized and centralized learning of robot manipulation policies using the Flower federated learning framework across 8 clients (4 sim, 4 real).

Collected and synchronized large-scale multimodal datasets (RGB, joint states) with ROS 2, Gazebo and deployed policies on real-world Elephant myCobot 280 arms via PyMyCobot and Intel RealSense.

Benchmarked federated diffusion policies achieving stacking success comparable to centralized models while preserving data privacy.