flower labs
London, UKFeb 2024 – May 2025
Industrial Project — Federated Diffusion Policies for Robot Learning
Key Achievements
End-to-End Neural Pipeline for Autonomous Cube Stacking
Built an end-to-end neural pipeline combining a ResNet34 encoder with conditional diffusion models, achieving 10Hz real-time inference for autonomous cube stacking tasks.
Conditional Diffusion ModelsResNet34Real-time Inference
Federated Learning Across Distributed Clients
Implemented federated learning across 8 distributed clients using the Flower framework, maintaining 92% of centralized model performance while preserving data privacy.
Federated LearningFlower FrameworkDistributed Computing
Automated Expert Demonstration Collection
Automated expert demonstration collection and integrated a PyTorch diffusion policy training loop, reducing training data requirements by 30% without sacrificing task accuracy.
PyTorchDiffusion PolicyData Collection
Technical Skills
Research
- Conditional Diffusion Models
- Federated Learning Systems
- Real-time Neural Inference
- Vision Encoders
Engineering
- Flower Framework
- PyTorch
- Distributed Computing