DREAM – Distributed, Robust and Efficient AI for Autonomous Vehicles

This project addresses challenges with federated learning for large-scale real-time use in autonomous vehicles. Autonomous vehicle systems rely on large amounts of multimodal sensor data, where centralized processing is impractical due to privacy concerns and communication limitations. The project aims to make federated learning more efficient by using self-supervised learning to train models with mainly unannotated data. The project will further investigate knowledge distillation as a possible solution for transferring knowledge between different models with varying architectures when the platform changes. 

To streamline data exchange in federated learning, methods for compression and aggregation are studied. A large multimodal dataset will be collected using a fleet of vehicles equipped with sensors. This dataset will initially be shared within the project consortium but with the goal of eventually make the dataset public. 

The project contributes to advanced digitalization of Swedish industry. The work is important for the Swedish automotive industry, but the knowledge and expertise gained will be applicable in all domains related to federated learning, multimodal data and where there is a large number of nodes with limited embedded resources. The project is based on a strong collaboration between Zenseact, Scaleout, AI Sweden and RISE, and is expected to lead to more robust and efficient distributed AI systems.