DREAM – Distributed, Robust and Efficient AI for Autonomous Vehicles

Purpose and goal

The purpose of DREAM is to advance federated learning for autonomous vehicles by improving efficiency, robustness, and adaptability in large-scale, real-time use. The project explores self-supervised learning to reduce annotation costs, knowledge distillation to adapt models when sensors and hardware change, and communication optimisation to enable deployment in fleets. The overall goal is to develop safer, generalizable AI that strengthens Sweden’s automotive competitiveness.

Expected effects and result

The project will deliver concrete advances including a large multimodal dataset for federated self-supervised learning, validated methods for knowledge transfer across heterogeneous platforms, and efficient communication protocols for vehicle fleets. Results will be shared openly, supporting research and industrial innovation beyond the consortium. The expected effect is enhanced road safety, reduced costs, and strengthened global leadership for Sweden in digitalized and sustainable mobility.

Planned approach and implementation

The project is structured into five integrated work packages addressing use-case definition, federated self-supervised learning, communication efficiency, testbed deployment, and project management. The work will follow a customary project structure with interrelated work packages. Key activities include large-scale data collection, development of advanced methods, and real-world validation. Implementation is carried out in close collaboration between RISE, Zenseact, Scaleout, and AI Sweden.