Federated Fleet Learning - System topology

Important results from the project

The project met its goals partially. We delivered important results in system architecture, HIL validation, and cloud-vehicle data/model flows, and improved understanding of how safety-critical constraints affect federated learning in vehicles. The WP1 and WP2 targets of a functioning HIL setup with successful model training and scaling was achieved. At the same time, full in-vehicle training did not fit the timeframe, leading to a joint shift toward active learning and fleet analytics.

Expected long term effects

In the long term, the project is expected to enable a more scalable and data-driven vehicle development loop through better event selection, more efficient data collection, and clearer cloud-fleet feedback mechanisms. This is expected to increase development speed, customer satisfaction and road safety. The project also provides practical methods and decision support that reduce risk in future initiatives related to fleet learning, active learning, and operational fleet analytics.

Approach and implementation

The project followed a staged model: first an HIL phase to validate core federated-learning concepts on relevant hardware, then vehicle-side integration in a production-like environment, and finally a controlled shift toward active-learning and analytics flows as technical and safety constraints became clear. Collaboration between Zenseact, Volvo Cars, and AI Sweden was central, enabling both technical progress and a pragmatic reprioritization.