Background and project description

Autonomous vehicles generate massive amounts of multi-modal sensor data, including camera images, lidar point clouds, radar measurements, GPS information, and vehicle control signals. These heterogeneous data sources provide complementary information that is essential for robust perception, localization, and decision-making. However, transferring such large volumes of data to centralized servers is often impractical due to bandwidth limitations, storage costs, privacy concerns, and regulatory constraints. Federated Learning (FL) offers a distributed and privacy-preserving framework that enables multiple vehicles, fleets, or organizations to collaboratively train machine learning models without sharing their raw data. While FL has shown significant promise for autonomous driving applications, its effectiveness is often limited by the availability of high-quality labeled data. Deep learning-based perception modules require high-quality annotations, which are costly and complex to obtain. Self-supervised learning (SSL) offers a solution by leveraging mostly unlabeled data with minimal labels. Early studies show that federated self-supervised training can achieve performance comparable to centralized approaches, with potential improvements as larger unlabeled datasets are used.

This thesis project aims to advance federated learning for autonomous vehicles by integrating self-supervised methods with robust aggregation techniques to develop models that are efficient, generalizable, and capable of handling both common and rare driving scenarios, while reducing reliance on manual annotation and avoiding the costs of central data storage.

The thesis is part of the research project DREAM – Distributed, Robust and Efficient AI for Autonomous Vehicles. The topic is highly relevant for enabling scalable and efficient AI development in next-generation autonomous driving systems.

Main Tasks

• In this master thesis project, you will focus on: 

• A novel self-supervised learning approach to exploit all available data on the central server, even with limited labels

• A hybrid federated learning scheme combining self-supervised and supervised techniques, adapted to local and global learning rounds

• Validation through extensive comparisons with fully supervised learning within the same federated scheme.

• Demonstration of the efficacy of combining self-supervised and supervised learning on the Zenseact Open Dataset (ZoD) under various federated scenarios.

• Present findings to the project partners

Qualifications

We are looking for one or two highly motivated students with a good general background in machine learning and computer vision. The following skills would be essential:

• Deep learning

• Computer vision

• Python programming

• Reading scientific papers

• Handling complex systems

• Federated learning (would be a bonus)

Conditions.

• Location: RISE, Kista, Stockholm

• Applications are reviewed on a rolling basis, apply as soon as possible, but no later than August 31st, 2026.

• Starting date: As soon as possible, not later than September 1st, 2026.

• Credits: 30 points

• Compensation: 39990 SEK upon a successful completion of a high-quality thesis.

Supervisors:

• Sima Sinaei (RISE)

• Henrik Abrahamsson (RISE)

Welcome with your application!

Send in your application (CV, motivation letter, transcript of records) no later than August 31st.

For any questions, please contact:

• Sima Sinaei, sima.sinaei@ri.se

• Henrik Abrahamsson, henrik.abrahamsson@ri.se