30hp - Representation Learning for Improved Spatial Understanding in Autonomous Driving

Ingress: TRATON is a group of strong brands with a shared mission: transforming transportation together to create the future of sustainable transport solutions. Within TRATON, we include MAN, Scania, Volkswagen Truck & Bus, and International. A thesis project at TRATON is an excellent way to build valuable connections for your future working life. Many of our current employees started their career with a thesis project.

Background: Autonomous driving systems rely on high-quality labeled data, yet annotation is costly, slow, and can struggle with rare traffic scenarios. Self-Supervised Learning (SSL) offers a promising alternative by learning general representations directly from large amounts of unlabeled data. Depending on the choice of pre-text tasks, these representations can enhance spatial understanding, domain robustness, and downstream perception while reducing dependence on manual labels. Recent research in SSL for autonomous driving highlights 3D/4D occupancy prediction, generative reconstruction, and contrastive methods as promising pre-text training tasks. Works like UnO [1], DiO [2], and GASP [3] explore occupancy-centric objectives and foundation model distillation for spatial and temporal reasoning. In contrast, BEV-MAE [4] uses Masked Autoencoders (MAE) to reconstruct masked inputs, learning geometry and semantics through generative pre-training.

Target: Explore representation learning through SSL as a promising pre-training strategy for autonomous driving perception, with evaluation on downstream tasks, including for example 4D occupancy prediction [5]. The goal is to develop and evaluate pre-training strategies that produce transferable, scene-aware representations capturing space, geometry, and potentially also kinematics which are useful across tasks like occupancy flow, segmentation, and detection, aiming for stronger generalization to rare and challenging conditions.

Assignment: The assignment is divided into sub-tasks: 1.    Formulate the research questions and decide on appropriate benchmark to test 2.    Research promising SSL objectives for autonomous driving, utilizing sensor input modalities such as camera, LiDAR, and/or radar 3.    Implement pre-training and fine-tuning pipelines and downstream evaluation metrics according to relevant benchmarks 4.    Write and deliver a thesis report and presentation, documenting methods, experiments, ablations, conclusions, and insights

Education: Master (civilingenjör) in computer science, robotics, engineering physics, electrical engineering, or applied mathematics, preferably with specialization in artificial intelligence algorithms or computer vision. Pre-requisites include tested programming experience using Python (incl. PyTorch, TensorFlow, or Jax), and knowledge of machine learning theory.

Number of students: 1 Start date: January 2026 Estimated time needed: 20 weeks

Contact persons and supervisors: Jesper Eriksson, Industrial PhD student in Autonomous Research, AI Technologies, jesper.x.eriksson@scania.com, +46 70 087 84 33 Carol Yi Yang, Industrial PhD student in Autonomous Research, Perception, carol-yi.yang@scania.com, +46 70 081 17 51 Truls Nyberg, Industrial PostDoc in Autonomous Research, Autonomous Motion, truls.nyberg@scania.com 08 – 553 535 27

Application:  Enclose CV, cover letter and transcript of records.

A background check might be conducted for this position. We are conducting interviews continuously and may close the recruitment earlier than the date specified.