Title:	Deep Learning for Autonomous Driving
DNr:	Berzelius-2026-281
Project Type:	LiU Berzelius
Principal Investigator:	Lars Hammarstrand <lars.hammarstrand@chalmers.se>
Affiliation:	Chalmers tekniska högskola
Duration:	2026-10-25 – 2027-05-01
Classification:	10207
Homepage:	https://github.com/zenseact
Keywords:

Abstract

Our application is intended to provide GPU resources for research conducted by at least 10 PhD students. The research focuses on deep learning for autonomous driving and spans three main areas: 1) neural rendering and neural simulation, 2) multimodal self-supervised learning and 3D scene understanding, and 3) planning, Vision-Language-Action (VLA) models, and federated learning. In neural rendering and simulation, the research focuses on reconstructing dynamic driving scenes from real-world sensor data, novel-view synthesis, and generating realistic observations from viewpoints that differ substantially from those recorded by the vehicle. The projects use techniques including 3D Gaussian Splatting, NeRFs, uncertainty estimation, and generative diffusion models. These methods can support scalable simulation and generation of safety-critical scenarios for autonomous driving. The multimodal self-supervised learning projects aim to exploit large amounts of unlabeled automotive data to learn representations and improve 3D scene understanding. The research includes self-supervised 3D occupancy prediction using supervision generated by Vision Foundation Models (VFMs), multimodal learning from heterogeneous sensors, self-distillation approaches such as JEPA, and methods for improving robustness to rare and challenging scenarios. The planning-related projects investigate learning-based approaches for autonomous driving decision-making, including imitation learning, reinforcement learning, Vision-Language-Action models, and federated learning. These projects study trajectory prediction and planning, reasoning for safety-critical decisions, the connection between linguistic reasoning and motion planning, and privacy-preserving policy learning across heterogeneous vehicle fleets. A common requirement across these projects is access to substantial GPU compute and storage. The research involves large-scale autonomous driving datasets, computationally intensive model training and inference, large numbers of parallel experiments, and extensive storage of intermediate representations, pseudo-labels, rendered data, annotations, and model checkpoints. Access to Berzelius is therefore essential for conducting these projects efficiently at the required scale.