Self-supervised Representation Learning for LiDAR Point Clouds - A Design Science Study of a Self-Supervised Model for Perception in Autonomous Driving

Abstract

Autonomous driving systems are large, complex software systems where accurate environmental perception is foundational to safe navigation. Perception systems often rely heavily on supervised deep learning models trained on large volumes of manually annotated datasets, resulting in performance heavily tied to the quality of the annotations. Self-supervised learning (SSL) offers a promising alternative by deriving supervisory signals directly from raw, unlabeled data, yet its application to LiDAR point clouds of autonomous driving data remains largely underexplored. We investigate whether a JEPA-based architecture, adapted to operate on LiDAR data, can learn high-quality representations without manual labels, and what that implies for the autonomous driving system during its evolution. We examine the meaning for software quality as well as the impact on the engineering process of building and updating the system. Through an iterative design process, we find that our SSL model substantially outperforms a fully supervised baseline in low label-budget regimes, and that fine-tuning the pre-trained backbone recovers nearly identical detection performance under the full label budget. Our results suggest that SSL pre-training is a viable architectural strategy for reducing annotation dependency and improving maintainability through backbone reuse, though these benefits come with meaningful upfront engineering complexity that should be weighed in practice.