Vernata: Self-Supervised Learning of LiDAR Point Representations

2026-08-07

AI Summary

This work addresses the scarcity of annotated outdoor LiDAR point clouds, a challenge stemming from the high cost of 3D labeling, by proposing a multimodal multi-teacher self-supervised distillation framework built upon the Sonata architecture. The approach enhances robustness to point cloud sparsity through sparse-view augmentation, incorporates a memory bank mechanism to stabilize training under resource constraints, and—uniquely—leverages high-resolution 2D image semantics to provide fine-grained self-supervised guidance for LiDAR point clouds. Evaluated on the TartanGround and Waymo datasets, the method achieves mIoU scores of 54.7 (+12.1%) and 57.1 (+14.7%), respectively, while maintaining strong performance even under weak modality conditions such as missing color or normal information.

Abstract

LiDAR serves as a primary sensing modality for robots operating in outdoor environments. However, the performance of deep learning models in this domain is severely limited by the scarcity of labeled data, a direct result of the high cost of 3D annotation. Self-supervised learning addresses this scarcity by learning general-purpose features from unlabeled data. In this work, we present a multi-modal, multi-teacher distillation framework for self-supervised learning on outdoor LiDAR point clouds. Building upon the Sonata architecture, we introduce Vernata, consisting of three extensions: sparse view augmentation to improve robustness against varying point densities, a memory bank mechanism to stabilize resource-constrained training, and cross-modal distillation utilizing dense, high-resolution 2D image features to enable fine-grained semantic guidance. We evaluate our method on the GrandTour, TartanGround, and Waymo datasets, as well as data collected from our own robotic platforms. Our experiments demonstrate a significant performance improvement over Sonata baselines, yielding mIoU scores of 54.7 on TartanGround (+5.9 points, +12.1%) and 57.1 on Waymo (+7.3 points, +14.7%). Finally, we show that the self-supervised approach maintains strong performance even in reduced-modality settings (lacking color or normals), achieving competitive mIoU scores of 49.4 and 50.2 on the respective datasets.

Problem

• LiDAR
• self-supervised learning
• 3D annotation scarcity
• outdoor point clouds
• labeled data scarcity

Innovation

• self-supervised learning
• LiDAR point clouds
• cross-modal distillation
• sparse view augmentation
• memory bank