Point clouds represent 3D spatial information using a set of discrete points relative to the sensor’s origin. Such a representation captures global structure, but lacks local semantics, making it difficult to extract features rich in geometric detail. This aspect, however, remains largely overlooked in self-supervised pre-training tasks for LiDAR point clouds. To overcome this limitation, we propose a computationally efficient, yet highly effective approach: estimating local point cloud geometry from LiDAR data, specifically surface normals and curvature, and directly utilizing these features as network input. We demonstrate that incorporating our local geometric features improves the model’s ability to learn a more discriminative latent representation through a self-supervised contrastive pre-training. Our evaluation demonstrates that a model pre-trained with our proposed local features outperforms a model pre-trained on standard global features for a LiDAR semantic segmentation task.