LiAuto-GeoX: Efficient Grounded Driving Transformer

Jiawei Lian1,2,*, Haoyi Sun2,*, Yang Wu1,*, Lifu Mu2,*, Siyuan Wang1,2,
Le Hui3,4,†, Ning Mao2,†,‡, Tao Wei2, Pan Zhou2, Kun Zhan2, Jian Yang1,†
1Nanjing University of Science and Technology 2Li Auto Inc.
3Northwestern Polytechnical University 4Department of Computing, The Hong Kong Polytechnic University
*Equal Contribution †Corresponding Author ‡Project Leader
arXiv Code
Overview of LiAuto-GeoX

Overview of LiAuto-GeoX. LiAuto-GeoX provides an efficient driving geometry model for surround-view 3D reconstruction and downstream autonomy tasks. It strikes a compact yet effective accuracy–efficiency balance, accommodates diverse camera configurations, and transfers the learned dense geometry to tasks such as pose estimation, depth estimation, 3D reconstruction, trajectory prediction, occupancy prediction, and future occupancy forecasting, achieving excellent performance across all of them.

Abstract

Dense 3D reconstruction has demonstrated immense potential for spatial understanding, yet its viability as a real-time, onboard representation for autonomous driving remains an open challenge. Existing large-scale visual geometry models typically require substantial computational resources and lack the long-range geometric fidelity, surround-view consistency, and real-time efficiency demanded by dynamic driving environments. To bridge this gap, we present LiAuto-GeoX, an efficient grounded driving transformer designed for deployable, ego-centric 3D scene understanding. Our approach begins by learning a high-capacity driving geometry model from large-scale surround-view data, utilizing sparse LiDAR priors to provide robust geometric grounding in distant, ambiguous, or structure-sparse regions. We then instantiate this capability into a highly compact 155M-parameter onboard model through a novel geometry-preserving distillation framework. This framework employs mask-guided depth-aware distillation to retain fine-grained metric structures by emphasizing geometrically informative regions, and relative-pose relational distillation to enforce cross-view spatial consistency through pose-induced geometric relations. Extensive evaluations reveal that LiAuto-GeoX runs at 220 FPS on KITTI while maintaining high-fidelity dense reconstruction, enabling real-time deployment. The learned geometry transfers seamlessly to downstream autonomy tasks, achieving 90.6 PDMS in trajectory prediction, 24.63 mIoU in occupancy prediction, and 47.67 IoU in future-frame prediction. These all demonstrate that efficient dense 3D reconstruction can transcend its traditional role as a perception target to serve as a scalable, foundational geometric representation for next-generation autonomous driving.

Method

Overall pipeline of LiAuto-GeoX. LiAuto-GeoX first trains a high-capacity teacher model to learn dense driving geometry from calibrated multi-view RGB inputs, then distills its geometric capability into a compact student model with task supervision, token-mask condition, and relational constraints. During inference, only the student is deployed to produce dense 3D reconstructions under flexible surround-view camera configurations.

3D Reconstruction Visualization

nuScenes Scenes

DDAD Scenes

Waymo Scenes

OpenScene Scenes

PandaSet Scenes

Lyft Scenes

Argoverse 2 Scenes

Deployment Inference Speed

Qualitative camera trajectory predictions

GeoX Traj

VGGT Traj

Night Scene Trajectories

More Trajectory Predictions

BibTeX

@article{lian2026geox,
  author    = {Lian, Jiawei and Sun, Haoyi and Wu, Yang and Mu, Lifu and Wang, Siyuan and Wei, Tao and Hui, Le and Mao, Ning and Zhou, Pan and Zhan, Kun and Yang, Jian},
  title     = {LiAuto-GeoX: Efficient Grounded Driving Transformer},
  journal   = {arXiv},
  year      = {2026},
}