LiAuto-GeoX: Efficient Grounded Driving Transformer
Researchers have developed LiAuto-GeoX, an efficient transformer model for real-time 3D scene understanding in autonomous driving. This model leverages sparse LiDAR data for geometric grounding and employs a novel distillation framework to create a compact 155M-parameter onboard version. LiAuto-GeoX achieves high-fidelity reconstruction at 220 FPS on the KITTI dataset and shows strong performance in downstream tasks like trajectory and occupancy prediction. AI
IMPACT Enables real-time, onboard 3D scene understanding for autonomous vehicles, potentially improving safety and efficiency.