We propose contribution-based trimming strategy to refine messy Gaussians to be geometrically accurate. This strategy has a potential to be integrated with any Gaussians!

We propose a practical LAtent World Model (LAW) to enhance end-to-end autonomous driving in a self-supervised manner without leveraging time-consuming video generation, achieving strong and realtime performance in both open-loop and close-loop benchamrks.

The first voxel-based state space model for LiDAR-based 3D object detection in driving scenes, achiving state-of-the-art performance in both Waymo and nuScene benchmarks.

Drive-WM is the first multi-view world model for planning in autonomous driving.

MixSup achieves strong performance with a few accurate box labels and cheap cluster labels. PointSAM is developed to generate cluster labels. It achieve on-par performance with SoTA 3D segmentation methods in nuScenes without any 3D annotations!

FSDv2 is an improved version of FSD, removing the handcrafted heuristics in FSD. FSDv2 achieves strong performance in Waymo, nuScenes, and Argoverse 2 dataset, and are fully open-sourced!

CTRL is the first open-sourced LiDAR-based 3D object autolabeling system, surpassing the performance of human annotators!

FSF explores multi-modal 3D object detection with fully sparse architecture by seamlessly integrating 2D instance segmentation and 3D instance segmentaion in a unified framework.

FSD++ extends FSD into the multi-frame setting. In addition to the spatial sparsity, FSD++ emphaiszes temporal sparsity.

FSD first proposes the concept of LiDAR-based "fully sparse detection", achieving state-of-the-art performance in both conventional benchmark and long-range (>200m) LiDAR detection benchmark.

SST emphasize the small object sizes and sparsity of point clouds. Its sparse transformers enlight new backbones for outdoor LiDAR-based detection.

RangeDet greatly narrows the performance gap between range view based LiDAR detection and voxel/BEV based LiDAR detection.