KD-SLAM: One KD-Tree to deskew them all

PCA KD-trees are a remarkably versatile structure: equivariant under rigid transformations, they encode geometry at every level from root to leaf. Can a single such structure carry the full LiDAR SLAM pipeline – deskewing, odometry, place recognition, and multi-view registration – without switching representation? This paper is the answer: a composable 2D/3D, CPU/CUDA, rigid and continuous-time mapping suite that runs in real time on a 10 W embedded processor. Evaluated on VBR and KITTI. The code is available at https://github.com/rvp-group/kd_slam2. A video is here https://youtu.be/c-sCCt9hMmI