SG²SLAM: geometry-informed 3DGS SLAM beats baselines on KITTI and Waymo
zhenjun_zhao · x · 2026-10-29
SG²SLAM couples streaming spatial geometry with a ray-conditioned 3DGS kernel for monocular outdoor navigation, replacing per-frame 2D-3D association with chunk-level metric depth priors that jointly supervise tracking and mapping.
- Ray-conditioned kernels extend each primitive with a viewing-direction subspace for monocular viewpoints
- Outperforms MonoGS, S3PO-GS, DiskChunGS and GigaSLAM on KITTI rendering (competitors often OOM)
- Best ATE RMSE on several Waymo sequences (e.g. 0.18m vs MonoGS 6.96m)
- Code to be released upon acceptance
More from Embodied
- NVIDIA Inception's Global Head of Physical AI to join SF robotics dinner on Nov 2 — AnneliesGamble · 2026-10-29
- REK 2.1 lands Nov 20: win the sim playoffs to control a real robot — cixliv · 2026-10-29
- $80-150 AI-driven KVMs let VLMs bare-metal control any computer, no install — Porespellar · 2026-10-29
- Physical AI Startup Antioch Raises $32M to Build Autonomy Teams, Hiring Across NYC — IanAndrewsDC · 2026-10-29
- MotionBricks Demo Generates 3D Robot Motion From Target Poses, All on CPU in Browser — victormustar · 2026-10-29
- Apple's Robot Demo Wins Laughs With Its Hilariously Tiny Grab Motion — RileyRalmuto · 2026-10-29