Co-SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM

Wang et al. introduce Co-SLAM, a real-time neural RGB-D SLAM system designed around a hybrid scene encoding that combines a multi-resolution hash grid with a coordinate-based one-blob encoding. The hash grid contributes fast convergence and high-frequency detail, while the smooth coordinate encoding encourages surface coherence in regions that the camera has not yet observed — a known weakness of purely parametric grids, which tend to leave unobserved space noisy.

The hybrid representation feeds a shallow MLP that predicts signed distance and colour for queried 3D points. Camera tracking is performed by inverting this rendering through a photometric and depth loss; the system then folds tracking and mapping into a global BundleAdjustment over all keyframes rather than the small active-keyframe window used in iMAP and NICE-SLAM. A ray sampling strategy tuned to the hybrid representation keeps the bundle adjustment tractable at interactive rates.

On Replica, ScanNet, and TUM-RGBD the system runs at 10–17 Hz and matches or exceeds prior neural SLAM methods in reconstruction quality (PSNR, SSIM, LPIPS) and tracking accuracy. The global BundleAdjustment is shown to be essential in long sequences where local-window optimisation alone allows drift to accumulate, while the smoothness contribution of the one-blob encoding keeps unobserved regions plausible rather than empty. Co-SLAM is part of a wave of neural SLAM systems that culminated in GaussianSplatting-based variants such as GS-SLAM and Gaussian Splatting SLAM, but it remains a strong baseline for hybrid implicit representations and for studying how surface-completion priors interact with high-frequency parametric encodings in incremental reconstruction.