MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors

1 min read

MASt3R-SLAM is a real-time monocular dense SLAM system built around MASt3R, a network that predicts aligned pointmaps and matches for a pair of images. It assumes no fixed or parametric camera model beyond a single camera centre. The system adds pointmap matching, camera tracking with local fusion, graph construction and loop closure, and second-order global optimisation.

Findings

  • Globally consistent poses and dense geometry at 15 FPS (abstract); with known calibration, a small change gives the best published results on several benchmarks.
  • Pow3R-SLAM reproduces its published numbers on 16 sequences and extends it to RGB-D.

Limits

  • Abstract only so far. Monocular, so scale is free (Pow3R-SLAM’s unscaled error comparison makes this concrete).

Relevance

The base system for a family of foundation-model SLAM work, including underwater attempts. Its robustness on in-the-wild video is the claim to test in murky water. See foundation models for robotics and xu-2026-aquaflow.