MASt3R-Nav: Pixel-level 3D mapping eliminates need for global geometry

Traditional robot navigation relies on globally-consistent 3D maps that are computationally expensive and brittle in dynamic environments. On the other hand, topological graphs using images or objects sacrifice geometric accuracy, limiting them to simple teach-and-repeat tasks. MASt3R-Nav, presented at ICRA 2026 by researchers including Vansh Garg, Rohit Jayanti, and others, strikes a balance by building a map from image sequences using pixel correspondences in relative 3D coordinate systems of individual image pairs. This approach avoids the need for global geometric consistency while maintaining precise spatial understanding.

The system constructs a pixel-level graph from inter-image connectivity and then sparsifies intra-image connectivity to enable global path planning. This yields a 'WayPixel Costmap' – a dense, pixel-relative representation that serves as a conditioning variable for a learned controller to predict trajectory rollouts. The authors demonstrate that this costmap is a more accurate conditioning signal than image- or object-level alternatives. MASt3R-Nav was validated on four distinct navigation tasks in simulation and through real-world demonstrations, achieving superior performance. The work opens the door to more efficient and robust visual navigation for field robots and autonomous systems.