Mapping and localisation

Geometric mapping and localisation: SLAM, 3D reconstruction from images, and the estimation underneath both. Feed-forward reconstruction models (VGGT, MapAnything, MASt3R-SLAM) are changing what one camera can do, and certifiable estimators (TEASER++, GNC) are changing what a robot can promise about its answer.

Concepts

• Robust estimation: outlier rejection, certificates, and what adversarial outliers do to them.
• Collaborative SLAM, filed under multi-robot systems.
• Underwater perception, filed under field robotics.

Entities

Datasets, platforms, tools and groups that come up in this subfield: COLMAP, lidar, ScanNet, Replica, uHumans2. The entity index has the rest.

Open questions

• Can a certificate travel with a measurement, so the robot receiving it knows how much to trust it?
• Dense learned depth is often wrong with confidence. Where should a SLAM system stop trusting it?
• Which faults can no residual reveal, because the robot never moves in the direction that would show them?

Papers

Other subfields: perception, semantic mapping, planning and navigation, control, robot learning, multi-robot systems, safety and verification, security, Human-robot interaction, field robotics.