humanoidsdata.com

Search

Search companies, datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Hardware & control

SLAM

Simultaneous localisation and mapping

SLAM is the joint estimation of a moving robot or sensor rig's pose and a map of its environment from sensor observations. It addresses the coupled problem of needing a map to localise while needing pose estimates to build that map.

Also known as: simultaneous localization and mapping

Updated

Estimating pose and a map together

In SLAM, new sensor observations help estimate both where the sensor is and what its surroundings are like. The map is not just a visualisation attached to a trajectory: it provides a reference that can be used again for localisation. The ORB-SLAM3 paper makes this map-reuse objective explicit in its discussion of visual SLAM and odometry.

SLAM is a problem class, not the name of one sensor or algorithm. ORB-SLAM3 illustrates several camera-based configurations, including monocular, stereo, RGB-D and visual-inertial operation. Its particular representations and sensor support should not be treated as mandatory features of every SLAM system.

Revisited places can constrain accumulated drift

A loop closure associates a current observation with an earlier mapped place. That relationship can support correction of accumulated trajectory and map error. ORB-SLAM3 describes place recognition, geometric consistency checks and subsequent optimisation, as well as merging maps from different sessions.

Revisiting a place is not an automatic accuracy guarantee. The system must establish the correct correspondence; weak visual information or an incorrect match can undermine estimation. ORB-SLAM3 identifies low-texture environments as a failure case for its feature-based approach. Its benchmark results do not establish the accuracy of an unrelated robot, camera or environment.

Visual-inertial odometry focuses on motion estimation and may discard older scene information. Some odometry systems add loop closing or mapping, so the practical distinction depends on what history they retain and reuse, not merely the software's label.

A SLAM map is not automatically a task-ready digital scene

A map sufficient for localisation may not contain the dense geometry, object identities, physical properties or contact surfaces needed for simulation and manipulation. Buyers of humanoid or egocentric datasets should ask what the exported map actually contains, rather than assuming that a file called a map is a complete reconstruction.

For trajectory use, record the coordinate frame, scale convention, calibration, mapping version and whether estimates were produced online or revised after loop closure. Preserve reset boundaries and map identifiers when tracking restarts. A retrospectively optimised trajectory can be useful supervision, but it must not be confused with the pose estimate a deployed controller could access at that instant.

Sources