SLAM (Simultaneous Localization and Mapping)

Definitional Boundaries & Taxonomy

Simultaneous Localization and Mapping (SLAM) is the foundational computational problem of constructing or updating a map of an unknown environment while simultaneously keeping track of an agent's location within it. In essence, it is the classic "chicken-and-egg" problem: to localize precisely, the robot requires a map, but to construct a coherent map, the robot must know its precise location. At an architectural level, SLAM is traditionally bifurcated into two primary subsystems: the Front-End (sensor data processing, feature extraction, data association, and outlier rejection) and the Back-End (state estimation, map optimization, error minimization, and loop closure computation).

Taxonomically, SLAM systems are classified primarily by their perceptive modalities and the underlying filtering paradigms:

• Visual SLAM (vSLAM): Relies on monocular, stereo, or RGB-D cameras (e.g., ORB-SLAM3, SVO, LSD-SLAM). Monocular setups suffer from scale ambiguity, requiring IMU integration.
• LiDAR SLAM: Utilizes 2D or 3D Laser Scanners, ideal for structural geometries and lighting-invariant operations (e.g., LOAM, Cartographer, FAST-LIO, LeGO-LOAM).
• Inertial/Odometry Fusion (VIO/LIO): Visual-Inertial Odometry and LiDAR-Inertial Odometry, coupling high-frequency IMU data with low-frequency perceptive scans to overcome featureless domains and aggressive motion blur.
• Semantic SLAM: Incorporates object recognition to map not just geometry, but the meaning of the space, moving from dense point clouds to object-oriented topological and hierarchical maps.
• Radar SLAM: Utilizing mmWave radar for extreme weather robustness (fog, dust, rain), albeit with high noise profiles and sparse returns.