FullSLAM

Fullslam, often written full-SLAM or full Slam, is a term in robotics and computer vision referring to a class of SLAM approaches that aim to estimate a complete, globally consistent map of the environment together with the full trajectory of the sensing platform. In a full-SLAM system, all available observations are incorporated into a single probabilistic model, producing a joint posterior distribution over robot pose and map features rather than separate, incremental estimates.

Approaches to full-SLAM include graph-based methods, where poses and landmarks are nodes and constraints from odometry and observations are edges; optimization techniques such as pose graph optimization, bundle adjustment, or dense graph methods are used to enforce consistency and reduce drift. Sensor modalities commonly used include stereo, monocular vision, LiDAR, and radar, often fused in a unified estimation framework. Full-SLAM contrasts with partial or local SLAM variants that may relinearize or reset portions of the map, leading to drift or inconsistency.

Applications of full-SLAM span autonomous vehicles, drones, service and industrial robots, and augmented reality. Challenges include computational cost, data association, loop-closure detection, nonlinearity, and the need to store and manage large maps. Variants of full-SLAM encompass visual-SLAM, LiDAR-SLAM, and multi-sensor configurations; some systems perform online graph optimization or batch optimization. The term is sometimes used informally to describe any SLAM implementation that emphasizes global consistency and comprehensive mapping.