Simultaneous localization and mapping (SLAM) is the problem of building a map of an unknown environment while at the same time tracking the position of the robot or device within that map. The two tasks depend on each other. To place itself, a robot needs a map. To build a map, it needs to know where it is. SLAM solves both together using sensor data such as camera images, laser scans and motion measurements. It is a foundation of autonomous robots, self-driving cars, drones, robot vacuum cleaners and augmented reality, and it is a central topic in robotic perception.
The problem
Imagine being dropped into an unfamiliar building with a flashlight and a notepad, with no GPS. As you walk, you sketch a floor plan and mark where you are. Your estimate of your own position drifts with every step, and so the sketch gets distorted. When you return to a place you have seen before, you can recognize it and correct the whole drawing. SLAM faces exactly this challenge.
A SLAM system continuously estimates two things.
- The trajectory. The sequence of the robot’s poses, meaning positions and orientations, over time.
- The map. A representation of the surroundings, such as landmarks, a grid of occupied and free cells, a dense 3D model or a collection of features.
Because sensors and motion estimates are noisy, both are uncertain, and errors in one contaminate the other. Early researchers, including Smith, Self and Cheeseman in the 1980s, showed that the errors in robot pose and map features are correlated and must be estimated together instead of separately [1]. The acronym SLAM was coined in the 1990s, and Durrant-Whyte and Bailey’s two-part tutorial in 2006 summarized the first two decades of work [2].
How SLAM works
A SLAM system usually has a front end and a back end.
- Front end. Processes sensor data into measurements. It extracts features or scan matches, estimates motion between successive frames (odometry) and finds data association, deciding whether an observation corresponds to something already mapped.
- Back end. Takes these measurements and computes the best estimate of the trajectory and map, taking account of uncertainty.
- Loop closure. Detects when the robot has returned to a previously visited place, and adds a constraint that corrects accumulated drift across the whole map.
- Map management. Stores, updates and prunes the map, and may relocalize the robot if it gets lost.
Main estimation approaches
### Filtering methods
Early SLAM used the extended Kalman filter, which keeps a joint estimate of the robot pose and landmark positions with a covariance matrix. It works for small maps but scales poorly, because the cost grows with the number of landmarks. Particle filter methods, such as FastSLAM, represent the robot’s trajectory with samples and landmarks with small filters, and are widely used for grid maps.
### Graph-based optimization
Modern systems usually represent SLAM as a graph. Nodes are robot poses, and edges are constraints from odometry and observations. The back end finds the arrangement of poses that best satisfies all constraints, using nonlinear least squares optimization. Libraries such as g2o, GTSAM and Ceres support this. Graph methods scale to large environments and make it easy to add loop closures [3]. See also sensor fusion, since SLAM often fuses several sensors.
### Learning-based SLAM
Neural networks now assist or replace some parts: learned features and descriptors for matching and place recognition, depth prediction, and end-to-end odometry. Neural scene representations, such as neural radiance fields and Gaussian splatting, have been combined with SLAM to produce detailed maps; see 3D reconstruction. Learning improves robustness in difficult scenes but needs careful testing for generalization.
Types of SLAM by sensor
- Visual SLAM. Uses one or more cameras. Monocular systems estimate motion and structure up to an unknown scale, while stereo and RGB-D cameras resolve scale. Well-known open-source systems include ORB-SLAM, which tracks image features across frames.
- Visual-inertial SLAM. Adds an inertial measurement unit, which improves robustness to fast motion and gives metric scale. It is standard on drones and augmented reality headsets.
- Lidar SLAM. Matches laser scans to build precise geometric maps, common in autonomous vehicles and warehouses. Lidar is less affected by lighting but heavier and more costly.
- Radar and sonar SLAM. Used where light-based sensors fail, such as in smoke, dust or underwater.
- Multi-sensor and multi-robot SLAM. Several sensors, or several robots sharing a map, improve coverage and reliability.
Map representations
- Landmark maps. Sparse sets of distinctive points or objects with positions.
- Occupancy grids. A grid in which each cell stores the probability of being occupied, widely used for navigation.
- Point clouds and meshes. Dense 3D geometry for detailed models.
- Topological maps. Graphs of places and connections, helpful for route planning.
- Semantic maps. Maps annotated with object and region labels, which link SLAM to scene understanding and object detection.
The map is often used by motion planning to compute safe paths.
Applications
- Mobile robots. Warehouse vehicles, cleaning robots and delivery robots navigate buildings without markers.
- Autonomous driving. Building and using high-definition maps and localizing within them.
- Drones. Flying indoors, in tunnels and forests where satellite navigation fails.
- Augmented and virtual reality. Headsets and phones track their position to anchor virtual content.
- Surveying and construction. Handheld and vehicle-mounted scanners create floor plans and 3D models.
- Underwater and space exploration. Mapping with no GPS.
- Mining, inspection and search and rescue. Mapping hazardous sites.
- Medical endoscopy. Mapping organs from camera video.
Challenges
- Drift and loop closure errors. Small errors accumulate, and a false loop closure can ruin a map.
- Dynamic environments. Moving people and vehicles break the assumption of a static world.
- Perceptual aliasing. Different places that look alike, such as identical corridors, confuse recognition.
- Lighting and weather. Cameras fail in darkness or glare, and lidar in rain or fog.
- Scale and long-term operation. Maps must be kept consistent over large areas and over months, as environments change.
- Computation. Real-time SLAM on small, low-power devices demands efficiency.
- Robustness and failure detection. Systems must know when they are lost and recover. Cadena and colleagues argued that SLAM still needs improvements in robustness, scalability and high-level understanding before it can be called solved for all uses [3].
- Privacy. Maps of homes and workplaces can reveal sensitive information.
FAQs
What is SLAM? It is the process by which a robot or device builds a map of an unknown place while simultaneously estimating its own position within it.
What is loop closure in SLAM? It is the recognition that the robot has returned to a previously visited location, which allows the system to correct accumulated drift.
What is visual SLAM? It is SLAM that uses cameras as the main sensor, extracting features from images to estimate motion and build a map.
What is the difference between SLAM and GPS? GPS gives position from satellites and works poorly indoors. SLAM works from onboard sensors, needs no external infrastructure and also builds a map.
Where is SLAM used? In robot vacuums, warehouse robots, drones, self-driving cars, augmented reality devices and surveying tools.
More in Robotics and Autonomous Technology
Key facts
- The foundations of SLAM were laid by Smith, Self and Cheeseman in the 1980s, who showed that robot pose and map features are correlated and should be estimated together1.
- Durrant-Whyte and Bailey's 2006 two-part tutorial gave a widely cited overview of the history and main methods of SLAM2.
- Cadena and colleagues' 2016 survey described how SLAM research moved from filtering methods to graph-based optimization and questioned whether the problem was solved for long-term, large-scale use3.
Related concepts
Quick checkWhat is SLAM?Show answer
SLAM, short for simultaneous localization and mapping, is the computational problem of building a map of an unknown environment while at the same time estimating where the robot or device is within that map, using sensor data such as cameras or lidar.
Sources
- Randall Smith, Matthew Self and Peter Cheeseman, Estimating Uncertain Spatial Relationships in Robotics, Autonomous Robot Vehicles, https://doi.org/10.1007/978-1-4613-8997-2_14
- Hugh Durrant-Whyte and Tim Bailey, Simultaneous localization and mapping: part I, IEEE Robotics and Automation Magazine, https://doi.org/10.1109/MRA.2006.1638022
- Cesar Cadena et al., Past, Present, and Future of Simultaneous Localization and Mapping: Toward the Robust-Perception Age, IEEE Transactions on Robotics, https://arxiv.org/abs/1606.05830
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
