Simultaneous Localization and Mapping (SLAM)

01Definition: a chicken-and-egg problem

Why mapping and localization must be solved together

Robotics researchers Hugh Durrant-Whyte and Tim Bailey framed the core question in their foundational 2006 tutorial: can a mobile robot placed at an unknown location in an unknown environment incrementally build a consistent map of that environment while simultaneously determining its own location within it? They called solving this a "holy grail" for making a robot truly autonomous.

The reason it can't be split into two separate steps is often described as a chicken-and-egg problem: building an accurate map requires knowing where the robot is at each moment, and figuring out where the robot is requires already having a map to compare sensor readings against. Durrant-Whyte and Bailey's own explanation goes further — every landmark a robot observes is measured relative to its own uncertain position, so the resulting map estimates are all statistically correlated with that same position error.

Map and position have to be estimated jointly — neither can be solved first.

From odometry drift to the birth of SLAM

SLAM exists because the alternative, dead reckoning, breaks down quickly. A 1997 paper on odometry error modeling put it plainly: because of wheel slippage, floor roughness, and small errors from discretized wheel-increment sampling, odometry accumulates position error "in an unbounded fashion" the longer a robot moves.

The field traces back to the 1986 IEEE Robotics and Automation Conference in San Francisco, where researchers including Peter Cheeseman, Jim Crowley, and Hugh Durrant-Whyte began applying statistical estimation methods to mapping and localization. The acronym "SLAM" itself wasn't coined until nearly a decade later, in a 1995 paper on localizing automatic guided vehicles.

How this differs from GPS

Satellite positioning (GPS) is another way a robot can find its own position.

02Three algorithm families, and loop closure

Where GPS can't reach and no fixed infrastructure such as floor markers is in place, a robot has to solve this joint estimation problem with only its own onboard sensors. SLAM algorithms fall into three broad families: EKF-SLAM, FastSLAM, and graph-based SLAM.

How the processing pipeline works

Graph-based SLAM is the dominant one of the three today — here's how its processing pipeline works, from raw sensor data to an updated map.

At the core of graph-based SLAM is loop closure: detecting from sensor data that the robot has returned to a place it visited before, then using that match to correct accumulated drift across the entire graph at once.

That correction is what keeps a robot's map from drifting further out of shape the longer it explores.

Loop closure doesn't always work correctly. When different places look alike — perceptual aliasing — the system can mistake one location for another and insert a wrong constraint, making the map worse instead of better. A 2026 benchmark paper (a preprint, not yet peer-reviewed) tested ORB-SLAM3 across kilometer-scale synthetic outdoor sequences and found that in a repetitive desert environment, enabling loop closure raised one run's position error from 4.12 meters to 37.85 meters because the system accepted a spurious match.

From ORB-SLAM to foundation-model SLAM

The graph-based family has kept evolving well past its 1990s origins — absorbing photorealistic 3D-rendering techniques from computer-vision research, and, most recently, building on general-purpose 3D-reconstruction foundation models instead of a SLAM-specific network trained from scratch.

03LiDAR SLAM vs. Visual SLAM

Algorithm families are one axis; SLAM also splits by which sensor does the sensing.

LiDAR sensor cost and the competition to supply robotics customers are covered in more depth in our sensors & perception market article.

Combining both: hybrid systems

Some systems combine both sensing approaches. An RGB-D camera pairs an ordinary color camera with a depth sensor in the same unit, giving Visual SLAM's lower hardware cost some of LiDAR's ability to measure geometry directly, without a separate laser scanner.

KinectFusion, an early system built on Microsoft's Kinect depth camera, showed in 2011 that a full real-time 3D surface reconstruction was possible from exactly this kind of low-cost combined sensor.

04From robot vacuums to warehouse robots

Consumer robots: laser and camera mapping

The scale and cost of a SLAM implementation vary enormously with what it's built into — and the laser-versus-camera split from section 03 shows up directly in consumer products.

Among home robots, Neato Robotics' XV-11 (which began shipping in June 2010) is an early example of a consumer robot mapping a room with a rotating laser sensor, and iRobot's Roomba 980 (released September 2015) was the first Roomba built on iRobot's own camera-based SLAM, which the company calls vSLAM.

Configurable and open-source SLAM stacks

Some industrial and research robots let a user choose the SLAM approach for the job. AgileX Robotics' official LIMO manual documents two LiDAR SLAM workflows: Gmapping, described for smaller spaces, and Google's graph-optimization-based Cartographer, which performs real-time loop closure.

Nav2, the standard navigation stack of the ROS 2 robot framework, includes SLAM Toolbox (created by Steve Macenski) as its default SLAM implementation. Nav2's official SLAM tutorial demonstrates the stack on TurtleBot3, a common research platform.

Fixed infrastructure as an alternative to SLAM

Not every autonomous mobile robot (AMR) actually runs SLAM.

Amazon says Hercules, one of its warehouse drive units, reads a grid of encoded floor markers for localization, while Pegasus operates in the same kind of structured field. These systems show that an AMR (autonomous mobile robot) can navigate autonomously using fixed infrastructure; SLAM is one alternative when a robot must localize in spaces that are not prepared with those markers.

Mobile-robot market growth

ABI Research identifies SLAM as one of the technologies enabling mobile robots to expand beyond their traditional stronghold in warehousing and logistics into more varied, less structured environments. The firm's December 2024 forecast shows some of the fastest growth in exactly the newer segments where that flexibility matters most: agricultural robot shipments are projected to grow from 7,000 units in 2023 to 129,000 by 2030 (a 45% CAGR), delivery robots from 14,000 to 147,000 (37%), and restaurant robots from 6,000 to 78,000 (43%). As ABI Research robotics analyst George Chowdhury put it, mobile robots are "expanding automation to previously under-automated sectors" — deployments that depend on a robot being able to localize itself in a space nobody mapped out for it in advance.

05Open technical challenges

As those deployments spread, SLAM's remaining limits matter more. The perceptual-aliasing failures covered in section 02 aren't the only open problem — two areas in particular still limit what SLAM can reliably do today.

One active response to the semantic and generalization limits above is already showing up in section 02's algorithm timeline: foundation-model SLAM, which reuses a general-purpose 3D-reconstruction model instead of training a SLAM-specific network from scratch.

06Easier ways to put SLAM to work

Those open problems are for researchers to chip away at — most teams just need SLAM that works today. Building a SLAM stack from scratch, or configuring an open-source one like the ROS options in section 04, isn't the only path anymore. Three kinds of ready-made options now handle the SLAM itself so developers don't have to.

07Sources