Sumário
With contributions by Brenno Caudato
SLAM, short for Simultaneous Localization and Mapping, is a fundamental technique in robotics and computer vision. Its goal is to enable robots, autonomous vehicles or devices equipped with sensors to create a map of an unknown environment while simultaneously determining their own position within that environment.
The big challenge for SLAM is that localization and mapping are interdependent tasks: to locate itself, the robot needs a map; but to build a map, it needs to know where it is. This dilemma is known as the “chicken-and-egg problem” in the context of mobile robotics. SLAM solves this dilemma by combining both tasks into a single process, allowing them to be performed simultaneously and incrementally as the robot explores its environment.
The importance of SLAM goes beyond navigation. It is essential for domestic robots, drones, autonomous vehicles and even augmented reality systems to operate in dynamic and unknown environments, without relying on pre-existing maps or external infrastructure such as GPS.
Benefits of SLAM
SLAM greatly expands the possibilities for using robots and autonomous devices. Thanks to its ability to create maps in real time, the system can be deployed in new, complex or constantly changing locations, such as factories, hospitals, homes, and even outdoor environments.
In addition, it allows for rapid reaction to changes, detection of unexpected obstacles and dynamic recalculation of routes, making navigation much safer and more efficient. In industrial environments, for example, SLAM allows automated vehicles to avoid people or objects that are out of place, without compromising productivity.
Additionally, SLAM provides greater localization accuracy, even in environments where GPS signals are weak or non-existent, such as tunnels, underground parking lots or indoors. This is essential for applications such as inspection robots, mapping drones and augmented reality systems, where accuracy is critical to the success of the task.
Finally, SLAM contributes to cost reduction and scalability of autonomous solutions. By eliminating the need for dedicated infrastructure and allowing automatic adjustment to the environment, the technology makes the implementation of robots and intelligent systems in different sectors more accessible, democratizing access to advanced automation.
Common Use Cases for SLAM
One of the most common uses of SLAM is in mobile robots, such as smart vacuum cleaners and industrial robots, which need to accurately locate themselves while mapping their surroundings to avoid obstacles and plan efficient routes. Autonomous vehicles, such as automobiles and drones, also rely on SLAM to navigate dynamic environments, identify obstacles, and continually update the map as they move.
In augmented reality (AR) and virtual reality (VR), SLAM is essential for accurately superimposing digital objects onto the real world. Devices like smartphones and AR headsets (such as Apple Vision) use visual SLAM to identify surfaces, measure distances, and ensure that virtual objects remain stable as the user moves through the environment.
In addition, SLAM is widely used in industrial inspection, mapping of indoor and outdoor environments, agriculture, and even in medical applications, such as navigation of surgical instruments in minimally invasive procedures. In industrial environments, SLAM allows automated guided vehicles (AGVs) to navigate factories and warehouses, optimizing the logistics flow without the need for fixed tracks or beacons.
Another important application is in the creation of detailed 3D maps of environments, whether for digital reconstruction, topographic surveying or construction planning. The combined use of different sensors, such as cameras, LiDAR and IMUs (Inertial Measurement Units, electronic devices that measure and report the specific force, angular rate and, sometimes, orientation of a body, using a combination of accelerometers and gyroscopes) further expands the possibilities of SLAM, making it a versatile and indispensable technology for automation and machine intelligence in various sectors.
Structure of SLAM Algorithms: Frontend and Backend
SLAM algorithms consist of two main components: the frontend and the backend. Each performs distinct but complementary functions in processing sensor data and estimating the position and map of the environment.
Frontend: Feature Acquisition and Extraction
The frontend is responsible for receiving raw data from sensors, processing it, and extracting relevant information (features) from the environment. Data acquisition can be done with several types of sensors, including cameras (monocular, stereo, or RGB-D), LiDAR, radars, ultrasound, and IMUs.
In the feature extraction stage, the algorithm identifies points, lines, planes, or other characteristic elements in the environment, which are then associated (during the data association stage) with other elements found in previous frames of the image, ensuring tracking continuity.
Finally, the odometry estimate is made, where the algorithm calculates the relative movement of the sensor between consecutive frames, serving as a basis for mapping and localization.
Popular Approaches in the Frontend
There are multiple approaches to frontend data capture, each with its own characteristics and use cases. Some of the most popular are:
- Visual-Inertial Odometry (VIO): algorithms that combine camera data with IMU information to estimate movement in six degrees of freedom (6DoF). This approach is widely used in mobile phones for AR/VR applications (e.g., in Apple’s ARKit framework), in augmented reality glasses, and in drone navigation, offering accuracy and robustness even in environments where GPS does not work.
- RGB-D SLAM: Uses cameras that capture color images (RGB) and depth information (D), allowing the extraction of three-dimensional information from the environment, facilitating navigation and mapping in indoor environments.
- LiDAR Odometry: Relies on LiDAR sensors to obtain 3D point clouds of the environment. It can be combined with IMUs for greater accuracy. It is widely used in autonomous vehicles (such as Waymo’s autonomous taxis) and in industrial applications, such as inventory volumetrics.
Backend: Optimization and State Estimation
The backend receives the observations processed by the frontend and performs global optimization of the map and sensor trajectory. It uses mathematical techniques to minimize accumulated errors, correct drifts, and ensure map and location consistency.
The main functions of the backend include graph optimization, when the estimated positions of the sensor and visual landmarks are adjusted to minimize the overall error, and error minimization, which reduces uncertainty in the estimates. The backend also performs loop closure detection, when the system identifies a return to a previously mapped location, correcting accumulated deviations and improving map accuracy.
Algorithms vs Systems, What’s the Difference?
When discussing SLAM technology, it is common to come across terms such as algorithms and systems. The difference between them lies mainly in the scope and complexity of each concept.
A SLAM algorithm is basically the set of mathematical and computational methods that solve the central problem of SLAM: how a robot can simultaneously map an unknown environment and discover its own position within it. It is responsible for processing sensor data, identifying reference points (features), estimating trajectories and correcting accumulated errors, all in a mathematical and abstract way.
A SLAM system is the practical and complete implementation of this concept, bringing together not only the algorithm, but also all the components necessary for SLAM to work in a robot, drone, autonomous vehicle or other device. A SLAM system integrates sensors (such as cameras, LiDAR, IMUs), software modules for data acquisition and synchronization, processing steps (frontend and backend), communication interfaces and, of course, one or more SLAM algorithms at its core.
SLAM systems are designed to operate in real time, handle different types of sensors, perform calibration and deliver ready-to-use results such as maps, trajectories and precise location.
The 5 Most Popular SLAM Systems
There are several SLAM systems in use, each with different strengths, limitations, and recommendations regarding the sensors used. The choice depends on the proposed challenge and the application environment. Some of the most popular are:
ORB-SLAM3
Widely recognized for its accuracy and versatility, supporting everything from monocular to stereo cameras, RGB-D and integration with IMUs, which makes it ideal for visual-inertial (VIO) applications. It excels in scenarios such as mobile robotics, drones and AR/VR applications, thanks to its ability to operate in indoor and outdoor environments, small or large. However, it may struggle in environments with low texture or lighting, in addition to requiring greater computational power for large maps.
RTAB-MAP
Widely used when seeking flexibility in sensor selection. It can work with stereo, RGB-D and LiDAR cameras, as well as integrate visual odometry or LiDAR data. Its main advantage is the ability to create and manage large maps in real time, and is widely used in inspection robots, indoor mapping and industrial applications. On the other hand, it may require detailed parameter adjustments and consume more resources in very large environments.
Cartographer
Developed by Google, it is a reference for applications that prioritize the use of LiDAR, both 2D and 3D, with efficient integration of IMUs. It is very robust for real-time mapping, especially in mobile robots and autonomous vehicles. However, its flexibility for purely visual sensors is limited and, occasionally, loop closure detection can cause jumps in pose estimation.
LIO-SLAM
One of the best options for those who need maximum precision in complex and dynamic environments. It combines 3D LiDAR and IMU data, and can also incorporate GPS to eliminate drift over long periods. Widely used in autonomous vehicles, drones and industrial robots, LIO-SLAM requires careful calibration of the sensors and can be sensitive to vibrations or improper mounting.
LOAM
It is considered a reference in LiDAR-based SLAM, being efficient and accurate even when the sensor is in motion or in dynamic environments. It is widely adopted in autonomous vehicles and for 3D mapping of complex environments, but it is less suitable for scenarios with few objects or flat surfaces, in addition to not offering native support for visual sensors.
Systems Comparison
To make it easier to compare SLAM systems, it is helpful to view their features, typical applications, strengths, and limitations side by side. Here is a summary table with this information:
This comparison demonstrates how each system meets different needs, from applications in controlled indoor environments to large outdoor areas, including systems that require high precision, sensor flexibility or real-time performance. Thus, the choice of the most appropriate algorithm depends on both the project demands and the characteristics of the environment and the available sensors.
Conclusion
SLAM has transformed how machines and autonomous devices interact with the world around them. By enabling robots, vehicles, and even smartphones to build maps and locate themselves in real time, this technology has paved the way for innovative applications in areas such as robotics, autonomous vehicles, augmented reality, industrial inspection, agriculture, and even medicine.
The evolution of algorithms and systems has contributed to making SLAM increasingly accurate, efficient and accessible. Today, it is possible to find SLAM embedded in compact devices, operating in real time and integrating different types of sensors, such as cameras, LiDAR, sonar and IMUs.
By offering autonomy to systems that need to operate in unknown environments, without depending on ready-made maps or external infrastructure, SLAM guarantees flexibility and reduces costs, making automation viable in sectors and situations that previously seemed impossible.
In short, SLAM is one of the foundations of modern robotics and will continue to be a key part of the evolution of autonomous systems, connecting the physical world to the digital world in an increasingly natural and efficient way.
