In the past, household robot vacuum cleaners might encounter situations where they suddenly spin in circles in the middle of the living room while cleaning, or crash into the door frame even when the bedroom door is open. Behind all these scenarios lies the same problem: the robot does not know where it is.
Early mobile robots relied on wheeled odometry, counting wheel rotations to calculate how far they had traveled. However, wheel slippage, water on the floor, or crossing carpets could cause the wheels to spin idly. Each step introduces a small error, and as these errors accumulate, the robot becomes completely disoriented.
This is why SLAM technology is now essential.
SLAM stands for Simultaneous Localization and Mapping. It refers to a subject equipped with specific sensors that builds an environmental model and estimates its own motion during movement, without prior knowledge of the environment. If the sensor used is a LiDAR, it is called "LiDAR SLAM"; if the sensor is a camera, it is called "Visual SLAM".
Core Functions of SLAM:
Localization and Mapping: Enabling the robot to know "where I am". Without a map, a robot relying solely on odometry is prone to drift and will lose track of its position after just a few steps.
Navigation and Planning: Enabling the robot to know "how to reach the target point".
Environment Understanding: Enabling the robot to distinguish between static and dynamic objects. The map records the static environment, while dynamic objects are treated as temporary obstacles during operation.
Relocalization: Allowing the robot to recover its position after getting lost.
Principles of LiDAR SLAM
The LiDAR continuously emits laser beams in all directions to measure the distance to walls and furniture. It stitches together concentric rings of contour points to build an environmental map while simultaneously estimating the robot's own position.
LiDAR SLAM relies on LiDAR to directly acquire environmental distance information, featuring high ranging accuracy, insensitivity to lighting changes, and stable and reliable mapping and localization. However, traditional LiDAR suffers from issues such as bulky size and noise generated by rotating motors.
The industry is now shifting towards solid-state LiDAR, which replaces traditional mechanical rotating components with MEMS micro-mirrors or Flash array sensors. With no physical rotation, the LiDAR becomes more compact, boasts a longer lifespan, and operates with lower noise.
Principles of Visual SLAM
The camera continuously captures environmental images and extracts stable feature points from the frames. By tracking the positional changes of the same feature points across two consecutive frames, it calculates the camera's own translation and rotation. Simultaneously, it reconstructs the 3D positions of these feature points to build an environmental map while estimating the robot's own position.
Visual SLAM uses cameras as sensors, offering compact hardware that can simultaneously perform additional functions such as environment recognition, photography, and video recording. However, its localization relies on environmental textures, making it prone to drift in low-light conditions or textureless environments like plain white walls.
VIO (Visual-Inertial Odometry) is currently the most mainstream solution for the engineering deployment of Visual SLAM. In addition to the camera, it incorporates an IMU to output acceleration and angular velocity. The camera has a low frame rate but provides clear vision; however, it tends to fail due to motion blur during rapid movements. The IMU has a high sampling rate and reacts quickly, but it drifts over long periods of integration. Tightly coupling the camera and IMU creates a perfect complement: short-term rapid movements are tracked by the IMU's dead reckoning, while long-term drift is corrected by the camera. VIO is the mainstream solution for facial effects when scrolling through videos, smartphone ranging functions, autonomous drone flight, and robot spatial localization.
Given that both LiDAR SLAM and Visual SLAM have their respective pros and cons, deployed robots now rarely rely on a single sensor. Instead, multi-sensor fusion SLAM has become the standard. It uses Visual SLAM as the primary force for global localization and scene perception, while pairing it with low-cost TOF sensors or solid-state LiDAR for local obstacle avoidance. This approach integrates the advantages of both, compensating for the defects inherent in the single-sensor methods mentioned above.
Author: Xiao Xinji