LiDAR

LiDAR (Light Detection and Ranging) is an active exteroceptive sensor: it emits laser pulses and times the returns, directly measuring range to surfaces. A scanning LiDAR outputs dense 3D point clouds at typically 10–20 Hz with excellent range accuracy (on the order of ±2 cm), and — because it brings its own illumination — it works in darkness and is largely insensitive to ambient lighting.

What a LiDAR measurement contains

Each fired pulse returns a range along a known beam direction; range plus direction gives a 3D point in the sensor frame. A full sweep bundles these into a point cloud where each point typically carries:

Two main hardware families:

Motion distortion

A recurring theme unique to scanning sensors: at 10–20 Hz, one sweep takes on the order of 50–100 ms to acquire, and the platform moves meaningfully in that time. Points from the start and end of the sweep are therefore expressed in different sensor poses, and treating the scan as rigid smears the geometry. Every serious LiDAR pipeline de-skews: each point is re-projected using an interpolated pose for its individual timestamp, with the motion coming from an IMU or a constant-velocity model. This is one of the main reasons LiDAR pairs so naturally with an IMU.

Camera vs. LiDAR

Compared with cameras, the trade-offs are sharp:

CameraLiDAR
DepthIndirect (triangulation)Direct, metric
LightingFails in darkness / HDR scenesIndependent of ambient light
DataDense texture, color, semanticsSparse geometry, no color
Failure casesLow texture, blurRain/fog/dust, geometrically degenerate scenes (long corridors, open fields)
Cost/weight/powerLowHigher

LiDAR’s failure modes deserve emphasis because they are geometric, not photometric: a long featureless corridor or tunnel constrains the sensor in some directions and not others (scan matching slides freely along the corridor axis), open fields offer few structures to match, and rain, fog, dust, or glass produce spurious or missing returns. Degeneracy detection and handling is its own topic at Level 9.

The LiDAR SLAM stack

LiDAR SLAM has its own algorithmic stack, developed in Level 9:

Because camera and LiDAR failure modes are complementary — LiDAR gives geometry where vision has no texture; the camera gives texture and semantics where geometry is degenerate — the roadmap’s arrow from this node points to Visual-LiDAR fusion: systems like LVI-SAM, R3LIVE, and FAST-LIVO fuse both (plus IMU) for robustness that neither sensor achieves alone.

Why it matters for SLAM

LiDAR is the sensor of choice when metric accuracy and lighting robustness are non-negotiable — autonomous vehicles, surveying, industrial robots. For a visual-SLAM learner, understanding LiDAR matters twice over: it defines the strongest alternative paradigm to visual SLAM (useful for choosing sensors honestly), and it is half of the visual-LiDAR fusion systems that increasingly represent the state of the art in robust field robotics.