How LiDAR Works — Laser 3D Mapping

Photo: Sami Aksu / Pexels
LiDAR — Light Detection and Ranging — is a remote sensing technology that fires laser pulses at the environment and measures how long they take to return. Multiply by the speed of light, divide by two, and you have a precise distance. Do this millions of times per second while scanning a beam, and you get a dense 3D point cloud of everything around you.
LiDAR is the sensor that lets a Waymo car see at night, that lets your iPhone Pro autofocus instantly in the dark, and that revealed 60,000+ hidden Maya structures under Guatemalan jungle canopy. It's also the sensor Elon Musk publicly called a "fool's errand." Both can be true.
Time of Flight vs FMCW
There are two main physical principles for LiDAR.
Time of Flight (ToF) — fire a very short laser pulse (nanoseconds long), use a sensitive photodetector to time the return. Distance = (c × Δt) / 2. Simple, well-understood, dominant in shipping LiDARs as of 2026.
FMCW (Frequency-Modulated Continuous Wave) — emit a continuous laser whose frequency is being swept (chirped). The returning light's frequency is shifted slightly due to round-trip delay (and any Doppler from a moving target). Beating the returning light against the local source reveals both distance and velocity simultaneously. FMCW is harder to engineer (needs a stable single-mode laser, coherent detection) but is immune to interference from other LiDARs, can see through some fog, and directly measures velocity — useful for autonomy.
Aeva is the main commercial FMCW LiDAR pushing into automotive; most others (Hesai, Luminar, Innoviz, Cepton) are ToF.
Wavelengths: 905 nm vs 1550 nm
Two near-infrared wavelengths dominate.
| Wavelength | Pros | Cons | Used by |
|---|---|---|---|
| 905 nm | Cheap silicon detectors, mature | Eye-safety limits laser power → shorter range | Most consumer LiDARs, low-cost automotive |
| 1550 nm | Eye-safe at much higher power → longer range, sun-immune | More expensive (InGaAs detectors) | Luminar, Aeva, premium automotive |
The eye-safety story is the key tradeoff. At 905 nm, the lens of the eye focuses light onto the retina, so the maximum permissible laser power is low. At 1550 nm, water in the cornea absorbs the light before it reaches the retina, so you can pump out ~40× more power and still be eye-safe — which translates directly into hundreds of meters more range.
Mechanical, MEMS, and Solid-State Scanning
Early LiDARs (Velodyne HDL-64, used on the DARPA Grand Challenge winners and the iconic Waymo "spinning hat") were mechanical: an entire optical assembly rotated 360° at 5–20 Hz. Robust, but bulky, expensive, and a moving-parts reliability problem.
Modern designs have moved toward fewer or no moving parts:
- MEMS scanning — tiny micromirrors steer the beam. Cheap, compact, but limited field of view per unit.
- Optical phased arrays — same beam-forming principle as Starlink's antennas, but for light. Fully solid-state, electronically steerable. Still maturing.
- Flash LiDAR — illuminate the whole scene at once, detect with a focal-plane array. Simple, no scanning, but range-limited.
- Hybrid spinning + solid-state — Hesai AT128, Luminar Iris, etc.
The iPhone 12 Pro and newer iPad Pros use a small direct-time-of-flight (dToF) flash LiDAR with a VCSEL array and SPAD detector. It maxes out around 5 m, but it's enough for AR, autofocus assist, and 3D room scanning.
Where LiDAR Is Used
| Application | Why LiDAR | Notable examples |
|---|---|---|
| Self-driving cars | Direct 3D, works at night | Waymo, Cruise, Zoox, Mobileye |
| Driver assist | Detect cut-ins and pedestrians at range | Volvo EX90 (Luminar), Mercedes Drive Pilot |
| Smartphones / AR | Depth for AR, autofocus | iPhone Pro, iPad Pro (dToF) |
| Archaeology | See through forest canopy | Caracol, Tikal, Angkor surveys |
| Forestry | Canopy height, biomass | NASA GEDI on ISS |
| Atmospheric science | Cloud height, aerosol profiles | CALIPSO satellite |
| Bathymetry (shallow seas) | Green (532 nm) penetrates water | Coastal surveys |
| Drones / surveying | Sub-cm topography | DJI Zenmuse L2, Phoenix LiDAR |
| Industrial robots | Obstacle avoidance | Warehouse AMRs |
The archaeology story is particularly fun. In 2018, the PACUNAM LiDAR Initiative mapped 2,100 km² of Guatemalan jungle and revealed more than 60,000 previously unknown Maya structures — causeways, fortifications, agricultural terraces, urban sprawl — under dense canopy. LiDAR's ability to filter out vegetation by looking only at "last returns" (pulses that penetrate gaps in the leaves and bounce off the ground) reshaped Mesoamerican archaeology overnight. Similar surveys have transformed our maps of Angkor and Amazonian pre-Columbian civilizations.
LiDAR vs Radar vs Cameras
Self-driving cars stack multiple sensors because each has different failure modes.
| Sensor | Resolution | Range | Bad weather | Native depth | Cost |
|---|---|---|---|---|---|
| LiDAR | High (dense 3D) | 100–300 m | Degraded in heavy fog/rain | Yes | $500–$10,000 |
| Radar (mmWave) | Low–medium | 200+ m | Excellent (sees through fog) | Yes (+velocity) | $50–$500 |
| Cameras (RGB) | Very high | Limited by resolution | Poor in low light | No (stereo or ML needed) | $5–$200 |
| Imaging radar (4D) | Medium | 300+ m | Excellent | Yes (+velocity) | $200–$1,500 |
This is why the "vision-only vs sensor fusion" debate exists. Tesla famously bet on cameras alone, arguing that humans drive with two eyes and a brain, and that adding LiDAR is expensive and redundant. Waymo, Cruise, Zoox, Mobileye, and essentially everyone else disagree and keep LiDAR because:
- Direct 3D measurement avoids hard inference problems.
- Works in pitch darkness.
- Failure modes are different from cameras — useful redundancy.
Whether vision-only can match LiDAR-based stacks remains contested. As of 2026, no robotaxi service operates at scale without LiDAR — see our how self-driving cars work article for the broader picture.
The Major LiDAR Companies
| Company | Headquarters | Headline product | Notable customers |
|---|---|---|---|
| Hesai | China | AT128, AT512 | Largest by volume in 2024 |
| Luminar | USA | Iris (1550 nm) | Volvo EX90 |
| Innoviz | Israel | InnovizTwo | BMW iX |
| Aeva | USA | Aeries II (FMCW) | Daimler Truck |
| Cepton | USA | Vista series | General Motors (canceled 2023) |
| Mobileye | Israel | EyeQ + LiDAR | Multiple OEMs |
| Ouster (merged Velodyne) | USA | Digital flash | Robotics, mapping |
Hesai's emergence as the dominant player by shipped volume reflects China's rapid EV/autonomy buildout — most domestic Chinese EVs ship with at least one LiDAR by 2026.
Signal Processing and Noise Rejection
Raw LiDAR data is noisy. Multiple factors degrade the 3D point cloud:
Ambient light: Sunlight interferes with detection. The detector sees scattered sunlight plus signal echo. Modern detectors use high-speed gating (turn on/off nanoseconds after firing) to reject most ambient light, but sunset/sunrise conditions still challenge 905 nm systems. 1550 nm is naturally sun-immune.
Multiple returns: A laser pulse can scatter multiple times—bouncing off vegetation, then ground, creating ghost detections. Advanced systems distinguish "first return" (closest object), "last return" (often ground), and intermediate returns. Filtering to last return works well for archaeology (vegetation suppression) but fails on reflective surfaces.
Cross-talk: In high-density scanning, signals from adjacent beam elements can interfere, creating false positives. FMCW's coherent approach naturally rejects this; ToF systems need careful filtering.
Motion distortion: If the LiDAR or vehicle is moving, scan lines are displaced relative to each other. Ego-motion compensation (subtracting vehicle velocity/rotation) is essential for clean point clouds.
Range-dependent noise: Noise increases with range — a target 200 meters away has lower SNR (signal-to-noise ratio) than one 50 meters away. This limits effective range for small, dark objects.
The output of signal processing is a labeled point cloud: each point has (x, y, z), intensity (return signal strength), and optionally velocity, classification (road, vegetation, building), and confidence.
Source: Thakur, S. et al. (2020) — "LiDAR Point-Cloud Signal Processing: Methods and Challenges" — IEEE Sensors Journal.
Eye Safety and IEC 60825 Classification
Eye safety is the primary regulatory concern. IEC 60825-1 establishes laser safety classifications and maximum permissible exposure (MPE).
Class 1 — safe under all normal use; no hazard, no precautions needed. Most automotive LiDARs aim for Class 1.
Class 3R — moderate risk; viewing beam is harmful, but viewing reflected light is usually safe. Lower-cost LiDARs might fall here.
Class 3B/4 — higher risk; avoid all viewing.
The eye-safety tradeoff is fundamental:
- 905 nm: The lens focuses this wavelength onto the retina → low power limit (MPE ~1 mW averaged) → shorter range.
- 1550 nm: Water in the cornea absorbs this → no focus onto retina → higher power limit (MPE ~40–50 mW) → longer range and sun-immunity.
Luminar's choice of 1550 nm is largely about this eye-safety advantage: longer range + sun immunity + higher power tolerance. The drawback is cost (InGaAs detectors are more expensive than silicon).
Source: IEC 60825-1:2014 — "Safety of laser products — Part 1: Equipment classification and requirements."
Point Cloud Processing and 3D Segmentation
A 64-channel LiDAR at 10 Hz produces ~650,000 points per second. What do you do with them?
Segmentation: Label points (road, car, pedestrian, building, vegetation). Classical methods use features (intensity, local density); modern methods use neural networks (PointNet, PointNet++) to directly process point clouds.
Clustering: Group points into objects. DBSCAN (density-based) or Euclidean clustering identifies distinct targets. A cluster might be a vehicle, pedestrian, or debris.
Bounding box prediction: Fit 3D boxes around clusters to estimate object size, location, and orientation.
Tracking: Link detections across frames. A vehicle detected at (100m, 10m) at t=0 and (99m, 10m) at t=0.1s is probably the same vehicle. Kalman filtering or newer deep-learning trackers maintain identity.
Obstacle avoidance: For robotics/drones, compute safe paths through point clouds without collision.
Self-driving cars run all of this in real time (~100 ms latency budget), which is why LiDAR processing is compute-intensive. Modern automotive stacks use specialized hardware (NVIDIA Drive, Tesla's custom chips) to process LiDAR + radar + camera data in parallel.
Source: Qi et al. (2017) — "PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation" — IEEE CVPR.
Performance Comparison in Adverse Weather
Real-world testing reveals how sensors fail:
Rain: LiDAR works through light rain reasonably. Heavy rain scatters light, reducing effective range by 50–70%. Radar degrades less.
Fog: LiDAR degrades significantly; 1550 nm better than 905 nm, but both suffer. Radar largely unaffected.
Snow: Reflective snow can overwhelm LiDAR; radar also struggles. Cameras handle snow reasonably if snow isn't blocking the lens.
Mud/dust clouds: LiDAR works fine; cameras can't see. This is a rare but real scenario in off-road autonomy.
The lesson: sensor fusion is essential. No single sensor is universally robust. Redundancy and diversity of failure modes are why leading robotaxi services layer LiDAR + radar + camera + GPS/HD maps.
Cost Breakdown and Market Trajectory
Mechanical LiDAR (Velodyne HDL-64): ~$85,000 (2010s) → now rare due to reliability and cost.
Automotive solid-state LiDAR (2026 volumes):
- Hesai (ToF, 128 channels): ~$400–600 in volume
- Luminar (1550 nm, 200+ points/pixel): ~$1,000–1,500
- Innoviz (ToF, compact): ~$500–700
- Smartphone LiDAR (flash, ~5 m range): ~$50 (cost to smartphone manufacturer)
Cost driver breakdown:
- Laser/emitter: 15–20%
- Detector array (photodiodes/SPAD): 20–30%
- Optics/scanning mechanism: 20–25%
- Electronics/DSP: 15–20%
- Packaging/assembly: 10–15%
Future trajectory: Solid-state designs (flash, optical phased arrays, MEMS) eliminate mechanical parts, enabling sub-$200 cost by 2029 for automotive-grade LiDARs. Price pressure from Chinese suppliers (Hesai, Innovusion) is accelerating this.
Market projections: LiDAR for automotive TAM (total addressable market) reaches ~$10 billion by 2030 if adoption accelerates beyond current 50–60% of new EVs.
Source: Yole Group, "LiDAR for Automotive and Industrial Applications" market reports (2024–2025).
4D Imaging Radar: Emerging Competitor
Imaging radar (4D radar) is an emerging technology that combines traditional radar (excellent in fog/rain, velocity measurement) with imaging capability (medium spatial resolution, Doppler velocity per pixel).
Advantages over LiDAR:
- Excellent performance in heavy rain/fog
- Simultaneous velocity measurement (no Doppler ambiguity)
- Lower power, cheaper (potentially)
Disadvantages:
- Resolution much lower than LiDAR (sparse)
- Immature technology (most systems still in development)
- Unproven in production autonomy
Companies pushing 4D radar: Arbe, Uhnder, Aptiv, Autoliv. As of 2026, no robotaxi uses 4D imaging radar as a primary sensor, though trials are underway.
Whether 4D radar will displace LiDAR remains open. Current evidence suggests sensor fusion (LiDAR + 4D radar + camera) might be the mature solution.
Source: IEEE Transactions on Vehicular Technology (2023) — "4D Imaging Radar: Tracking and Segmentation Performance" — comparative analysis.
LiDAR in Autonomous Drones and Robotics
Beyond self-driving cars, LiDAR has become indispensable in aerial robotics and industrial automation. Autonomous drones use LiDAR for several critical functions: obstacle avoidance, simultaneous localization and mapping (SLAM), and autonomous navigation in GPS-denied environments (indoors, urban canyons, under forest canopy).
The DJI Zenmuse L2 (launched 2022) is a representative commercial drone payload — combining a LiDAR with RGB camera and thermal camera. Its LiDAR generates a real-time 3D map while the drone navigates, enabling seamless obstacle avoidance and flight path planning without operator input. This is critical for large-area surveying (agriculture, utility infrastructure inspection) where manual piloting would be impractical.
In warehouse robotics, mobile autonomous robots (AMRs) rely on LiDAR for navigation. A robot in a dynamic warehouse — with humans, forklifts, and moving pallets — must detect obstacles in real time and replan routes instantly. LiDAR's millisecond latency (compared to computer vision's processing lag) makes it ideal for safety-critical applications.
The SLAM problem — building a map while simultaneously localizing within it — is particularly well-suited to LiDAR. Traditional visual SLAM (using cameras) requires sufficient lighting and textured surfaces. LiDAR SLAM works day or night and doesn't care about texture. A robot in a featureless warehouse or dark storage area can still localize using LiDAR point clouds and scan matching.
Industrial applications extend to autonomous forklifts, last-mile delivery robots (Starship, Spot), and humanoid robots being deployed in factories and warehouses. Each integrates LiDAR as part of their sensor stack because direct 3D measurement is irreplaceable for safe autonomous navigation in human environments.
The cost of LiDAR in robotics is less prohibitive than in automotive — a $3,000 LiDAR on a $100,000+ industrial robot is a rounding error. This is why robotics adoption of LiDAR happened earlier and more uniformly than in passenger vehicles.
Source: Thrun, S., et al. (2006) "Stanley: The Robot That Won the DARPA Grand Challenge" — foundational robotics/autonomy paper.
Emerging LiDAR Technologies: Solid-State and Quantum Approaches
The limitations of current LiDAR systems are driving innovation in two directions: solid-state scanning (eliminating moving parts) and quantum sensing (improving signal-to-noise).
Solid-State Advances
Optical phased arrays (OPA) steer a coherent beam using phase shifters (similar to antenna arrays). Because there are no mechanical parts, OPA LiDARs can be made extremely compact and reliable. Challenges remain in beam efficiency and cost, but by 2026, multiple startups (Echodyne, Luminar with Iris, Aeva) are shipping OPA-based systems or on clear deployment timelines. An OPA-based system could eventually cost $50–100 if volumes scale, making affordable automotive LiDAR realistic.
Tunable semiconductor lasers allow electronic wavelength tuning. A single OPA transceiver could switch between 905 nm (cost) and 1550 nm (range) depending on conditions — adaptable autonomy sensing in a package the size of a stamp.
Quantum LiDAR
Quantum sensing uses entangled photons to beat the shot-noise limit (the noise floor of classical optics). In principle, quantum-entangled photons can achieve 2× better range for the same average laser power, or detect fainter objects than classical LiDAR. The NIST and BAE Systems have demonstrated laboratory quantum LiDAR systems.
Practical quantum LiDAR faces enormous challenges:
- Entangled photons are expensive to generate
- Detectors must be cryogenically cooled or extremely sensitive
- Atmospheric decoherence limits range
- Integration into mobile platforms is years away
By 2030–2035, quantum LiDAR might be viable for specialized military/aerospace applications. Mass-market automotive quantum LiDAR is a 2040+ possibility if cost and reliability can be solved. For now, quantum LiDAR remains a research domain.
Source: Giovannetti, V., et al. (2011) "Quantum Metrology" — Nature.
Source: Zhang, Z., et al. (2023) "A Quantum Optical Transceiver for Distributed Quantum Sensing" — IEEE Sensors Journal.
Frequently Asked Questions
Q: Will LiDAR function reliably through dense fog and heavy rain?
A: Partially in both cases. Light fog/rain are acceptable; LiDAR degradation is 30–50%. Heavy fog/rain severely reduces range and accuracy. 1550 nm systems perform better than 905 nm. Radar is more robust in severe weather—this is why robotaxi stacks combine LiDAR with radar for redundancy.
Q: Is LiDAR actually eye-safe for human operation?
A: Yes. Commercial automotive LiDARs are designed and certified to Class 1 eye-safe standards under normal operation. Some camera sensors can be damaged by close-range LiDAR (cameras lack eyelids), but human eyes are protected by automatic reflex closure and corneal absorption at 1550 nm. Intentional direct staring into a LiDAR beam should be avoided, but accidental exposure is safe.
Q: Can LiDAR penetrate glass windows or water?
A: Glass: partially. IR light reflects off glass surfaces, confusing detection. LiDAR through windshields loses signal strength and accuracy. Water: bathymetric LiDAR uses green wavelength (532 nm) to penetrate 10–20 meters of shallow water. Standard near-IR LiDARs cannot penetrate water appreciably.
Q: How precisely can LiDAR measure distance and objects?
A: Automotive LiDARs achieve 1–3 cm range precision (standard deviation) at normal distances (~100 meters). Accuracy (systematic error) is ~5–10 cm. Surveying-grade systems achieve sub-centimeter precision and accuracy. Precision degrades with range and target reflectivity; dark, distant objects are measured less accurately than bright, near ones.
Q: Why does Tesla avoid LiDAR while competitors use it?
A: Elon Musk argues that the world is built for human vision; a sufficiently capable vision stack (cameras + neural networks) should equal LiDAR-based autonomy while costing less. Other robotaxi leaders (Waymo, Cruise, Zoox, Mobileye) disagree, citing LiDAR's direct 3D measurement, night operation, and different failure modes as essential for safety redundancy. No major robotaxi operates at scale without LiDAR as of 2026.
Q: Will LiDAR become affordable enough for mass-market vehicles?
A: Yes. Solid-state automotive LiDARs are already $400–600 in high volume from Chinese manufacturers. Sub-$200 systems are plausible by 2028–2029 if manufacturing scales. Cost trajectory follows Moore's Law (halving every ~3 years). At $100–150 per unit, LiDAR will be standard on most vehicles, not a premium feature.
The takeaway
LiDAR (Light Detection and Ranging) is a laser-based 3D sensing technology that fires millions of pulses per second and measures their round-trip time to build dense 3D maps of the environment. The technology spans two physical principles—time-of-flight (ToF) and frequency-modulated continuous wave (FMCW)—and operates at wavelengths (905 nm and 1550 nm) that balance cost, range, and eye safety. LiDAR is already transforming multiple domains: autonomous vehicles use it to navigate without cameras in pitch darkness; archaeologists use it to reveal lost civilizations under jungle canopy; smartphones use compact flash LiDAR for AR and autofocus. The sensor fusion debate—whether LiDAR is necessary for autonomy or whether cameras alone suffice—remains contested; but leading robotaxi services unanimously deploy LiDAR alongside radar and camera. Cost is declining rapidly due to solid-state designs and Chinese manufacturing scale, suggesting LiDAR will become commodity sensor hardware within the decade. The fundamental question for autonomous systems is not "LiDAR or not" but "LiDAR plus what other sensors" to achieve redundancy and robustness across failure modes.
Read next:
Follow new articles
Add SciHub101 to your feed reader to see new science and technology stories.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. AI tools assist with drafting and research. Source documentation in the existing archive is uneven, and articles are being reviewed and corrected. The Editorial Team byline identifies editorial responsibility; it does not certify specialist or independent human review. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles

How Neuralink Works — Brain-Computer Interface
Inside Neuralink's N1 implant — 1024 electrodes, a surgical robot, and the race to give paralyzed patients control of computers with thought alone.

How Tor Works — Anonymous Internet Explained
Tor anonymizes web traffic by bouncing it through volunteer relays in encrypted layers. Here's how onion routing works, and what it doesn't protect.

How Do Self-Driving Cars Actually Work?
Self-driving cars combine cameras, radar, lidar, and AI into one of the most complex robots ever built. Here's how the sensors, software, and decisions actually work.

Computer Vision Object Detection Explained: From YOLO to Vision Transformers
Object detection went from slow two-stage pipelines to real-time single-pass networks and transformer detectors. Here is how each generation works and what the benchmark numbers mean.

How Image Recognition Works: The Science of Convolutional Neural Networks
In 2010 the best image classifier missed the right label in its top five guesses about 28% of the time. By 2015 that was 3.57%. Here is what changed, and how a CNN actually sees.

Complete Technology Guide 2026 — How Modern Tech Works
The complete guide to technology in 2026: how the internet, semiconductors, and quantum computing work, AI hardware, cybersecurity, and what's coming next.