LiDAR vs Time-of-Flight
LiDAR vs Time-of-Flight: The Sensor That Sees the Room
LiDAR and time-of-flight both measure distance by timing light, but they are not the same sensor. Direct ToF times each laser pulse; indirect ToF measures a phase shift. One reaches 200 metres, the other reads a room at 240 Hz. This is how to choose by range, accuracy, sunlight, and surface, with Example Labs' warehouse robot as the worked example.
A robot that cannot measure the room cannot move through it. It can count its own wheel turns, the encoder from Article #25 tells it how far its motors have turned, but the encoder measures the motor, not the world. The world is the pallet at twelve metres, the person stepping out at three, the glass door at five.
To see those, the robot needs a depth sensor. And the depth sensor market has a naming problem: two technologies that both measure distance by timing light, sold under names that make them sound like different species, when the real difference is how they time it, and what that costs.
This article is the selection guide. It is about the two ways of timing light, the three architectures that use them, the surfaces that break all of them, and the arithmetic that decides which one goes on your robot.
The words, in plain English
LiDAR. Light Detection and Ranging. A remote-sensing technology that estimates range by illuminating an object with a collimated laser beam and detecting the reflected light with a photodetector. The name is also written LIDAR, and it stands either for "Light Imaging Detection And Ranging" or for "LIght and raDAR". LiDARs can use ultraviolet, visible, or infrared light, and they can target a wide range of materials, metallic and non-metallic objects, rocks, vegetation, rain, clouds, but excluding highly specular materials.
Time-of-flight (ToF). The measurement principle both technologies share: send light out, time how long it takes to come back, and divide by two. Light travels at a known constant speed, so the round-trip time gives the distance: d = (c × t) ÷ 2. A wall 4 metres away sends the pulse back after 26.7 nanoseconds. To measure that wall to 1 centimetre, the sensor has to time the echo to 67 picoseconds, which is why time-of-flight lidars have special timing electronics.
Direct ToF (dToF). Fire very short laser pulses and time each return directly. The key component is the Single-Photon Avalanche Diode (SPAD), a detector sensitive enough to register a single photon, which lets dToF systems work with weak return signals from distant or low-reflectivity targets. To pick the real signal out of ambient light and electronic noise, dToF commonly uses Time-Correlated Single Photon Counting (TCSPC): the system records the arrival times of many photons over repeated pulses and builds a histogram. Noise photons arrive at random times and spread out evenly, while photons from the laser pulse pile up at the same time bin, so the most frequent time difference marks the most likely round-trip time.
Indirect ToF (iToF). Send continuously modulated light and measure the phase shift between the emitted and received signals. The phase shift is proportional to the travel time. Measuring a phase is easier than timing picoseconds, but a phase repeats every full turn, so the distance wraps around: at 10 MHz the wave repeats every 15 metres, so a wall at 16 metres reads as 1.01 metres. A higher frequency measures more finely but repeats sooner: at 100 MHz, every 1.5 metres. Real sensors combine two or more frequencies to tell the cases apart.
Scanning lidar. A lidar that focuses all of its laser energy into a single narrow beam and moves that beam across the field of view, by a spinning head, a rotating mirror, or a MEMS micro-mirror, painting the scene point by point.
Flash lidar. A lidar that floods the entire scene with a wide flash of infrared light and captures the whole frame at once, with a two-dimensional array of sensors in which every pixel measures its own time of flight simultaneously. No moving parts. Like a camera taking a flash photograph.
Multiple returns. One laser pulse can produce several echoes: the first return is the top of whatever the pulse hit, canopy, roof, ground in the open; intermediate returns are the canopy interior; the last return is the deepest thing the pulse reached, which over vegetation is often the ground. A single pulse may reflect from the upper, middle, and lower tree canopy as well as the ground, and because each reflection has its own X, Y, and Z coordinate, algorithms can classify points into canopy layers and ground.
Stereo vision. The third option: two cameras a known distance apart (the baseline), matching the same visual features in the left and right images and measuring the horizontal shift (the disparity) to triangulate depth. No light emitted at all, which is its strength and its weakness.
Point cloud. The output of a depth sensor: a table of measurements, each row a return with an X, Y, and Z. A lidar point cloud is not a 3D picture. It is a table of pulse measurements, and almost every question people ask of it, how tall is that tree, where is the ground, how dense is the survey, depends on columns other than X, Y, and Z.
One principle, two ways of timing it
Every sensor in this article measures the same thing: how long light takes to go out and come back. The difference is how the timing is done, and the two ways trade against each other in every dimension that matters.
Direct ToF times the pulse. The laser fires very short pulses, and the receiver records the exact moment each reflection arrives. It delivers robust, low-power distance measurement, works well at long range, from a few metres to hundreds of metres, and resists multipath interference and crosstalk better than indirect methods. The trade-offs lie in the hardware: dToF needs very fast timing circuits to resolve nanosecond-scale delays, which raises design complexity and cost, and SPAD arrays are difficult to scale because each pixel needs its own timing electronics. This is how flash lidar and most automotive and drone lidar operate, and how the LIDAR-Lite measures a single point.
Indirect ToF measures the phase. The light's brightness rises and falls smoothly at a frequency, and the sensor compares the wave that comes back with the one sent out. Measuring a phase is easier than timing picoseconds, and iToF sensors can use the conventional CMOS image sensor structure to deliver high-resolution 3D images, which is why iToF is common in phones, robots, and gesture sensors. The cost is the ambiguity: the phase repeats, so a wall beyond the repeat distance wraps around to a short reading, and real sensors must combine two or more frequencies to tell the cases apart. iToF is also the more multipath-sensitive of the two: a pixel receives light from multiple paths, and the measured phase is a mixture of them.
Sony's own summary, from the manufacturer of both kinds: dToF is capable of measurement at long distances, and iToF offers high speed and high image resolution. iToF image sensors are suitable for near- to mid-range applications that require high resolution. SPAD ToF depth sensors are suitable for long-range applications indoors and outdoors, including construction sites that are dangerous to enter, and railway station platforms, which are prone to the influence of outdoor light and a lot of noise.
The three architectures
Scanning lidar concentrates all of its laser energy into one tiny spot, which allows detection of distant objects as far as 200 to 300 metres, and offers high resolution by increasing the number of points scanned. The cost is the steering: a motor, a MEMS mirror, or an optical phased array, plus the motion distortion that comes with it. A scan takes time, and the world moves during the scan, at 5 Hz, each scan is 0.2 seconds old by the time it is finished, and a robot moving at 1.5 m/s has moved 30 centimetres inside its own scan. Mechanical spinning lidar rotates a stack of lasers 360 degrees, the Velodyne HDL-64E, the roof-mounted "coffee can" on early self-driving prototypes, spins 64 layers through a 27° vertical angle at 10 to 30 Hz and sees to around 120 metres. It is the most popular and widely adopted lidar for a reason: one sensor gives the whole 360° view. It is also the one with moving parts, and moving parts are a failure mode and a cost, back in 2017, these systems could sell for $100,000. The industry has since moved to semi-solid-state steering (a small rotating mirror or MEMS micro-mirror), which is what most lidar in production cars uses today.
Flash lidar has no moving parts and captures the whole scene at once, which makes it excellent for tracking fast-moving objects without distortion, there is no rolling-shutter effect, because every pixel measures its time of flight simultaneously. The cost is the photon budget, and it is fundamental. Flash lidar must illuminate the entire target area rather than just 1 ÷ (256 × 256) of it, so for a 256 × 256 image it requires a laser at least 65,000 times more powerful than a scanning lidar to yield the same number of incident photons per pixel, and because the reflected pulse energy is distributed over the entire focal-plane array, another 65,000 times on top of that for the same return-signal strength per pixel. The result: flash lidar's range is limited, usually less than 50 metres, and its wide-angle receiver sees a lot of sunlight, which can easily drown out the laser signal outdoors. It also has significantly worse range resolution than scanning lidar. NASA's answer to the resolution problem is processing: its flash lidar acquires an image with a single laser pulse at video rates (30 frames per second), and an embedded algorithm blends consecutive frames to enhance the 3D resolution by up to eight times. The European Space Agency's flash lidar, built for spacecraft docking, is very precise and very quick, 5 cm accuracy in typical docking situations, and highly effective up to around 1.2 km, where the image quality drops. Space is kind to flash lidar: no atmosphere, no sunlight scattering, no rain. Your warehouse is not space.
The ToF depth camera is the third architecture, and it is usually iToF: a small sensor that instantly acquires 3D information over a short range, at high speed and high resolution. The TeraRanger Evo 64px gives a 64-pixel (8 × 8) depth map at up to 130 frames per second, from 10 centimetres to 5 metres, in a 12-gram module. It is not a lidar in the scanning sense, it does not build a map of the room. It answers one question, fast, about what is in front of it: is something there, and how far away.
The two datasheets, read properly
The numbers that anchor this article come from two datasheets, and they are worth reading the way an engineer reads them, including the footnotes.
Garmin LIDAR-Lite v3 (and v3HP). A single-point dToF lidar: range 0.05 to 40 metres (70% reflective target), resolution ±1 cm, accuracy ±2.5 cm at distances of 2 metres and above (below 2 metres, ±5 cm typical, with nonlinearity present below 1 metre), mean ±1% of distance maximum, update rate greater than 1 kHz typical (the v3 non-HP: up to 500 readings per second). A 905 nm, 1.3 W edge-emitting laser, Class 1 eye-safe, 5 V, 85 mA typical during measurement, 34–38 grams, about $130. It does not rotate, it measures straight ahead, one point, very fast. The footnote that matters: on very dark or absorptive surfaces, effective range drops to 5–15 metres. On retroreflective targets (road signs, reflective tape), range can exceed 40 metres. The range on the datasheet is a range at a reflectivity, not a range.
Terabee TeraRanger Evo 60m. An infrared ToF sensor (LED-based, iToF ranging with automatically varying integration time): range 0.5 to 60 metres indoors, 0.5 to 10–60 metres outdoors, update rate up to 240 readings per second, output resolution 0.5 cm below 14 metres and 2 cm from 14 metres, accuracy ±4 cm in the first 14 metres, 1.5% above 14 metres, field of view approximately 2°, 5 V, 90–330 mA, 9 grams plus a 3-gram backboard, 29 × 29 × 22 mm, eye-safe, under $150. The footnote that matters: the specifications are derived from tests in controlled conditions, and the combination of very bright sunshine and low-reflectivity targets (such as grass) can reduce maximum range to 10 metres or less. The 60 metres is an indoor number, on a cooperative target, in controlled light.
Read together, the two datasheets tell the same story: at 12 metres, the LIDAR-Lite is accurate to ±2.5 cm typical (with a ±1% of distance maximum mean, ±12 cm at 12 metres), and the Evo 60m is accurate to ±4 cm. At 18 metres, the LIDAR-Lite's mean error is ±18 cm and the Evo's is ±27 cm. Neither is a millimetre instrument at range, and the "millimetre-level accuracy" claim for lidar belongs to the survey-grade scanning systems, not to these two.
The Evo family shows the speed–range trade directly: the Evo 600Hz runs at a fixed 600 readings per second, but its range falls to 0.75–8 metres and its accuracy to ±12 cm. You can have speed, or range, or accuracy, pick two, and the datasheet will tell you which one you gave up.
Sunlight, and the surfaces that lie
ToF is smaller, cheaper, lower power, and works in the dark, but it has two enemies, and both are in your building.
Sunlight. Indirect ToF loses range in strong sunlight because the sun is also an infrared source, and it raises the receiver's baseline and reduces the signal-to-noise ratio. The Evo 60m's 60 metres becomes 10 metres or less in bright sunshine on grass. The dToF systems hold up better, the LIDAR-Lite's 1.3 W laser and correlation ranging are built to pick a signal out of noise, but they degrade too: on a very dark surface, its effective range falls to 5–15 metres. At the automotive end, the wavelength choice is the same fight: 905 nm lasers are inexpensive and work with standard silicon detectors, but because the light reaches the retina, eye-safety limits cap the power, which historically held back range, current 905 nm sensors such as Hesai's ATX are rated to about 200 metres. 1550 nm lasers do not reach the retina, so they can safely use more power and reach farther, often 200 metres or more, but they need more expensive detectors, and water absorbs 1550 nm light far more than 905 nm, which can hurt performance in rain. Luminar targets 200 metres for objects reflecting only 10% of the incident light. The most stringent requirement for a vehicle lidar is detecting over 200 metres or more on a freeway at 75 mph, and the cost has fallen from about $75,000 per sensor in 2007 to roughly $200 for high-volume units in 2025.
Surfaces. Very dark, shiny, or transparent surfaces return poor or false readings, and the failure modes are physical, not software bugs:
- Transparent (glass, clear plastic): the beam may pass through, reflect from a background surface, or produce partial reflections from the front surface. The sensor may report the background distance or an unstable intermediate value. A glass door at 5 metres reads as the wall behind it at 8 metres, and the robot drives into the door.
- Glossy (polished metal, still water, glass panes): specular reflection sends the light away from the receiver unless the geometry is favourable. The beam strikes at a diagonal angle and bounces away, creating a blind spot, unless the deflected beam happens to strike another object, in which case the light travels a longer V-shaped or zigzag path, the time of flight is artificially inflated, and the sensor plots a "ghost geometry" much further away than any actual object.
- Very dark (black rubber, felt, black cloth): the surface absorbs the infrared light, the return signal is weak, and the sensor may report max-range, no-return, or noisy distances even when an object is close. Current commercial ToF cameras typically struggle with reflectivity below 20%.
- Multipath (corners, cavities, translucent objects): the light bounces off multiple surfaces before returning. In a corner, the reflected light in all directions is relatively dense, which affects the depth of the object boundary. With a translucent object, part of the light is reflected on the front surface while the other part passes through and reflects from the background, so the measured depth is a mixture of two distances. Multipath is almost impossible to eliminate completely; the mitigations are to remove high-reflective objects from the scene, keep the camera away from translucent objects and corners, and reject outliers.
The rule that follows: test at your surface, in your light. A datasheet range measured on a white wall indoors tells you nothing about a black rubber mat in sunlight. If a part is shiny, no filter will invent the missing returns; you change the lighting, the angle, or the sensor technology.
The third option: stereo vision
Stereo vision is cheap, high resolution, and gives colour, and it needs light and texture, and struggles at night. It recovers depth from two images taken by cameras a known distance apart: depth Z equals focal length times baseline divided by disparity, Z = f·B/d. Because disparity sits in the denominator, depth error grows with the square of distance, a one-pixel matching error barely matters at one metre but becomes large at twenty metres.
With a 5 cm baseline and the dot located to a tenth of a pixel, the error is 0.3 cm at 1 metre, 2.6 cm at 3 metres, and 18.3 cm at 8 metres. Time of flight's error stays about the same (here 1.5 cm, for 100 ps of timing jitter). This quadratic error growth is the defining limitation of stereo vision, and it drives most design choices: a wider baseline improves accuracy at distance but enlarges the blind zone near the cameras and increases occlusion.
Many robots use a baseline of roughly 6 to 12 cm for manipulation and wider rigs for driving.
Textureless and repetitive surfaces defeat passive matching, because many candidate pixels look alike, white walls, glossy floors, and fences are the classic failure cases. Active stereo projects an infrared dot pattern to add texture, which helps on blank walls and tabletops indoors. And stereo has one advantage no active sensor can match: because it relies on ambient light and natural scene texture instead of projecting infrared patterns or laser pulses, multiple cameras can operate side by side without interfering with each other, indoors and outdoors, even in direct sunlight.
A ZED stereo camera gives depth from a few centimetres to 35 metres, with full-resolution colour and depth from the same sensor pair, but avoid very low-light environments, or add scene illumination.
The three technologies, in one line each: lidar measures the room by timing its own light, and reaches furthest; ToF measures the room by timing its own light, and answers fastest at short range; stereo measures the room by comparing two pictures, and costs the least, but only where there is light and texture to compare.
Where the sensor meets the filter
A depth reading is not a position. It is a measurement, with an error, and the error has a shape: the LIDAR-Lite's ±2.5 cm, the Evo's ±4 cm, the stereo camera's error that grows with the square of the distance. Article #4's Kalman filter is what turns those measurements into a position the robot can trust: it weights each reading by its noise, it predicts where the robot should be from the IMU, and it corrects the prediction with the measurement, and it does this at the IMU's 100+ Hz rate, not the sensor's.
The filter cannot invent a reading the sensor never produced, and it cannot fix a reading the sensor got wrong, a glass door that reads as 8 metres when it is 5 metres is a measurement the filter will believe, because the sensor believed it first. That is why the sensing review starts with the surfaces, not the filter.
Article #2's event camera is the other half of the same argument: the event camera reports change, not frames, and it sees the person stepping out of the aisle the moment they move, while the lidar is still painting its scan, point by point, and the ToF camera is still waiting for its next frame. The depth sensor tells the robot where things are. The event camera tells it what just changed.
The filter fuses them. And Article #9's on-device model is what consumes the result: the depth map is the input to the obstacle avoidance, the path planning, and the grasp, and the model's accuracy is bounded by the sensor's accuracy, not by the model's size.
Example Labs is a made-up company we use for worked examples.
DEMO
Example Labs' warehouse robot, from Articles #10 and #20: an autonomous mobile robot (AMR) with the two-link bench arm on top, working a 20 m × 12 m warehouse aisle. It must see a pallet at 12 metres, a forklift at 18 metres, a person stepping out at 3 metres, and a glass door at 5 metres, and it moves at 1.5 m/s. Every number below was run in Python before writing; the demo's arithmetic is the article's own, verified before writing, and the robot and its warehouse are the worked example, not a measured project.
Step 1: The timing budget: what a centimetre costs
Light travels at 299,792,458 m/s. The round trip to a wall and back:
- 0.5 m → 3.3 ns
- 1 m → 6.7 ns
- 5 m → 33.4 ns
- 12 m → 80.1 ns
- 18 m → 120.1 ns
- 40 m → 266.9 ns
- 200 m → 1,334.3 ns
To resolve 1 cm of distance, the timing must resolve 66.7 picoseconds. To resolve the LIDAR-Lite's ±2.5 cm, it must resolve 166.8 picoseconds. That is why the timing electronics are the expensive part of a dToF sensor, and why the cheap lidars do not time anything at all, they shine a laser dot and triangulate it with a camera a few centimetres to the side, and their error grows with the square of the distance.
Step 2
At 12 metres, the LIDAR-Lite is accurate to ±2.5 cm typical (with a ±1% of distance maximum mean, ±12 cm at 12 metres), and the Evo 60m is accurate to ±4 cm. At 18 metres, the LIDAR-Lite's mean error is ±18 cm and the Evo's is ±27 cm. Both are fine for "something is there", and neither is fine for docking.
The pallet at 12 metres is a detection problem, not a measurement problem, and both sensors solve it. The difference is the sunlight at the dock door: the Evo's 60 metres is an indoor number, and in bright sunshine on a low-reflectivity target it collapses to 10 metres or less, so the forklift at 18 metres, seen through the dock door in sunlight, is invisible to the Evo and visible to the LIDAR-Lite.
The selection rule: the sensor that works in your light, on your surface, at your distance, not the sensor with the biggest number on the datasheet.
Step 3: The angular problem: the chair leg at 18 metres
A scanning lidar at 10 Hz takes 400 points per turn, 0.9° apart. That spacing is in angle, so the points spread out with distance: 7.9 cm apart on a wall 5 metres away, 18.9 cm at 12 metres, 28.3 cm at 18 metres. A 10 cm chair leg at 18 metres falls between the points, the lidar never sees it. At 0.9° spacing, a 10 cm object is invisible beyond 6.37 metres.
The fix is angular resolution, not range: at 0.1° spacing, the points are 3.1 cm apart at 18 metres, and the chair leg is seen. The question to ask about a scanning lidar is not "how far does it see" but "how far apart are its points at the distance I care about", and the answer is the angular spacing times the distance.
Step 4: The motion problem: the robot outruns its own map
The AMR moves at 1.5 m/s. At a 10 Hz scan rate, it moves 15 cm between scans; at 5 Hz, each scan is 0.2 seconds old by the time it is finished, and the robot has moved 30 cm inside its own scan, the first and last points of the same scan describe a robot 30 cm apart. At 240 Hz (the Evo 60m), it moves 0.6 cm per reading; at 600 Hz (the Evo 600Hz), 0.25 cm.
The flash lidar captures the whole frame at once, so it has no motion distortion at all, which is why it is the right sensor for a fast-moving object, even though its range is shorter. The update rate is not a specification. It is the distance the robot travels blind between measurements.
Step 5: The surfaces: the glass door and the black mat
The glass door at 5 metres: the ToF sensor sees through it and reports the wall behind at 8 metres. The robot drives into the door. The black rubber mat: it absorbs the infrared light, and the sensor reports max-range or no-return, the robot reads a hole in the floor where there is a mat.
Neither failure is fixable by filtering, because the sensor's reading is wrong in a way the filter cannot detect. The fix is a second modality: a stereo camera, which works in daylight and sees the door frame, and the event camera from Article #2, which sees the person stepping out the moment they move. The depth sensor tells the robot where things are.
The second sensor tells it what the first one got wrong.
Step 6: The selection, per zone
- The aisle (12–18 m, sunlight at the dock door): a scanning dToF lidar, the LIDAR-Lite class for a single-point range, or a 360° scanner for the full view. It is the only sensor in the comparison that holds its range in sunlight on a dark surface, and its ±2.5 cm typical accuracy at 12 metres is the best of the three at that distance.
- The near field (0.5–2 m, the arm picking a box): an iToF depth camera, the Evo 64px class, 8 × 8 pixels at up to 130 frames per second, ±5 cm at 0.1–1 metre. It answers fast, at short range, where the lidar's angular spacing would miss the box's edge.
- The glass, the black mat, and the person: a stereo camera as the second modality, plus the event camera from Article #2. Stereo is the only sensor that works in direct sunlight without emitting anything, and the event camera is the only one that sees the change the moment it happens.
The choice is per zone, not per robot, the same way Article #25's encoder choice was per joint, not per robot.
The decision test
Five questions, in order. If you cannot answer the first, the rest do not matter.
- What is the distance, the accuracy, and the frame rate, as a triplet? A sensor that meets two of the three is the wrong sensor, and no amount of filtering recovers missing accuracy. Write the requirement as a triplet: distance, accuracy, frame rate.
- What is the surface, and what is the light? Test at your surface, in your light. A white wall indoors tells you nothing about a black rubber mat in sunlight. If the surface is glass, shiny, or very dark, the sensor's reading is wrong in a way the filter cannot detect, and you need a second modality.
- How far apart are the points at the distance you care about? Angular spacing times distance. A 10 cm obstacle is invisible beyond 6.37 metres at 0.9° spacing. The question is not "how far does it see" but "how far apart are its points".
- How far does the robot travel between measurements? At 1.5 m/s, a 10 Hz scan is 15 cm of blindness; a 5 Hz scan smears the map by 30 cm. The update rate is the distance the robot travels blind.
- What does the sensor do when it cannot see? Does it report max-range, no-return, or a wrong number? A sensor that reports "no return" is honest. A sensor that reports the wall behind the glass door is lying, and the filter will believe it.
Objections
"But lidar is expensive, can't we just use a camera?" You can, where there is light and texture. Stereo is cheap, high resolution, and gives colour, but its error grows with the square of the distance, it needs texture to match, and it struggles at night. The camera is the right sensor for the glass door and the person. It is the wrong sensor for the pallet at 12 metres in the dark. The question is not which sensor is cheapest. It is which sensor is right for the zone.
"But the ToF sensor's datasheet says 60 metres." It says 60 metres indoors, on a cooperative target, in controlled light. In bright sunshine on grass, it says 10 metres or less, in the footnote. The range on a datasheet is a range at a reflectivity, in a light, on a surface. Read the footnote before you read the headline.
"But we can just filter the bad readings out." You can filter noise. You cannot filter a wrong reading that looks like a right one, the glass door that reads as 8 metres, the black mat that reads as a hole, the corner that reads as a ghost. The filter weights each reading by its noise, and a confident wrong reading gets a high weight. The sensing review starts with the surfaces, not the filter.
"But flash lidar has no moving parts, so it must be better." It is better at one thing: capturing the whole scene at once, with no motion distortion. It is worse at two things: range, and range resolution, because spreading the light across the whole scene means 65,000 times less power per pixel than a scanning lidar at 256 × 256. The right sensor for a fast-moving object at short range is the flash lidar. The right sensor for a stalled car at 200 metres is the scanning lidar. The choice is per zone.
Sources
- GoPhotonics, "What is ToF (Time-of-Flight) LiDAR?"
What we used
A time-of-flight lidar measures distance by timing how long light takes to travel to an object and back, d = (c × t) ÷ 2; ToF is popular because it is compact, fast, and works without moving parts in many designs; depending on how it is built, a system can measure objects a few centimetres away for gesture control, or targets hundreds of metres away for automotive LiDAR; direct ToF (dToF) measures the flight time of light pulses head-on, the laser fires very short pulses, and the receiver records the exact moment each reflection arrives; the key component is the Single-Photon Avalanche Diode (SPAD), sensitive enough to register a single photon, which lets dToF systems work with weak return signals from distant or low-reflectivity targets; dToF commonly uses Time-Correlated Single Photon Counting (TCSPC), recording the arrival times of many photons over repeated pulses and building a histogram, noise photons arrive at random times and spread out evenly, while photons from the laser pulse pile up at the same time bin; dToF delivers robust, low-power distance measurement, works well at long range from a few metres to hundreds of metres, and resists multipath interference and crosstalk better than indirect methods; the trade-offs are very fast timing circuits to resolve nanosecond-scale delays, which raises design complexity and cost, and SPAD arrays that are difficult to scale because each pixel needs its own timing electronics
- Sony Semiconductor Solutions, "ToF (Time of Flight) Technology, Industrial Use"
What we used
ToF depth sensors measure the distance to a target object by emitting light from a source, usually a laser or LED, and detecting the reflected light, based on the time elapsed between the two; direct ToF involves a direct measurement of the time until reflection is detected, allowing it to also be used for measurement of targets at long distances; indirect ToF measures distance by accumulating the electric charge generated by the reflected light and discerning any phase shift between emitted and reflected light; dToF uses SPAD pixels with avalanche multiplication to amplify electrons from a single incident photon, so the sensors can detect even a faint light while also being capable of measuring longer distances; iToF can use the conventional CMOS image sensor structure and deliver high-resolution 3D images, and excels especially at near- to mid-range sensing with high-speed acquisition of high-resolution 3D information; SPAD ToF depth sensors are suitable for long-range applications indoors and outdoors, including construction sites that are dangerous to enter and railway station platforms prone to outdoor light and noise
- Tran Khanh Thanh, "Mobile robots from scratch, part 8: lidar, how it measures"
What we used
The direct way to measure a distance with light is to send a short pulse and time how long it takes to come back, d = c·t ÷ 2; a wall 4 m away sends the pulse back after 26.7 ns, and to measure to 1 cm the lidar has to time the echo to 67 ps; the echo gets weaker with distance, roughly as 1/d² for a wall, at 4 m it is 16 times weaker than at 1 m; cheap lidars often don't time anything, they shine a laser dot and a small camera a few centimetres to the side sees where the dot lands, and the error grows with the square of the distance (with a 5 cm baseline and the dot located to a tenth of a pixel, the error is 0.3 cm at 1 m, 2.6 cm at 3 m, and 18.3 cm at 8 m, while time of flight's error stays about the same, here 1.5 cm, for 100 ps of timing jitter); the other way is amplitude-modulated continuous wave, the light's brightness rises and falls smoothly at a frequency f, and the echo is delayed, so its wave is shifted by a phase φ = 4πfd ÷ c; a phase repeats every full turn, at 10 MHz the wave repeats every 15 m, so a wall at 16 m reads as 1.01 m, and at 100 MHz every 1.5 m, so real sensors combine two or more frequencies; a 2-D lidar takes 4,000 ranges a second and turns 10 times a second, 400 points a turn, 0.9° apart, which is 7.9 cm apart on a wall 5 m away, and at 20 Hz 200 points, 15.7 cm apart at 5 m, while at 5 Hz 800 points but each scan is 0.2 s old by the time it is finished
- Hansard et al., "An Overview of Depth Cameras and Range Scanners Based on Time-of-Flight Technologies" (arXiv:2012.06772)
What we used
Time-of-flight cameras measure the depths of scene points by illuminating the scene with a controlled laser or LED source and analysing the reflected light; the two principles are pulsed-light cameras, which measure directly the time taken for a light pulse to travel from the device to the object and back, and continuous-wave modulated-light cameras, which measure the phase difference between the emitted and received signals and obtain the travel time indirectly; LIDAR stands either for "Light Imaging Detection And Ranging" or for "LIght and raDAR", and estimates range by illuminating an object with a collimated laser beam and detecting the reflected light; because LIDARs use a fine laser beam, they can estimate distance with high resolution, and can target metallic and non-metallic objects, rocks, vegetation, rain, clouds, and so on, but excluding highly specular materials; a scannerless device diverges the light from a single emitter so the entire scene of interest is illuminated, and the reflected light is imaged onto a two-dimensional array of photodetectors, a TOF depth camera
- Garmin, "LIDAR-Lite v3HP Operation Manual and Technical Specifications" (April 2020)
What we used
Range 40 m (70% reflective target); resolution ±1 cm; accuracy below 2 m ±5 cm typical (nonlinearity present below 1 m); accuracy at 2 m and above ±2.5 cm typical; mean ±1% of distance maximum; ripple ±1% of distance maximum; update rate (70% reflective target) greater than 1 kHz typical; size 40.18 × 54.99 × 35 mm, weight 38 g; Class 1 laser product, classified EN/IEC 60825-1 2014; wavelength 905 nm
- SparkFun / Garmin LIDAR-Lite v3 datasheet
What we used
Range 0 to 40 m, accuracy ±2.5 cm at distances greater than 1 m, power 4.75–5 V DC, current consumption 105 mA idle and 130 mA continuous, rep rate 1–500 Hz, laser wavelength 905 nm at 1.3 W peak power, beam divergence 4 mrad × 2 mrad, optical aperture 12.5 mm, interface I2C or PWM, dimensions 20 × 48 × 40 mm; the LIDAR-Lite v3 uses correlation ranging, a signature matching technique that estimates time delay by electronically sliding a stored transmit reference over the received signal to find the best match; on very dark or absorptive surfaces, effective range drops to 5–15 m, and on retroreflective targets (road signs, reflective tape) range can exceed 40 m; price $129.95 (Adafruit)
- Terabee, "TeraRanger Evo 60m Specification Sheet"
What we used
Principle: infrared time-of-flight; range 0.5 m up to 60 m (indoor), 0.5 m up to 10–60 m (outdoor); update rate up to 240 readings per second; output resolution 0.5 cm below 14 m, 2 cm from 14 m; accuracy ±4 cm in the first 14 m, 1.5% above 14 m; field of view approximately 2°; supply voltage 5 V DC ±5%; supply current 90–330 mA; interfaces USB 2.0 Micro-B, UART, I2C; weight 9 g (sensor) + 3 g (backboard); dimensions approximately 29 × 29 × 22 mm; eye safety yes (CE certified); specifications are derived from tests in controlled conditions, and the combination of very bright sunshine and low-reflectivity targets (such as grass) can reduce maximum range to 10 m or less. The Evo 60m uses LED technology rather than laser, which gives it a field of view, instead of measuring distance based on a very small point of light, the sensor measures over an area
- Terabee, "TeraRanger Evo 600Hz Specification Sheet"
What we used
Principle: infrared time-of-flight; range 0.75 m up to 8 m; update rate fixed at 600 readings per second; output resolution 0.5 cm; accuracy ±12 cm; field of view approximately 2°
- Terabee, "TeraRanger Evo 64px Specification Sheet"
What we used
64-pixel (8 × 8 matrix) 3D depth map; range 0.1 m up to 5 m (close-range mode) and 0.5 m up to 5 m (fast mode); update rate up to 80 fps and up to 130 fps; output resolution 0.5 cm; accuracy ±5 cm from 0.1 m to 1 m, ±10 cm below 3 m, ±15 cm above 3 m; field of view 15°; weight 12 g (sensor) + 3 g (backboard)
- MDPI Sensors, "Radiometric Calibration of an Inexpensive LED-Based Lidar Sensor" (2020, 20/18/5215)
What we used
The TeraRanger Evo 60 m by Terabee is an LED-based lidar sensor with an advertised range of up to 60 m and reported centimetre-level precision, economically priced under $150; a key difference from traditional laser-based lidar sensors is the use of an ITOF ranging method, which features an automatically varying integration time, if the return signal power is low, the integration time increases to generate a more reliable range measurement; the maximum range of 60 m is only achieved in optimal conditions, i.e., when observing a highly reflective target in low ambient light conditions; the specifications indicate a change in resolution and accuracy at 14 m, and testing with several Evo 60 m sensors confirms this threshold exists
- Laser Focus World, "Real time comes to three-dimensional ladar"
What we used
Single-point MEMS scanning and single-pulse flash are two approaches to real-time 3-D ladar; flash ladar uses a focal-plane array, a 2-D array of photodiodes that has no moving parts and is therefore more rugged, though it has significantly worse range resolution than scanning ladar; for a 256 × 256 image, a flash ladar would require a laser at least 65,000 times more powerful than a MEMS ladar to yield the same number of incident photons per pixel, and because the reflected pulse energy is distributed over the entire focal-plane array, another 65,000 times on top of that for the same return-signal strength per pixel
- RF Wireless World, "Flash LiDAR vs. Scanning LiDAR: Key Differences & Trade offs"
What we used
Scanning lidar paints the scene point by point, focusing all of its laser energy into a single narrow beam and moving it across the field of view; it detects distant objects as far as 200 to 300 metres and offers high resolution by increasing the number of points scanned; flash lidar operates like a digital camera equipped with a flash, it floods the entire scene with a wide flash of infrared light, the reflection is captured by a 2D array of sensors, and every pixel measures its own time of flight simultaneously; it offers a high frame rate and captures the entire 3D scene instantly, with zero moving parts, so it is highly resistant to vibration and mechanical failure; spreading the light thin means its range is limited, usually less than 50 metres, and the wide-angle receiver sees a lot of sunlight, which can easily drown out the laser signal outdoors; scanning lidar can suffer motion distortion (a rolling-shutter effect), while flash lidar is immune to motion blur
- NASA Langley, "Super Resolution 3D Flash LIDAR"
What we used
Traditional scanning lidar techniques generate an image frame by raster scanning one laser pulse per pixel at a time, whereas flash lidar acquires an image much like an ordinary camera, generating an image using a single laser pulse; the embedded image processing algorithm simultaneously corrects, enhances, and derives relative motion, processing the image sequence into a high-resolution 3D synthetic image at video rates (30 frames per second), with up to eight times image resolution enhancement
- European Space Agency, "New LIDAR uses flash to take a snapshot"
What we used
Imaging flash lidars are a key enabling technology for future exploration missions and for space operations involving the rendezvous between two spacecraft in orbit; the flash lidar fires the laser pulse just once to highlight the entire area, and a large detector records the entire surface at once, like a camera taking a flash photograph; it is very precise and very quick, able to see with a 5 cm accuracy in typical docking situations, and highly effective up to around 1.2 km, when the image quality drops
- ThinkAutonomous, "A Hardcore Look at 9 types of LiDAR systems"
What we used
The Velodyne HDL-64E uses a spinning 64-layer laser detector (64 horizontal planes) to cover a vertical angle of 27°, spins at 360° at 10 to 30 Hz, and can see at around 120 m range; a mechanical rotating lidar provides a 360° view with one sensor, but it is moving, which introduces a risk of failure as well as a higher cost (back in 2017, these systems could be sold for $100,000), and they operate at 905 nm, which is good for long distances but has a risk of damaging the human eyes; solid-state lidar systems have no moving parts, with MEMS (a mirror moves to scan the environment) and OPA (optical phased array) as the two common approaches
- Lidar News, "Automotive Lidar: The Complete Industry Guide"
What we used
Most automotive lidar uses time of flight: the sensor sends a short laser pulse and measures the round-trip time to an object and back, and distance equals half the round-trip time multiplied by the speed of light; a second method, frequency-modulated continuous-wave (FMCW) lidar, sends a continuous beam whose frequency changes over time and can measure an object's speed directly, with every point, using the Doppler effect; 905 nm lasers are inexpensive and work with standard silicon detectors, but because this light reaches the retina, eye-safety limits cap the power, which historically held back range, current 905 nm sensors such as Hesai's ATX are rated to about 200 m; 1550 nm lasers do not reach the retina, so they can safely use more power and reach farther, often 200 m or more, but they need more expensive detectors, and water absorbs 1550 nm light far more than 905 nm, which can hurt performance in rain; automotive lidar is designed to meet Class 1 laser safety under IEC 60825-1, meaning it is eye-safe in normal use; mechanical spinning lidar rotates a stack of lasers 360 degrees, semi-solid-state lidar keeps the sensor body fixed and steers the beam with a small rotating mirror or a MEMS micro-mirror (most lidar in production cars today), and flash lidar lights the whole scene at once with no moving parts, though eye-safety limits spread its power across the full field of view, which shortens its range; automotive lidar cost has fallen from about $75,000 per sensor in 2007 to roughly $200 for high-volume units in 2025
- Laser Focus World, "1550 nm triple-junction laser diode excels for automotive lidar"
What we used
The most stringent requirement for a successful lidar on a vehicle is its capability to successfully detect over a range of 200 m or more on a freeway at a typical driving speed of 75 mph; most lidar systems available in the marketplace require many laser emitters and a complex system design to detect targets over 200 m away, as a single laser emitter can hardly reach over 100 m based on lasers that operate at short wavelengths such as 905 nm; Luminar is targeting a range of 200 m for objects reflecting only 10% of incident light
- Spatial Workflow, "LiDAR Point Clouds Explained: Returns, Classes and Intensity"
What we used
A row is one return: a detectable echo from one outgoing laser pulse; a pulse over bare ground gives one return, while a pulse through a canopy can give several, as energy reflects off leaves, branches, and finally the ground; first returns are the top of whatever the pulse hit, last returns are the deepest thing the pulse reached (over vegetation, often the ground), and intermediate returns are canopy interior; in a USGS 3DEP survey over Brooks Camp, Alaska, the returns were 59.5% first, 36.1% second, 4.4% third, and 0.1% fourth, with 76.5% of points coming from a pulse that produced more than one return; every ground point is a last return, and only 38.4% of last returns are classified ground
- ROCK Robotic, "Enhancing Aerial LiDAR with Multiple Returns: A White Paper"
What we used
Traditional discrete-return lidar sensors recorded only the first and last return from each laser pulse, so information about intermediate canopy layers was lost; advances in receiver electronics and waveform processing have made it possible to record multiple returns from a single pulse, and today's airborne lidar sensors can extract up to seven returns; multiple returns enable lidar to "see" through vegetation by recording successive reflections as the laser pulse passes through gaps in leaves and branches; a single pulse may reflect from the upper, middle, and lower tree canopy as well as the ground, and because each reflection has its own X, Y, and Z coordinate, algorithms can classify points into canopy layers and ground, enabling accurate digital elevation models even in dense forests
- Rohrbach, "How Vegetation Shapes LiDAR Ground Returns"
What we used
A laser pulse is not a thin line; it leaves the sensor as a cone and reaches the vegetation as a small illuminated area, the footprint, and part of the energy hits leaves and branches while part passes through gaps in the canopy and continues downward, with each surface that sends back enough energy producing an echo; in a 2017 study of Wytham Woods (Simpson, Smith, and Wooster), the terrain model from the leaf-on survey had a vertical error (RMSE) of 0.83 metres, while the leaf-off survey reached 0.22 metres, and the worst errors came from low undergrowth below 3.5 metres, such as ferns and brambles, where errors exceeded one metre; a pulse that lasts a few nanoseconds is more than a metre long in the air, so when two surfaces are close together, the echoes overlap and merge into one
- Webiano, "Stereo vision, Robotics"
What we used
Stereo vision is the technique of recovering three-dimensional depth from two images taken by cameras a known distance apart; depth Z equals focal length times baseline divided by disparity, Z = f·B/d; because disparity sits in the denominator, depth error grows with the square of distance, a one-pixel matching error barely matters at one metre but becomes large at twenty metres, and this quadratic error growth is the defining limitation of stereo vision; a wider baseline improves accuracy at distance but enlarges the blind zone near the cameras and increases occlusion; many robots use a baseline of roughly 6 to 12 cm for manipulation and wider rigs for driving; textureless and repetitive surfaces defeat passive matching because many candidate pixels look alike, white walls, glossy floors, and fences are the classic failure cases, which is why active stereo and learned methods have spread
- StereoLabs, "How does the ZED 3D perception work?"
What we used
ZED stereo cameras are passive stereo vision cameras that reproduce human binocular vision, embedding two synchronised image sensors separated by a fixed, factory-calibrated distance called the baseline; by matching the same visual features in the left and right images and measuring their horizontal shift (the disparity), the SDK computes the distance of every pixel through triangulation, building a dense three-dimensional model in real time; because ZED cameras rely on ambient light and natural scene texture instead of projecting infrared patterns or laser pulses, they are fully passive sensors, so multiple cameras can operate side by side without interfering with each other, indoors and outdoors, even in direct sunlight; passive stereo gives depth from a few centimetres (ZED X Nano) up to 35 m (ZED X and ZED 2i with narrow lenses), with full-resolution colour and depth from the same sensor pair; scene texture matters, homogeneous, textureless, or specular surfaces (white walls, green screens, glass, mirrors) can generate unstable or missing depth; avoid very low-light environments, or add scene illumination
- Cursa, "Time-of-Flight Optical Sensors: Direct Distance with Photons"
What we used
Optical ToF distance sensors emit modulated or pulsed infrared light and measure how long the photons take to return after reflecting from the scene; the module estimates the delay of the dominant return within its field of view, and if multiple surfaces contribute, the reported distance may be a compromise or may jump between targets; transparent surfaces (glass, clear plastic) let the beam pass through, reflect from a background surface, or produce partial reflections, so the sensor may report the background distance or an unstable intermediate value; glossy surfaces cause specular reflection that can send light away from the receiver, with intermittent returns as the robot moves; very dark or black surfaces absorb IR strongly, and the sensor may report max-range, no-return, or noisy distances even when an object is close; multi-path reflections (light bouncing off multiple surfaces before returning, e.g., shiny floor plus wall) bias distance high or alternate between paths as the robot moves, and are especially problematic in tight indoor spaces with glossy materials
- GoerMicro, "What is the multipath interference of indirect Time of Flight (iToF)?"
What we used
In theory, an individual pixel in an iToF camera receives light from one path, but in practice it receives light from multiple paths; with a specular object, the beam is blocked from returning to the sensor; with a translucent object, part of the light is reflected on the front surface while the other part passes through and is reflected by the background object, guiding the depth measurement error; with multiple obstacles in the path of light, multi-surface reflection usually occurs in corners or depressions and affects the depth of the object boundary; lens flare or ghosting is multipath interference caused by scattering between the optical internal elements of the lens; multipath interference is almost impossible to eliminate completely, but it can be minimised by removing high-reflective objects from the scene and keeping the camera away from translucent objects and corners
- Article #2, "Event-Based Vision: Cameras That Report Change, Not Frames" (Senova, 2026-11-19)
What we used
The sensor whose readings the filter must trust; the event camera reports change, not frames, and sees the person stepping out the moment they move
- Article #4, "Sensor Fusion and the Kalman Filter, in Plain Words" (Senova, 2026-11-19)
What we used
The filter that fuses the depth with the IMU; GPS 1–10 Hz, 5–10 m standard accuracy; IMU 100 Hz or higher, with drift growing with the square of time
- Article #9, "On-Device AI: Running the Model Where the Data Lives" (Senova, 2026-11-19)
What we used
The model that consumes the depth map; its accuracy is bounded by the sensor's accuracy, not by the model's size