Robot Sensors: A Practical Reference

This is the sensor that made autonomous driving imaginable. A rotating head of pulsed lasers and detectors sweeps the scene and times each return, which produces 360° point clouds with centimeter range accuracy at 10–20 Hz. Velodyne's spinning "coffee can" defined the category (and the DARPA-challenge aesthetic); modern descendants pack 32–128 beams into smaller, cheaper, sealed units. It measures geometry directly, with no inference and no learning involved, and that's still its main virtue.

One unit gives true 360° coverage and dense direct 3D geometry, it works in darkness, and the SLAM and detection ecosystems around it are mature. The weaknesses come from the hardware: it's a precision opto-mechanical assembly that spins continuously, so bearings and encoders wear out and shock and vibration shorten its life. The units are big and conspicuous, per-unit cost stayed high for years, and resolution concentrates in horizontal lines, so vertical detail gets sparse at range.

Pick spinning LiDAR when you need 360° geometry from a single mast-mounted unit at 30–200 m range with centimeter accuracy, and the platform can carry a fist-sized assembly that wears out: outdoor AMRs, off-road and mining vehicles, mapping rigs. Sub-$1k 32-beam units make it the default outdoor navigation sensor. Take it over solid-state whenever full surround coverage from one sensor matters more than point density in a wedge. Avoid it on high-shock or high-vibration platforms, in designs that must embed the sensor flush, and in automotive series production, where one or several solid-state units are the usual answer. If the job is surviving rain, fog, and dust rather than resolving shape, skip LiDAR entirely and use radar.

Range 30–200 m · range accuracy roughly ±2–3 cm · field of view 360° horizontal by 30–45° vertical · 32–128 beams · update rate 10–20 Hz · power draw roughly 8–20 W · unit cost under $1,000 for 32-beam units, into the tens of thousands for high-channel models.

Rain, snow, fog, and dust both return spurious points and absorb signal. Glass and mirrors either read as invisible or place the surface somewhere else entirely. Retroreflectors bloom and ghost. Wet black asphalt absorbs returns. Multiple LiDARs interfere with each other once you run them at scale. Bearing wear ends in dead sectors or total loss, usually gradually but sometimes all at once.

Velodyne HDL-64 (the iconic early unit, now retired), Ouster OS-series digital LiDAR, RoboSense and Hesai mechanical lines on countless robotaxis, delivery robots, and mining trucks; survey and mapping backpacks.

Prices went from $75k (HDL-64) to sub-$1k (Chinese 32-beam units) in fifteen years, one of the steepest cost declines in robotics, driven by Hesai/RoboSense automotive volume. Mechanical spinning units increasingly serve robotics and mapping while automotive design-ins shift to solid-state. What keeps the spinning form factor alive is 360° coverage per dollar, which nothing has beaten yet.

7 minVelodyne Laser TeardownHizook · 10k+ views

4 minLidar 101 from Velodyne LidarVelodyne Lidar · 100k+ views

6 minWhat Is LiDAR and how does LiDAR work?Phoenix LiDAR Systems · 50k+ views

What Is Lidar and How Does It Work? (Ouster) · Lidar for Autonomous Driving: Principles, Challenges, and Trends (arXiv)

These are LiDARs with (almost) nothing spinning. MEMS mirrors, rotating polygon prisms, optical phased arrays, or pure flash illumination steer or spread the beam electronically. The goal is automotive qualification: no wear parts, small enough to hide in a grille or behind a windshield, and manufacturable like electronics rather than instruments. Directional units trade the spinning sensor's 360° view for a forward wedge, so robots use several or add a spinning unit on top.

They tolerate vibration and wear, they're compact enough to embed, point density inside the field of view is high, and costs are falling on semiconductor-style curves. The weaknesses: each unit covers a limited field of view; MEMS and polygon versions still have micro-moving parts, so the "solid-state" label is elastic marketing; flash LiDAR's power budget caps its range; and optical phased arrays, the version with genuinely nothing moving, are still barely commercial after a decade of promises.

Pick solid-state when the sensor must be embedded, shaken, or automotive-qualified and a forward wedge of dense points covers the task: ADAS and highway autonomy, drones, forklift and AMR forward safety zones. Sub-$500 units ride Chinese ADAS volume. If you need wider coverage, use multiple units or add a spinner to cover the seams. Avoid it when one sensor must see 360° (a spinning unit still wins on coverage per dollar) or when your program cannot accept supply concentrated in two Chinese vendors. Treat "solid-state" labels skeptically, since MEMS and polygon units still carry micro-moving parts with shock limits.

Range roughly 30–200 m on MEMS and polygon units, 20–50 m on flash · range accuracy roughly ±2–3 cm · field of view typically 120° horizontal by 25° vertical per unit · update rate 10–20 Hz · power draw roughly 8–20 W · unit cost under $500 at ADAS volume.

These units have the same weather and surface problems as all ToF LiDAR (rain, fog, glass, retroreflector bloom). Windshield mounting adds dirt, condensation, and wiper streaks. MEMS mirrors have resonance and shock limits. A narrow field of view creates coverage seams between units, and objects in those seams go undetected. Fixed-pattern units also interfere with each other, and that interference is harder to dodge than it is between spinners.

Hesai AT-series and RoboSense M-series (the Chinese ADAS volume leaders, on dozens of EV models), Innoviz (BMW), Valeo Scala (the first-ever automotive LiDAR, Audi A8), Luminar (Volvo), AEyE; Livox's prism units on DJI platforms and countless robots.

Chinese ADAS design-ins turned LiDAR into a sub-$500 automotive component and pulled the whole category's costs down, while Western pure-plays consolidated hard (Ouster-Velodyne merger, multiple bankruptcies). Robotics buyers get the spillover: automotive-qualified units at consumer-electronics prices. The strategic caveat is that supply concentrates in two Chinese vendors.

7 minLidar Technologies 101Hesai Technology · 50k+ views

2 minThe MEMS Solid-State LiDAR`s Principle & Features, RS-LiDAR-M1 PreRoboSense · 10k+ views

4 minMEMS Mirror Technologies of Fraunhofer IPMS for LiDAR | Fraunhofer IPMSFraunhofer IPMS · 10k+ views

How Multi-Beam Flash Lidar Works (Ouster) · All-Solid-State Beam Steering via Integrated Optical Phased Arrays (Micromachines)

FMCW LiDAR borrows radar's approach. Instead of timing a pulse, frequency-modulated continuous-wave LiDAR chirps its laser and interferes the return with a local copy. Coherent detection buys two properties pulsed LiDAR can't offer: per-point instantaneous velocity via Doppler (no more inferring motion across frames), and near-immunity to sunlight and other LiDARs, because only light coherent with the local oscillator registers. The price is photonic complexity, since every channel needs a long-coherence laser and interferometric optics.

It measures velocity directly, so a moving pedestrian separates from the static world immediately. Interference immunity comes from the physics rather than from coding tricks, range at eye-safe power is better, and retroreflectors don't blind it. The weaknesses: each channel is a coherent photonic system, so silicon-photonics integration is the whole manufacturing story. Point rates have historically trailed ToF, processing coherent returns takes a lot of DSP, and the technology carries a decade of "next year" scheduling history.

Pick FMCW only when its physics is decisive: per-point velocity for separating movers from the static world at highway closing speeds (trucking is the proving ground), immunity to interference in dense multi-LiDAR fleets, or long range at eye-safe power. Even then, pick it only if your program can absorb few-vendor, young-silicon supply risk on a multi-year timeline. Avoid it for near-term production or anything needing high point rates and a mature ecosystem. ToF LiDAR (spinning or solid-state) remains the default until silicon-photonics integration actually lands, and Mobileye's wind-down is the cautionary reference.

Range roughly 200–300 m at eye-safe power, with long-range modes claimed past 400 m · range accuracy a few centimeters · per-point Doppler velocity resolved to roughly centimeters per second · update rate 10–20 Hz · 1550 nm wavelength · unit cost still in the thousands, from a handful of vendors.

FMCW shares fog and rain attenuation with all optical sensing, though coherent gain helps. Rough surfaces produce speckle noise. Doppler ambiguity requires careful chirp design. Laser coherence degrades with temperature and aging. Ecosystem risk counts as a failure mode too: there are few vendors and the silicon is young, which matters for any program betting on it.

Aeva (Daimler Trucks design-in, the flagship program), Mobileye's internal FMCW program (wound down in 2024), Scantinel, SiLC; academic silicon-photonic beam-steering demos aiming at chip-scale units.

The bet is that LiDAR ends up as a silicon-photonics chip. If that happens, FMCW's extra physics comes nearly free and ToF incumbents face a discontinuity. Until that integration matures, units stay expensive and design-ins stay scarce; Aeva's truck win is the category's existence proof. Watch silicon-photonics fab partnerships rather than demo videos.

3 minFMCW vs. ToF LiDARSiLC Technologies Inc. · 10k+ views

7 minWhat is FMCW Radar and why is it useful?Marshall Bruner · 100k+ views

Coherent Lidar for Ride-Hailing Autonomous Vehicles (arXiv) · Photonic-Electronic Integrated Coherent LiDAR Engine (arXiv)

mmWave radar puts radio waves at 60–77 GHz out of a chip the size of a fingernail. FMCW radar measures range and velocity directly (velocity via Doppler), it sees through the rain, fog, dust, and darkness that blind every optical sensor, and it costs tens of dollars because hundreds of millions of cars carry it. Modern "4D imaging" radars use large virtual antenna arrays to add elevation and enough angular resolution to sketch object shapes, which moves radar from a detection sensor toward a perception sensor.

It works in any weather and any lighting, gives velocity directly for each detection, and is cheap, solid-state, and automotive-hardened. It also penetrates plastic, so it mounts invisibly behind bumpers. The weaknesses: angular resolution is far below LiDAR, so classic units report "something, 40 m, closing at 3 m/s" rather than shapes. Multipath ghosts and clutter mean the tracking layer has to treat detections skeptically, stationary objects have historically been filtered out (the stopped-firetruck problem), and metal dominates returns while pedestrians return very little.

Pick radar whenever operation must continue through rain, fog, dust, or darkness, when direct velocity simplifies tracking, or when the BOM allows tens of dollars for long-range sensing. Use it as the redundancy layer beside cameras or LiDAR in safety cases, mounted invisibly behind plastic. It is the cheapest way to know something is closing at 3 m/s at 40 m. Avoid it as a sole perception sensor wherever classification or shape matters, because it cannot tell a pedestrian from a pole. Never ship the default static-return filtering unaudited. Pair it with a camera for semantics, and check whether imaging-radar resolution has caught up to your use case since you last looked.

Frequency 60–77 GHz · range up to 200–300 m on long-range units · range resolution roughly 4–40 cm depending on bandwidth · angular resolution 10–15° on classic units, near 1° on 4D imaging radar · velocity resolution around 0.1 m/s · update rate 10–20 Hz · unit cost $10–40 per chip.

Multipath reflections create ghost objects under bridges and in tunnels. Guardrails and manhole covers get reported as obstacles. Radar-to-radar interference grows with fleet density. Bumper paint and ice attenuate the signal. Angular ambiguity smears adjacent objects into one. Over-filtering static returns deletes real stopped obstacles, which is the failure behind several accident reports.

Every AEB/ACC-equipped car (Bosch, Continental, Denso corner radars), TI IWR/AWR chips powering robot and drone radars, imaging-radar vendors (Arbe, Zendar, Bosch Gen6), Google Soli's micro-gesture radar, industrial level sensors.

Radar is the cheapest long-range sensing per dollar in existence, on the back of TI/NXP/Infineon silicon economics: a single-chip radar is now a $10–40 part. Robotics adoption lags its merits mostly for software reasons (the data is sparse and unusual, and open-source tooling is thin). Imaging radar is the segment to watch. If resolution keeps climbing at radar prices, it will take over LiDAR's mid-range use cases from below.

18 minFMCW Radar for Autonomous Vehicles | Understanding Radar PrinciplesMATLAB · 100k+ views

2 minWhat is mmWave sensing ? | Mouser Electronics | Texas InstrumentsMouser · 50k+ views

17 minHow do automotive (FMCW) RADARs measure velocity?Marshall Bruner · 50k+ views

The Fundamentals of Millimeter Wave Radar Sensors (Texas Instruments) · 4D Millimeter-Wave Radar in Autonomous Driving: A Survey (arXiv)

Ultrasonic ranging is sonar for pennies. A piezo transducer chirps at 40–200 kHz and times the echo, which measures distance from centimeters to a few meters with zero optics, zero processing, and near-zero cost. It's the parking sensor in the world's cars and the bump-avoidance on a generation of hobby robots. It keeps surviving attempts to declare it obsolete because sound reflects off exactly the surfaces that light passes through, so nothing else detects a sheet of glass as reliably.

It detects glass, mirrors, and transparent obstacles that defeat every optical sensor, it doesn't care about lighting, it's trivially cheap and simple, and it's safe to use anywhere. The weaknesses: it's slow (limited by the speed of sound), the beam is wide and sloppy so it localizes poorly, maximum range is a few meters, and soft materials like foam, cloth, and snow absorb the ping and disappear. Temperature and wind shift the speed of sound, and multiple units need scheduling to avoid crosstalk.

Pick ultrasonic for last-centimeter work and transparent obstacles: docking, parking, and any indoor robot operating near glass walls, partitions, or mirrors that optical sensors read as open space. It costs a few dollars per unit, needs no compute, and doesn't depend on lighting. It is worth carrying as the cheap safety net even on sensor-rich platforms. Avoid it wherever you need range past a few meters, precise bearing, fast update rates, or detection of foam, cloth, and other sound absorbers; for those, short-range ToF or radar does the job. Tesla's removal-and-regret arc shows what happens when cameras are asked to cover the last centimeter alone.

Frequency 40–200 kHz · range roughly 2 cm to 5 m · accuracy about ±1 cm · beam width 30–60°, so bearing is coarse · update rate 10–40 Hz, capped by the round-trip time of sound · power draw well under 1 W · unit cost $1–20.

Specular surfaces at oblique angles bounce echoes away, so a smooth wall at 45° reads as empty space. Sound-absorbent obstacles such as clothing and foam return nothing. Ultrasonic cleaners, air nozzles, and other robots' pingers inject phantom echoes. Rain and wind noise degrade outdoor use. Contamination on the transducer face, like mud or ice, kills sensitivity without any warning.

Automotive park-assist rings (billions of units — Bosch, Murata), HC-SR04 modules on every Arduino robot, Tesla's ultrasonic suite (removed in 2022 when Tesla went camera-only), drone landing altimeters, tank-level industrial sensors.

Ultrasonic is a solved, commoditized technology at $1–20 per unit, and its niche looks permanent, because the glass-detection and last-centimeter problems don't yield to cameras. Tesla's removal and partial walk-back of parking performance illustrates the trade. In robot BOMs it persists as the cheap safety net nobody brags about, which is the right role for it.

9 minHow Do Ultrasonic Distance Sensors Work? - The Learning Circuitelement14 presents · 50k+ views

5 minThe Basics of Ultrasonic SensorsSame Sky · 10k+ views

8 minHow Do Parking Sensors Detect Obstacles?History of Simple Things · 10k+ views

Ultrasonic Sensing Basics (Texas Instruments)

Cameras deliver more information per dollar than any other sensor. Smartphone volume turned CMOS image sensors into $2–20 components that deliver megapixels of texture, color, and semantic detail no other modality approaches. A camera reads signs, distinguishes a plastic bag from a rock, and recognizes a specific person. It measures geometry only by inference, through learned depth, structure-from-motion, or known-object priors. Most of modern robotics' deep-learning work happens on camera data.

Semantic richness is unmatched, the sensors are tiny, cheap, and passive, the ecosystem of models and tooling is enormous, and it's the sensor of choice for learned end-to-end policies (and for imitation of human video). The weaknesses: there's no direct depth, and performance depends entirely on illumination and dynamic range. The interpretation compute is the true cost, since a $10 camera can demand a $1,000 GPU. Metric scale is unobservable without motion, stereo, or priors.

Put cameras on essentially everything. They are the only sensor that delivers semantics (reading labels, classifying objects, feeding learned policies), and at $2–20 the sensor cost is never the objection. Choose camera-primary stacks when the task is recognition-heavy, the BOM is tight, and you can fund the real cost, which is GPU compute and data pipelines. Avoid relying on monocular vision as the sole geometric sensor where wrong depth is a safety event: night operation, out-of-distribution scenes, metric-critical manipulation. In those cases add stereo, ToF, or LiDAR for measured geometry, unless you are deliberately making the camera-only model-progress bet with eyes open.

Resolution typically 1–12 MP · frame rate 30–60 fps, higher on global-shutter industrial parts · dynamic range around 70 dB standard and up to 120 dB on HDR automotive sensors · field of view 60–120° depending on lens · sensor power draw roughly 0.5–2 W · unit cost $2–20, with interpretation compute 50–100× the sensor price.

Glare, low sun, and tunnel exits swing the dynamic range past what the sensor can hold. Darkness defeats the camera unless you add illumination. Motion blur shows up at control-relevant speeds. Lens contamination is the number-one issue in fleets. Vibration produces rolling-shutter distortion. Learned depth hallucinates on out-of-distribution scenes, and confidently wrong geometry is the signature failure of this modality. Adversarial and spoofing edge cases are also a concern.

Tesla's eight-camera Autopilot suite (the camera-only flagship), every drone (DJI), warehouse AMR fiducial navigation, Sony IMX sensors as the de facto standard silicon, OAK/Luxonis and Basler in industrial vision, VLA-driven manipulation research running on plain webcams.

Sensor cost is a rounding error; compute and data pipelines are the spend. The camera-only versus fusion argument is at heart a bet that model progress outruns the cost of the sensors it replaces. Even fusion stacks are camera-first for semantics, so cameras are central either way. The open question is whether they're sufficient on their own.

9 minImage Sensors Explained: How CCD and CMOS Sensors works? CCD vs CMOSALL ABOUT ELECTRONICS · 100k+ views

11 minTypes of Image Sensors | Image SensingFirst Principles of Computer Vision · 100k+ views

7 minWhy Do Cameras Do This? | Rolling Shutter Explained - Smarter Every Day 172SmarterEveryDay · 1m+ views

Understanding Camera Sensors for Machine Vision (Edmund Optics) · Computer Vision: Algorithms and Applications (Szeliski)

Stereo cameras get depth the way biology does. Two cameras sit a known baseline apart, and matching pixels between the two views triangulates distance. Modern modules add an infrared texture projector so blank walls still match, and run the correspondence search in on-board ASICs, which delivers registered RGB-D streams over USB. Depth precision falls with distance squared and grows with baseline, and that physics sets each product's sweet spot and its ceiling.

It can work passively, so it holds up in sunlight where ToF struggles. Depth comes out dense and aligned to color, there are no moving parts, and commercial modules are well packaged with mature SDKs. The weaknesses: depth error grows quadratically with range, so centimeter accuracy at 1 m becomes decimeters by 5 m. Textureless and repetitive surfaces defeat matching unless you project a pattern, object edges leave occlusion shadows, calibration drifts if the rigid baseline flexes, and the compute is real (though increasingly on-module).

Pick stereo for outdoor depth in sunlight at roughly 0.5–15 m (drones, sidewalk and field robots, outdoor manipulation), where active-IR ToF washes out and LiDAR is overkill on cost or weight. Modern modules with on-board matching and projectors make it a $200–500 bolt-on that gives you registered RGB-D. Match baseline to working range, since precision falls off quadratically with distance. Avoid it past its baseline's sweet spot, on textureless indoor scenes without projection, and for short-range indoor grasping, where a ToF module is cheaper, denser, and lighter on compute. If you need real range outdoors, step up to LiDAR.

Working range roughly 0.5–15 m · depth error roughly 1% of range and quadratic with distance (a centimeter at 1 m, decimeters at 5 m) · baseline 50–120 mm on common modules · depth resolution 720p–1080p at 30–90 fps · power draw roughly 2–3 W over USB · unit cost $200–500.

Texture-poor surfaces such as white walls and shiny tables yield holes or garbage without the projector, and the projector itself washes out in direct sunlight. Repetitive patterns like fences and gratings cause false matches that report wrong depth with high confidence. Thermal or mechanical flex of the baseline mis-scales the world until you recalibrate, with no obvious symptom in the meantime. Specular and transparent objects break the Lambertian assumption that the whole method depends on.

Intel RealSense D400 series (the research default, surviving Intel's wind-down scare), Stereolabs ZED (outdoor/long baseline), Luxonis OAK-D (with on-board neural compute), Skydio's navigation cameras (the strongest example of stereo-only autonomous flight), Subaru EyeSight (automotive stereo's persistence).

$200–500 modules made stereo the default lab and startup depth sensor. Its commercial position is squeezed from two sides: ToF modules from underneath (cheaper, better indoors, phone volume) and learned monocular depth from above (free with the camera you already have). Sunlight performance and metric reliability keep stereo the outdoor workhorse. Skydio-class results show how far the modality stretches with enough software.

13 minDepth Camera - ComputerphileComputerphile · 100k+ views

17 minSimple Stereo | Camera CalibrationFirst Principles of Computer Vision · 100k+ views

3 minClumsy Cyclops- How does Stereo Vision work?i3D robotics · 10k+ views

Tuning Depth Cameras for Best Performance (RealSense) · Intel RealSense Stereoscopic Depth Cameras (arXiv)

Both of these measure depth at every pixel using their own light source. Time-of-flight cameras modulate infrared light and measure the phase shift at every pixel simultaneously; structured-light units project a known dot pattern and decode how it deforms. Both deliver dense depth indoors without much compute. The Kinect (structured light, 2010) accidentally launched a robotics research decade, and phone Face-ID modules industrialized the components. ToF has largely won the newer designs on range and robustness.

Depth comes out dense with minimal compute, because the sensor does the work. Short-range precision is excellent for grasping and bin-picking, the modules are small and cheap thanks to phone volume, and they work in total darkness. The weaknesses: active IR loses to sunlight outdoors, and range is limited by illumination power (typically 0.2–10 m). Multipath within a scene (corners, shiny bins) bends measurements, there are frame-rate and resolution trade-offs, and multiple units interfere unless you synchronize them.

Pick this family as the first answer for indoor depth at 0.2–5 m. Manipulation, bin-picking, and AMR obstacle sensing get dense, low-compute depth that works in darkness, and phone-volume ToF modules keep the price in stereo's range without stereo's matching compute. Step up to industrial structured-light scanners (Photoneo, Zivid class, $5–15k) when bin-picking needs sub-millimeter registration. Avoid it outdoors, since direct sunlight collapses performance, and distrust it on black foam, transparent packaging, and shiny bins. For the outdoor cases use stereo, and for the optically hostile ones you'll need added modalities or polarization tricks.

Working range 0.2–10 m, with 0.2–5 m the usual sweet spot · depth accuracy roughly 1% of range, sub-millimeter on industrial structured-light scanners · depth resolution VGA to 1 MP at 15–30 fps · field of view 70–120° · power draw roughly 2–5 W, most of it the IR illuminator · unit cost a few hundred dollars for phone-lineage modules, $5–15k for industrial scanners.

Direct sunlight swamps the return, so outdoor performance collapses. Dark and IR-absorbing materials such as black foam and some fabrics go invisible. Specular and transparent surfaces either mis-range or drop out entirely. Multipath in concave corners reports walls closer than they are. Cameras interfere with each other in multi-robot cells. Temperature drift shifts the phase calibration.

Microsoft Azure Kinect and its industrial successors, orbbec's Kinect-lineage sensors (now the volume leader), PMD/Infineon ToF in phones and robots, Photoneo and Zivid structured-light 3D scanners for industrial bin-picking (the high-accuracy end), iPhone Face ID and LiDAR-branded rear ToF.

Phone and automotive in-cabin volume made ToF chips a commodity. Industrial-grade structured-light scanners (Photoneo, Zivid at $5–15k) hold the precision bin-picking niche with sub-millimeter accuracy. For indoor manipulation this family is usually the right first answer, because the compute savings versus stereo pay for the sensor. Outdoors, it isn't in the running.

2 minHow the Kinect Depth Sensor Works in 2 MinutesCuriousInventor · 100k+ views

4 minWHAT IS ToF? All About Time-of-Flight and the FRAMOS ToF DevkitFRAMOS · 10k+ views

4 minWhat is Structured-Light 3D Scanning?Holocreators GmbH · 10k+ views

Time-of-Flight Camera: An Introduction (Texas Instruments) · How Structured Light Works (Zivid)

Event cameras don't produce frames. Each pixel independently fires an event the microsecond its brightness changes, so the output is a sparse asynchronous stream instead of images. That gives microsecond latency, ~120 dB dynamic range (sunlight and shadow in one view), no motion blur, and milliwatt power, which sounds like exactly what a fast robot needs. The catch is that thirty years of computer vision assumes frames, so event streams require rethinking everything from features to networks.

Temporal resolution and latency are measured in microseconds, dynamic range is extreme, there's no blur at any speed, and power and bandwidth stay low in static scenes. The weaknesses: there's no absolute intensity, so a stationary camera watching a stationary scene sees nothing. Algorithms and tooling remain research-grade, resolution and cost trail conventional sensors, texture and semantics are weak without a companion frame camera, and the killer app has stayed perpetually adjacent.

Pick an event camera only when a conventional sensor's physics is the binding constraint: microsecond latency for aggressive flight or high-speed tracking, 120 dB dynamic range for welding-arc or tunnel-mouth scenes, or milliwatt always-on sensing in wearables and IoT. Even then, pick it only if the team can carry research-grade tooling and train on thin datasets. Avoid it for mainstream perception. A global-shutter frame camera at a few hundred fps solves most "too fast" problems with mature tooling, and hybrid frame-plus-event chips are the sensible middle path when you genuinely need both.

Temporal resolution about 1 µs, with sub-millisecond latency · dynamic range around 120 dB · resolution 320×320 to 1280×720 on current sensors · sensor power in the milliwatts on static scenes, rising with event rate · unit cost a few thousand dollars for a development camera, less for the bare sensor.

Static scenes are invisible until something moves, though hybrid frame+event chips address this. Flickering lights such as LEDs and fluorescents flood the stream with useless events. Low light adds event noise. Per-pixel threshold mismatch requires calibration. Downstream algorithms trained on limited event datasets generalize unpredictably, so the ecosystem is a failure mode in its own right.

iniVation DVS and Prophesee-Sony IMX636/GenX320 (the commercial leaders), high-speed tracking and vibration-monitoring industrial pilots, drone obstacle-avoidance research (UZH's aggressive-flight demos), Samsung's SmartThings motion sensing, automotive in-cabin monitoring evaluations.

Sony's fab partnership with Prophesee signaled real intent, but a decade of "emerging" status reflects the missing volume driver: every proposed use has an adequate conventional alternative. The plausible break-in points are always-on low-power vision (wearables, IoT) and hybrid sensors where events augment frames. Price follows volume, and volume is still waiting on the app.

4 minWhat is Event-Based Vision | Metavision by PropheseePROPHESEE Metavision Technologies · 10k+ views

3 minEvent-based Vision Sensor (EVS) - Detecting Changes in Moving Subjects [Technology] | Sony OfficialSony · 10k+ views

3 minRapid, Dynamic Obstacle Avoidance with an Event-based CameraUZH Robotics and Perception Group · 100k+ views

Event-Based Metavision Sensors (Prophesee) · Event-Based Vision: A Survey (arXiv)

Thermal cameras see heat instead of light. Microbolometer arrays image long-wave infrared (8–14 µm), which everything above absolute zero emits, so they need no illumination at all. A pedestrian glows against a cold road at midnight, in fog, or through smoke, all conditions that defeat every visible-light sensor. Uncooled microbolometers made the capability affordable, and export regulations (thermal imaging is dual-use) still shape which resolutions and frame rates you can buy.

It sees warm-blooded life and machinery in zero light, smoke, dust, and most fog, it's passive and unjammable, and it doubles as a temperature-measurement instrument (overheating motors, battery fires). The weaknesses: resolution is low and the germanium optics are expensive. There's no texture or color to recognize things by, glass is opaque, and polished metal acts as a thermal mirror. Scene contrast collapses when everything equilibrates (rain-soaked scenes, 37 °C air), and frame rates are export-capped on many parts.

Pick thermal when the job is finding warm things where light fails: pedestrian detection at night, search-and-rescue and firefighting drones, security patrol, and condition monitoring of motors, substations, and battery installations. A ~$200 Lepton-class module covers detection, while recognition-grade work needs $1k+ resolution and germanium optics. It is also the only camera that doubles as a thermometer. Avoid it as a general perception sensor, since it gives no color or texture, glass is a wall, and dawn/dusk crossover erases contrast. Pair it with a visible camera rather than replacing one, and check export caps on resolution and frame rate before designing it in.

Wavelength 8–14 µm · resolution 160×120 on entry modules up to 640×512 and above · thermal sensitivity (NETD) typically under 50 mK · frame rate 30–60 Hz, or 9 Hz on export-capped parts · power draw roughly 0.15–1.5 W · unit cost about $200 for a Lepton-class module, $1,000 and up for recognition-grade resolution with germanium optics.

Thermal crossover at dawn and dusk erases the contrast between targets and background. Reflective metal surfaces show the sky's temperature rather than their own. Glass blocks the view entirely. Sun-heated clutter creates false positives for any "warm object" logic. Uniformity drift requires periodic shutter-based recalibration, which blinds the camera for a moment each time. Low resolution merges adjacent objects at range.

FLIR (Teledyne) Lepton/Boson modules, the dominant supplier; automotive night vision (Autoliv/Veoneer systems in premium cars), AEB-pedestrian research after high-profile nighttime failures, firefighting and SAR drones (DJI thermal payloads), battery-plant and substation monitoring robots.

Leptons brought entry thermal to ~$200, but automotive-grade resolution is still a $1k+ story driven by germanium optics and regulation. A genuine volume driver (NHTSA's pedestrian-AEB night requirements pushing thermal into mainstream cars) could finally bring those costs down. In robotics it is a specialist worth carrying wherever robots share darkness with living things.

4 minAsk an Expert: What is a Microbolometer?Sierra-Olympia Technologies · 10k+ views

15 min#1486 Micro Bolometer Thermal Camera SensorIMSAI Guy · 5k+ views

3 minHow Do Thermal Imaging Goggles Work?SciShow · 100k+ views

How Do Thermal Cameras Work? (Teledyne FLIR) · Choosing the Right Thermal Imager (Teledyne FLIR OEM)

An IMU is the robot's inner ear. Micromachined accelerometers and gyroscopes on a chip measure specific force and angular rate at hundreds to thousands of hertz, for under a dollar at phone volume. It's the only sensor that works everywhere and always (no line of sight, no lighting, no infrastructure), which makes it the backbone of every state estimator. Its flaw is equally universal: it measures derivatives, so position comes from double integration, and every bias error grows quadratically into drift.

It's tiny, cheap, kilohertz-fast, and works in any environment. It's the glue of every sensor-fusion stack (VIO, GNSS-INS, leg odometry) and the flight-control heartbeat of every drone. The weaknesses: on its own it drifts without bound, and consumer parts lose meters within seconds of dead reckoning. Bias wanders with temperature and time, vibration aliases into the signal (drone props are especially bad), and scale-factor and axis-misalignment errors need calibration. Grades are stratified across four orders of magnitude in price.

Every robot gets one, so the real decision is grade. Consumer parts (cents to tens of dollars) are correct whenever external fixes arrive every second or faster, as they do with VIO, GNSS-INS, or encoder odometry, which covers nearly all of robotics. Spend on tactical-grade ADIS-class parts ($1k–20k) only when dead-reckoning gaps stretch to tens of seconds or vibration is severe. Isolate it mechanically from props and motors before buying a better one. Don't trust any MEMS IMU standalone for position beyond a few seconds, and don't jump to FOG prices before checking whether better fusion or learned inertial odometry closes the gap in software.

Update rate 100 Hz to several kHz · gyro bias instability roughly 10–100 °/hr on consumer parts, 0.1–3 °/hr on tactical grades · unaided position error into the meters within seconds of dead reckoning · package around 3 × 3 mm, power in the milliwatts · unit cost cents to tens of dollars for consumer parts, $1,000–20,000 for tactical grade.

Vibration rectification biases estimates on any platform with motors, and it does so without any obvious symptom. Temperature transients swing bias faster than estimators can track it. Saturation during impacts corrupts everything downstream. Magnetometer-aided headings, the usual fix for yaw, break near steel and current. Estimator divergence after GPS or vision dropout is really the IMU's drift being unmasked.

Bosch/TDK-InvenSense/STMicro consumer parts in every phone and drone flight controller; tactical-grade MEMS (Honeywell HG series, Analog Devices ADIS) in robots and munitions at $1k–20k; VectorNav and Xsens integrated AHRS units as robotics staples.

Consumer silicon costs cents, and each 10× improvement in bias instability costs roughly 10× more, up through export-controlled navigation grades. The interesting frontier is software substituting for grade: learned IMU odometry and better fusion extract tactical-grade performance from consumer parts, which compresses the pyramid from below.

10 minHow MEMS Accelerometer Gyroscope Magnetometer Work & Arduino TutorialHow To Mechatronics · 500k+ views

6 minHow does an Accelerometer work? | 3D AnimationCircuitBread · 100k+ views

12 minIntro to Inertial Measurement Units (IMU)MicWro Engr · 10k+ views

Inertial Navigation Primer (VectorNav) · Using Inertial Sensors for Position and Orientation Estimation (arXiv)

Fiber-optic and ring-laser gyroscopes measure rotation using the Sagnac effect (counter-propagating light in a loop). They have no moving parts, and their bias instability is roughly a thousand times lower than MEMS. A navigation-grade INS can dead-reckon a submarine for hours or an aircraft across an ocean, drifting a nautical mile per hour or less. This is the sensor class that still works when GPS is jammed, denied, or underwater, which is why it is export-controlled and increasingly relevant.

Bias instability of 0.001–0.01°/hr is good enough for genuine long-duration dead reckoning and for gyrocompassing (finding north from Earth's rotation alone). The sensors are immune to jamming, spoofing, weather, and darkness, and they have decades of proven reliability. The weaknesses are the $20k–$1M+ price range, a mass and power budget measured in kilograms and watts rather than grams, and ITAR/export licensing that complicates everything. Performance is bought with fiber length or cavity precision, so costs don't fall the way consumer-electronics costs do.

Pick FOG/RLG when GNSS-denied dead reckoning for minutes to hours is a hard requirement (subsea and underground vehicles, defense platforms, operations near jamming), or when you need gyrocompassing and magnetometers are useless. The platform also has to be able to carry kilograms and a $20k–$1M line item. Factor export licensing into the schedule from day one. If external fixes (GNSS, vision, UWB, ranging) are reliably available, use tactical MEMS plus good fusion instead: it's 10–100× cheaper and covers outage gaps of seconds to minutes, which is most commercial autonomy.

Gyro bias instability 0.001–0.01 °/hr · navigation-grade drift about 1 nautical mile per hour unaided · gyrocompassing in minutes, with no external reference · mass in kilograms and power in the tens of watts · unit cost $20,000 to over $1M, and export-controlled.

Even navigation grade drifts eventually. It postpones the need for external fixes rather than eliminating it. Temperature gradients across the fiber coil induce bias, so careful thermal design is half the product. Shock beyond spec misaligns the optics. RLG dither mechanisms wear out. At the program level, the usual failure is procurement and export delay.

Honeywell HG9900 and Safran/iXblue fiber INS in defense platforms, iXblue Phins on subsea vehicles, KVH FOGs on autonomous trucks and mining robots, Boreas quantum-adjacent gyrocompasses; every airliner's IRS.

This is a defense-electronics oligopoly with stable, high margins, and demand is rising as GPS-denied operation goes from exotic to assumed (drones over contested territory). Two technologies could disrupt it: photonic-chip gyros and quantum interferometry both promise FOG performance at silicon prices, and both have been promising it for a while. The commercial pull worth watching is autonomous trucking's appetite for affordable 0.1°/hr parts.

6 minFiber Optical & Ring Laser Gyro working principle. What is Sagnac effect ?Aviation.Animation · 10k+ views

5 miniXblue insights - Fiber-Optic GyroscopesiXblue · 50k+ views

11 minThe GENIUS of Inertial Navigation Systems ExplainedFlyByMax · 1m+ views

Inertial Navigation Primer (VectorNav) · What Is Fiber-Optic Gyroscope Technology? (Exail)

GNSS gives you absolute position from orbit. Multi-constellation receivers (GPS, GLONASS, Galileo, BeiDou) deliver meters of accuracy anywhere on Earth for a few dollars, and RTK sharpens that to 1–2 centimeters by differencing carrier-phase measurements against a base station or network. RTK prices have collapsed (u-blox F9P-class modules run about $200), which turned centimeter positioning from survey equipment into a robotics commodity and quietly enabled the entire outdoor autonomy industry, from mowers to tractors.

This is the only sensor that gives drift-free absolute position, and RTK gets that to centimeter accuracy. It works globally, it's passive, it supports unlimited users, and it fuses easily with IMUs. The weaknesses start with needing sky view: urban canyons, tree canopy, and indoor operation degrade or kill it. RTK also needs a correction link and either a nearby base station or a network subscription, and convergence and re-acquisition take seconds after an outage. Jamming and spoofing are routine in several regions now, not theoretical.

Default to GNSS on any outdoor robot with sky view. Meter-grade receivers cost a few dollars, and $200 F9P-class RTK delivers 1–2 cm for mowers, tractors, survey drones, and construction robots wherever a correction link and a base or network exist. Always fuse with an IMU and wheel or visual odometry so outages degrade gracefully. Don't depend on it in urban canyons, under canopy, indoors, or anywhere jamming and spoofing are plausible. Design the denial case first, with LiDAR or visual localization carrying the robot through, because GNSS usually fails by handing you a confidently wrong fix rather than no fix at all.

Standalone accuracy 1–3 m, RTK 1–2 cm · update rate 1–20 Hz · RTK convergence and re-acquisition in seconds · usable baseline to a base station roughly 10–30 km · receiver power draw 0.1–0.5 W · unit cost a few dollars for meter-grade and about $200 for F9P-class RTK, plus a correction subscription.

Multipath off buildings yields confidently wrong fixes, which is the worst kind. RTK integer-ambiguity slips jump position by decimeters. Correction-link dropout degrades accuracy from centimeters to meters mid-task, with no warning. Jamming denies service and spoofing deceives the receiver, and both are increasingly common near conflict zones, ports, and VIPs. Solar storms and constellation faults are rare, but they hit an entire fleet at once. Leap-second and firmware edge cases have bricked receiver fleets before.

u-blox F9P/X20 (the robotics default), Trimble and NovAtel in agriculture and survey, John Deere StarFire (PPP at continental scale), drone RTK (DJI), Swift Navigation and Point One correction networks, every phone's meter-grade chip.

Receiver silicon is a commodity, so the business moved to corrections-as-a-service (subscription networks) and tight GNSS-INS integration. For robotics diligence, the thing to look at is denial-tolerance. Everyone assumes the sensor is there, so any outdoor autonomy plan gets judged by what happens when GNSS gives a bad fix or disappears.

5 minHow RTK works | Real-Time Kinematic for Precise GNSS PositioningEmlid · 100k+ views

4 minHow GPS Works | How GNSS Receivers Determine Your Location Using Satellite SignalsEmlid · 50k+ views

13 minAdventures in Science: How GPS WorksSparkFun Electronics · 50k+ views

What Is RTK? (Point One Navigation) · RTK Fundamentals (ESA Navipedia)

UWB is essentially GPS for indoors, built out of installed infrastructure. Ultra-wideband radios time nanosecond pulses between tags and fixed anchors, which gives 10–30 cm positioning through walls and clutter where GNSS can't reach. Phone adoption (Apple U1/U2, Android peers) commoditized the silicon. The broader family also includes visual fiducials (AprilTags/QR floors), BLE beacons (meters-grade), and magnetic-signature maps. All of them trade an infrastructure install for localization that pure SLAM would otherwise have to achieve with compute.

UWB gives absolute indoor position with bounded error and no drift, works through non-metal walls and in darkness, uses cheap tags, and needs little compute. It's a good fit for multi-robot coordination and asset tracking. The main weakness is the infrastructure tax: someone has to install, survey, and power the anchors. Metal-heavy environments produce bad multipath, there are capacity limits once you have many tags, and it competes with "free" SLAM that improves every year, which squeezes the niche from the software side.

Pick UWB when you control the facility and 10–30 cm absolute indoor position with zero drift is worth an anchor install: multi-robot coordination, forklift and asset tracking, indoor drone work. It also works as a deliberate bridge, where you ship reliable autonomy with anchors now and remove them as SLAM matures. Tags are $5-class, so the survey and anchor maintenance are the real cost. Skip it in metal-dense environments, where multipath bias eats the accuracy advantage, and in third-party spaces where you can't install infrastructure. If your SLAM already meets the accuracy bar, use that instead and save the install.

Position accuracy 10–30 cm · anchor-to-tag range roughly 10–50 m indoors · update rate 1–100 Hz, trading against how many tags share the airtime · band 3.1–10.6 GHz · tag cost around $5, anchors in the low hundreds of dollars each plus survey and install.

Non-line-of-sight paths cause bias: signals through bodies and through shelving arrive late, producing position errors of decimeters that still look statistically healthy. Anchor geometry degrades accuracy in corridors through dilution of precision. If an anchor drifts or gets moved, it corrupts the map without any obvious symptom. The UWB bands are getting crowded, so interference is a growing problem. Day to day, most of the trouble is dead tag batteries and anchor moves nobody surveyed.

Qorvo/Decawave DW3000 modules (the standard), Apple AirTag's finding network, warehouse forklift tracking (Zebra, Sewio), robot soccer and drone-swarm labs, Amazon's fulfillment-center positioning, automotive keyless entry (UWB's actual volume driver).

Phone and car-key volume made the chips a $5 part, so the business is systems and installation, which caps margins and scale. Strategically, UWB is a pragmatic bridge: teams ship reliable autonomy with anchors now, then remove them as SLAM matures. That sequencing is worth taking seriously in any warehouse-robotics diligence.

3 minHow UWB Ranging Works: TOF, TDOA, and Angle of Arrival ExplainedNXP Developer Zone · 10k+ views

2 minWhat Is Ultra-Wideband (UWB)? | Accurate, Secure Wireless Positioning ExplainedNXP Developer Zone · 10k+ views

5 minYour iPhone’s New 6th Sense – U1 & UWB ExplainedTechquickie · 100k+ views

How UWB Works (FiRa Consortium) · UWB TDOA Indoor Localization: From Sensor Placement to System Evaluation (arXiv)

Rotary encoders are optical or magnetic discs that report shaft angle, either incrementally (pulses counted from a reference) or absolutely (position known at power-on). Every servo loop in robotics closes around one. Resolution spans 12 bits (magnetic chips, cents) to 23+ bits (optical ring encoders, thousands of dollars). That choice sets a joint's precision ceiling, its torque-control quality (velocity comes from differentiating position), and its homing behavior at startup.

Encoders give direct, drift-free position at kilohertz rates with almost no processing. Magnetic chip encoders (AS5047-class) made 14-bit absolute feedback nearly free, and multiturn absolute types remove the need for homing routines. On the weakness side, incremental types lose position at power-off and miscount when there's noise. Resolution and accuracy are different specs: interpolated magnetic encoders wobble within a period. Mounting eccentricity usually dominates real-world accuracy, and through-shaft optical rings are easily degraded by contamination.

Every powered joint gets one, so the decision is which type. Magnetic absolute chips (AS5047-class, ~$5, 14-bit) are the default for cost-driven joints and QDD actuators. Optical rings (Renishaw, Heidenhain) buy the arcsecond accuracy that machine-tool-grade axes need. Multiturn absolute types are worth it wherever a homing routine at power-on is operationally unacceptable. If the joint has a high-torque transmission with meaningful compliance, spec dual encoders (motor side plus output side), which is increasingly the humanoid standard. Don't use exposed optical discs in dirty, oily, or wet environments; go magnetic or inductive instead. And avoid incremental-only feedback in fleet products, where a missed index turns into a field service call.

Resolution 12 bits on magnetic chips to 23+ bits on optical rings · accuracy roughly 0.1–0.5° for magnetic parts, arcseconds for precision optical · output rate in the kilohertz, fast enough to close a servo loop · power draw in the tens of milliwatts for chip encoders · unit cost about $5 for a 14-bit magnetic absolute chip, thousands for an optical ring.

Dust, oil, or condensation on optical discs drops counts with no warning, which is the usual cause of a joint that slowly drifts. Magnet misalignment or axial play corrupts magnetic readings. EMI on incremental lines injects phantom counts. Index-pulse glitches shift homing by one revolution. Connector failures from vibration look like controller faults. Differentiating position gives noisy velocity exactly when torque control needs it clean.

AMS/ams-OSRAM magnetic chips inside nearly every QDD robot joint, Heidenhain and Renishaw optical rings in machine tools and precision arms, US Digital and CUI in research robots, Nidec/Tamagawa resolvers-and-encoders in industrial servo motors by the million.

This is a mature component market, and the interesting part is how far absolute-encoder pricing has fallen: $5 chips now do what $500 units did, which makes cheap torque-controlled joints practical. Premium optical and inductive precision (Renishaw, Heidenhain) still holds machine-tool margins. Dual-encoder joints (motor side plus output side, to catch transmission error) are becoming the humanoid-grade standard, and they double the number of sockets per robot.

10 minWhat is the Difference between Absolute and Incremental Encoders?RealPars · 500k+ views

5 minOptical EncodersWisc-Online (Part of WisTech Open) · 10k+ views

2 minHow does an encoder work? | Encoders 101US Digital Encoders · 50k+ views

How Renishaw Optical Encoders Work (Renishaw) · Resolution, Accuracy, and Precision of Encoders (US Digital)

Resolvers are wound transformers whose coupling varies with shaft angle. There are no electronics, optics, or semiconductors at the sensing point, so they survive temperatures, radiation, and vibration that would kill an encoder. Modern inductive (eddy-current) encoders print the same physics onto PCBs, which gives you resolver robustness at encoder convenience and cost. They are quietly taking over EV traction motors and harsh-duty robotics.

The sensing element is close to indestructible; resolvers fly in engines, reactors, and space. They're absolute within a revolution and immune to the contamination that blinds optical discs, and the inductive versions are thin, hollow-shaft, and cheap. The weaknesses: resolvers need excitation and demodulation electronics (the R/D converter is the real cost) and their accuracy is modest, classic resolvers are heavy copper, and inductive encoders come from a small set of patent-holding vendors. Accuracy trails precision optical by about an order of magnitude.

Pick resolvers when the environment kills electronics at the sensing point: traction motors, engine and reactor mechanisms, aerospace actuators, or anywhere temperature, radiation, or vibration exceed encoder ratings. Pick inductive encoders for the growing middle ground of thin hollow-shaft joints, gimbals, and motor housings full of magnetic interference where magnetic chips misbehave. That's why they're becoming the default output-side encoder in humanoid joints. If accuracy is the whole point, use precision optical instead (it's about an order of magnitude better), and if a $5 magnetic chip in a benign environment does the job, use that; otherwise you're paying for robustness you don't need.

Absolute within one revolution · resolution 12–16 bits after resolver-to-digital conversion · accuracy roughly 0.1–0.5°, about an order of magnitude behind precision optical · excitation 5–10 kHz on classic resolvers · operating temperature past 200 °C, plus radiation and vibration tolerance · unit cost typically in the hundreds of dollars once the R/D electronics are counted.

The sensor rarely dies; the wiring does. In the field, most failures are excitation shorts, connector corrosion, and cable EMI. R/D converter faults show up as position noise. Inductive types err near large conductive masses and with target-gap runout. Temperature drift in the electronics, not the sensor, sets practical accuracy. Miswired phasing yields angles that are wrong but plausible enough to pass a casual check.

Resolvers in EV traction motors (billions of km proven), aerospace actuators and reactor mechanisms; inductive: Renishaw's induction lines, Zettlex/Celera IncOders in robot joints and gimbals, integrated inductive chips (Microchip, TI) spreading into cobots and humanoid joints.

Resolvers are a stable legacy market. The growth is in PCB-based inductive encoding, where EV volume is driving chipset costs toward magnetic-encoder levels while giving better immunity to the magnetic interference inside motor housings. For humanoid joints packed with magnets, that immunity is turning inductive into the default output-side encoder, so the chipset vendors are the ones to watch.

3 minHow does a Resolver work? - Technical animationlearnchannel · 100k+ views

6 minIntroducing the ResolverElektor TV · 10k+ views

9 minHow to Design an Inductive Position SensorRenesasPresents · 1k+ views

Resolvers: What They Are and How They Work (Dynapar) · IncOder Inductive Ring Encoders (Novanta / Celera Motion)

A linear encoder is a graduated scale (glass, steel tape, or magnetic stripe) fixed along the axis, plus a read head on the moving part. It reports true load position instead of inferring it from motor turns through a compliant, backlashed drivetrain. That difference between direct and inferred measurement is the whole value, and machine tools sell it as "closed-loop glass scales" and charge accordingly. Resolution runs from microns (magnetic tape) to nanometers (interferential glass).

The measurement leaves out drivetrain error (backlash, screw pitch error, thermal growth). Nanometer-class options exist for stages, and tape versions can be cut to any length on site. The weaknesses: the scale has to be mounted, aligned, and protected along the full travel, so installation is most of the work. Exposed scales collect shop contamination, cost scales with length, and most robot arms can't use them at all because their joints are rotary. That confines linear scales to gantries, stages, and machine axes.

Pick a linear scale when a linear axis has to hit tolerances the drivetrain can't guarantee: gantries, stages, and machine axes chasing tens of microns or better, where backlash, screw error, and thermal growth would otherwise show up in the part. Use glass scales for precision work, cut-to-length magnetic tape for long or dirty axes, and interferential scales for nanometer stages. If a motor-side encoder plus a stiff, well-mapped drivetrain already meets spec, stay with that, because installation, alignment, and full-travel protection are the real cost. And keep in mind that none of this applies to rotary-joint arms, which is most robots.

Resolution 1–5 µm on magnetic tape down to nanometers on interferential glass · accuracy roughly ±10 µm per meter on tape and ±1 to ±5 µm per meter on glass · travel from a few hundred millimeters to tens of meters with cut-to-length tape · traverse speed up to a few meters per second on sealed scales · unit cost in the hundreds of dollars for tape and $1,000–5,000 for a sealed glass scale, rising with length.

Contamination scratches or obscures the gratings, and sealed scales trade protection for friction. Mounting stress and thermal mismatch between the scale and the machine bend accuracy. The read-head gap drifts from wear or after a crash. Strong fields or physical dings corrupt magnetic scales. Incremental types lose their reference mark after a power cycle. Cable carriers fatigue on fast axes.

Heidenhain glass scales on every serious CNC and CMM, Renishaw tape systems on gantry robots and large-format machines, magnetic tape (SIKO, Balluff) on presses and actuators, interferential scales inside wafer steppers and metrology stages.

The market is close to a duopoly (Heidenhain, Renishaw), with instrument margins at the top and commodity magnetic tape at the bottom. Robotics exposure is indirect but real, since every gantry-based logistics system, quality 3D printer, and semiconductor handler carries them. The strategic point is that as robots chase machine-tool accuracy for machining and assembly, direct measurement gets popular again.

4 minLinear Encoders for Machine Tools | HEIDENHAINHEIDENHAIN · 100k+ views

8 minWhat is an ENCODER - Linear and Rotary encoder - Incremental and Absolute encoderJAES Company · 100k+ views

2 minM-DRO Guide to the Operation of Optical Linear Encoders, also known as Glass Scale Digital ReadoutsMachineDRO · 10k+ views

Linear Encoders Overview (HEIDENHAIN) · Linear Encoders for Numerically Controlled Machine Tools (HEIDENHAIN)

A six-axis F/T sensor is a machined flexure instrumented with strain gauges (or capacitive/optical gaps) that resolves the full contact wrench, three forces and three torques, at kilohertz rates. Mounted between the arm and the tool, it turns a position-controlled robot into a force-controlled one. That enables polishing at constant pressure, inserting pins that would otherwise jam, teaching by hand-guiding, and detecting collisions by their force signature.

You get complete, calibrated contact measurements right at the tool, which is what makes assembly, finishing, and force-feedback teleoperation work, and the sensors drop straight onto standard wrist flanges. The weaknesses: precision flexures are expensive (traditionally $2–15k), and overload during a crash is a warranty event, so the sensor that protects the process is itself fragile. Thermal drift means you have to re-zero regularly, the added compliance and mass sit right where the dynamics are most sensitive to them, and cheap versions have historically been a false economy.

Pick a wrist F/T sensor when the process needs calibrated wrench data: precision insertion beyond what passive compliance devices handle, polishing and finishing at a specified pressure, teleoperation with force feedback, and product test stands. Sub-$1k entrants have brought the entry price down from the traditional $2–15k. Spec overload protection if the cell can crash, and budget the re-zeroing into cycle plans. If coarse contact detection is enough, use current-based sensing instead (it's free and every cobot ships it). If the arm has built-in joint-torque sensing (iiwa, Franka class) that already meets your fidelity needs, use that. Save the dedicated sensor for tasks where force accuracy visibly pays.

Full-scale range roughly 100–600 N and 5–30 Nm on wrist-sized units · force resolution around 0.1 N · output rate 100 Hz to several kHz · overload limit typically a few times full scale · added mass 100–300 g and 30–50 mm of stack height at the wrist · unit cost sub-$1,000 for new entrants, $2,000–15,000 for the industrial standards.

Collisions overload the sensor, which is the most common way these die. Thermal drift gets misread as force, since strain gauges respond to temperature as well as strain. Cables fatigue at a wrist that moves constantly. Errors in gravity and payload compensation show up as phantom contact forces. Capacitive types are sensitive to temperature and humidity. Slow zero-drift corrupts long tasks without any obvious symptom.

ATI (the industrial standard, now Novanta), Robotiq FT-300, Bota Systems and Wacoh compact units, built-in joint-torque sensing on Kuka iiwa and Franka arms (the integrated alternative), Apple-supplier assembly lines full of them.

This is a profitable niche getting squeezed from two directions: integrated joint-torque sensing (every cobot now ships some version) and current-based sensorless estimation (free but crude). New entrants (Bota, Chinese vendors) have pulled entry prices under $1k. The bull case is humanoids doing contact-rich work, which could mean two wrists' worth of sensors per robot at volume pricing. The bear case is that "good enough" proprioception takes over the mid-market.

3 minATI Force/Torque Sensors used in Force Controlled AssemblyATI Industrial Automation · 10k+ views

2 minProduct Animation: 6-axis force/torque sensor FTSSCHUNK · 1k+ views

20 minThe Ingenious Design of Strain GaugesThe Efficient Engineer · 100k+ views

How Force/Torque Sensors Are Enhancing Automation (ATI Industrial Automation) · What Is a Force Torque Sensor? (Bota Systems)

Tactile skins are grids of pressure-sensitive taxels (capacitive, resistive/piezoresistive, magnetic, or optical) laminated onto fingertips, palms, and robot bodies to report where contact happens and how hard. A wrist F/T sensor gives you one number; a skin gives you a contact image, which shows slip precursors, grasp stability, and incidental collisions on a forearm. Humanoid programs have revived the field, because manipulation in clutter is contact-rich by definition and vision is blocked at the moment of grasp.

Skins give you contact location and pressure distribution that vision can't provide, and they detect slip before the object actually moves. Whole-body versions double as safety sensing. The weaknesses are serious: wiring hundreds of taxels through moving joints is a persistent engineering problem, and durability is worse still, since skins sit at the abrasion interface and wear out there. Calibration drifts with wear and temperature, soft substrates add hysteresis and crosstalk, and no standard product has consolidated the field, so most teams end up building their own.

Pick tactile arrays when the task is decided by information vision loses at the moment of grasp: slip onset on variable objects, contact location in clutter, grasp verification on deformable items. They also make sense for whole-body contact sensing on robots that work around people. Humanoid fingertips are the canonical case, at hundreds of dollars per finger with consumable-replacement economics. If you're running a cost- and uptime-driven industrial cell handling known objects, use a parallel jaw with current feedback or a wrist F/T sensor instead. Those give you the signal you need without a wear surface whose calibration drifts, and without a vendor-fragmented integration project.

Taxel counts from a few dozen on a fingertip to thousands across a body skin · taxel pitch typically 1–5 mm · per-taxel force range roughly 0.1–10 N · update rate 100 Hz to 1 kHz · unit cost in the hundreds of dollars per finger, with the contact surface a consumable.

Contact surfaces wear through and delaminate, so fingertip skins are consumables. Creep and hysteresis in the elastomer layers corrupt force readings. Taxels drop out as flex-cycle wiring fatigues. Temperature sensitivity gets misread as contact. High-impedance sensing lines pick up EMI. Sensitivity drifts gradually, which degrades grasp policies trained on fresh sensors without any obvious symptom.

Tesla Optimus and Figure fingertip arrays, Sanctuary's touch-enabled hands, pressure-mapping incumbents (Tekscan, Pressure Profile Systems), Xela and Contactile robotics skins, iCub's capacitive body skin lineage, BeBop Sensors' fabric-based arrays.

This was a research cottage industry for a long time, and humanoid demand is now pulling it toward real products, starting with fingertips at $100s per finger and consumable-replacement economics (good for vendors, a recurring cost for fleets). The unsolved product question is standardization. Whoever ships the durable, calibrated, easy-to-integrate "tactile module" that hand makers standardize on gets a strategic component slot.

3 minTactile Sensor-Enabled Gripping - How it worksContactile · 10k+ views

2 minHuman-Like Artificial Skin Sensor for Physical Human-Robot InteractionIEEE Spectrum · 5k+ views

3 minRobotSweater : Fabric Tactile Sensor "Skin"CMU Robotics Institute · 5k+ views

How Pressure Mapping Sensors Work (Tekscan) · Tactile Robotics: Past and Future (arXiv)

These sensors measure touch with a camera. A soft, opaque elastomer pad deforms against the object while an internal camera watches that deformation under structured illumination, which recovers contact geometry at optical resolution: surface texture, edges, shear fields, and incipient slip, at a level well past any taxel array. Because the output is an image, the whole deep-learning vision stack applies directly to touch. That's the main reason the approach has dominated manipulation research since GelSight started at MIT.

You get micron-scale contact geometry plus rich shear and slip signals, using commodity cameras and standard CNN tooling, and the sensors can be built from cheap components. The weaknesses: the gel is a consumable that tears, wears, and scuffs its optical coating. The camera needs standoff distance, so fingertips end up thick, which works against dexterity. Frame rate and processing latency both trail dedicated force sensors, each sensor needs its own photometric calibration, and absolute force is inferred from deformation models rather than measured directly.

Pick GelSight-class sensing for dexterity work where fine contact geometry pays: in-hand pose estimation, cable and fastener manipulation, texture and defect inspection by touch, and any learned manipulation stack. Since the output is an image, it plugs straight into vision tooling, and $300–500 modules put it within reach of any lab. Hold off in long-duty industrial cells until gel-tip life matches your maintenance intervals, because tips are consumables and sharp or abrasive parts wear them quickly. If you're closing a tight kilohertz force loop, the frame-rate latency will destabilize control, so use a conventional F/T sensor or taxel array there instead.

Sensing area roughly 15 × 20 mm per fingertip · spatial resolution in the tens of microns, fine enough to read surface texture · frame rate 25–60 fps, well below a dedicated force sensor's kilohertz · fingertip depth 25–40 mm, set by the camera standoff · unit cost $300–500 per module, plus replacement gel pads.

The gel wears and tears on sharp or abrasive parts, so fleet operators stock replacement tips. Scratches in the coating read as permanent phantom features. Aging illumination LEDs shift the calibration. Condensation or oil films alter the optics. Latency destabilizes tight force loops. Models trained on one gel batch degrade on the next, so manufacturing variance is itself a failure mode.

GelSight Inc (Mini as the commercial standard), Meta AI's Digit (open-hardware, low-cost push) and its Digit 360 successor, TacTip (Bristol's pin-based variant), DexIt/9DTact-style academic derivatives, deployment in Amazon and lab bin-picking research.

The $300–500 module price point made rich touch sensing affordable for any lab, and consumable gel tips give vendors a razor-and-blades revenue line. The open question is whether it's ready for industry: research labs use it heavily, but factories need tip life measured in months. If durability improves, vision-based touch could become the fingertip standard for the same reason it took over research, which is that its output is images, the same format the rest of the learned stack already uses.

2 minGelSight sensor gives robots touchMassachusetts Institute of Technology (MIT) · 50k+ views

3 minAdvancing robotics and touch perception | AI Research from Meta FAIRAI at Meta · 100k+ views

9 minDIGIT - a Low-Cost Compact High-Resolution Tactile SensorRoberto Calandra · 5k+ views

Tactile Sensing Technology (GelSight) · Classification of Vision-Based Tactile Sensors: A Review (arXiv)

Every motor already contains a rough force sensor. Torque is proportional to current, so the drive's own current measurement, combined with a dynamics model of the arm, estimates external forces with no added hardware. Quasi-direct-drive joints made this approach credible, because low gearing means external forces actually reach the motor. Every cobot's collision detection, hand-guiding mode, and gentle-contact behavior is some version of this technique.

It costs nothing, adds no mass, and needs no wiring, because the sensing is software. It gets better as dynamics models and learning improve, and it works at every joint at once rather than only at the wrist. The weaknesses: friction, gear losses, and temperature all corrupt the torque-current map, and high-ratio transmissions bury external forces under friction noise completely. Model errors in payload or inertia show up as phantom forces, resolution is about an order of magnitude below a real F/T sensor, and it senses joint torque rather than contact location.

Use it everywhere it comes free: collision detection, hand-guiding, and coarse contact estimation on any arm. On QDD and other low-gear joints, where external loads actually reach the motor, the force estimates are genuinely useful, which is how quadruped touchdown detection and cobot safety stops work at zero BOM cost. Before trusting it beyond that, measure where the friction floor sits for your transmission. Don't make it your only force sense for delicate low-speed contact, since stiction makes the estimates confidently wrong in exactly that regime, and don't rely on it behind high-ratio gearboxes that bury the signal. If the task needs real force numbers or contact location, pay for a wrist F/T sensor or a tactile sensor.

Added BOM cost zero · torque estimates at the 1 kHz servo rate, off a current loop running at 10–40 kHz · force resolution roughly an order of magnitude below a wrist F/T sensor · friction floor a few percent of rated torque on quasi-direct-drive joints, 20–30% behind a high-ratio gearbox · collision-detection latency typically 5–20 ms · coverage at every joint rather than only the wrist.

Friction hysteresis and stiction generate force estimates that are confidently wrong at low speeds, which is exactly where delicate contact happens. Unmodeled payloads bias everything. Temperature-dependent motor constants drift the calibration mid-shift. Gearbox wear changes the friction model over fleet life. Collision detection tuned tight enough to be safe will false-trigger on aggressive motion, and that sensitivity-versus-uptime trade never goes away.

Universal Robots' collision detection and freedrive, Franka's blended joint-torque/current approach, every QDD quadruped's contact estimation (foot touchdown from current spikes), Tesla Optimus and peers leaning on proprioception to minimize sensor count, admittance-control cobot lifting aids.

This is what compresses margins across the force-sensing market, because free and adequate competes hard with expensive and precise. Its ceiling is set by physics, since friction masks the signal, so the market settles into layers: current sensing for safety and coarse contact everywhere, dedicated F/T and tactile sensors where precision pays. When doing diligence on a "sensor-light" robot pitch, ask exactly where the friction floor sits for their transmissions.

7 minMotor Torque, Speed and CurrentAURA Robotics · 10k+ views

17 minMIT Mini Cheetah ActuatorSkyentific · 100k+ views

An Engineer's Guide to Current Sensing (Texas Instruments) · Robot Collisions: A Survey on Detection, Isolation, and Identification (Haddadin et al., IEEE T-RO)