Robot Actuators: A Practical Reference

This is the simplest motor that works. Current flows through a wound rotor, and carbon brushes riding on a copper ring handle commutation mechanically. You get two wires, torque proportional to current, and no drive electronics beyond an H-bridge. Brushed DC did most of the twentieth century's actuation, and it still owns the bottom of the market, where a motor has to cost a dollar and spin as soon as you connect a battery.

It's about as simple and cheap as actuation gets. Torque is proportional to current, no commutation electronics are needed, and reversing or controlling it is easy. On the weakness side, brushes wear out after hundreds to thousands of hours, they arc, and they generate EMI and dust. Heat ends up in the rotor, which is the hardest place to remove it from, so continuous torque is capped. At any performance level that matters, a BLDC now beats brushed DC on every axis except price and simplicity.

Pick brushed DC when the axis is intermittent-duty, cost-dominated, and coarse: a few hundred to a few thousand hours of life is acceptable, the BOM allows single dollars, and control is an H-bridge off a battery. Typical cases are toys, lids, low-end grippers, and prototype drivetrains. The coreless variant is worth its premium at low power when you need zero cogging and smooth motion, as in medical pumps and micro-mechanisms. Avoid brushed DC for continuous duty, tight thermal budgets, or any joint you plan to iterate toward higher performance, because rotor-trapped heat caps sustained torque and brush wear turns into a maintenance item. If you outgrow it, the usual upgrade is a small BLDC with a commodity FOC drive.

Brush life several hundred to a few thousand hours · peak efficiency roughly 50–75% in small sizes, higher on coreless types · no-load speed typically 3,000–15,000 rpm · continuous torque well under the stall rating, since the heat sits in the rotor · unit cost under $1 at volume, $100 and up for precision coreless motors.

Toys, seat adjusters, and appliance actuators by the billion; hobby gearmotors (Pololu-class) in education and prototyping robots; coreless brushed micromotors (Faulhaber, Maxon DCX) surviving in medical devices and precision niches where their zero-cogging smoothness matters.

maxon — premium coreless brushed and BLDC micromotors · FAULHABER — precision micro drive systems

The industry is fully commoditized, with sub-dollar units at volume. The premium coreless niche (Maxon, Faulhaber) holds its margins through quality and documentation. In robotics BOMs the brushed motor is usually the placeholder that gets designed out as requirements harden, though for low-duty, cost-dominated axes it stays the rational choice.

10 minHow does an Electric Motor work? (DC Motor)Jared Owen · 5m+ views

4 minBasics of Brushed DC Motors - Another Teaching Moment | Digi-Key ElectronicsDigiKey · 10k+ views

5 minBrushed vs Brushless Motors - What's the difference?CircuitBread · 100k+ views

Motor Tutorials: Coreless DC Motor and Drive Basics (FAULHABER)

This is the default motor of modern robotics. Permanent magnets sit on the rotor, the stator windings are commutated electronically, and field-oriented control (FOC) delivers smooth, precisely known torque at any speed including stall. Drone ESCs commoditized the electronics, and EVs industrialized the magnets and windings. Outrunner geometries (large-diameter and pancake-thin) maximize torque per mass, which is why they're used in essentially every quasi-direct-drive robot joint.

Efficiency is high (90%+), power density is good, and the only wear parts are the bearings. Torque is known from current, so you get force estimation for free, and the heat is generated in the stator where you can actually remove it. On the weakness side, it needs position sensing and a three-phase drive, so the electronics are part of the actuator rather than an accessory. Continuous torque is thermally limited far below peak, ungeared torque density is modest (which is why transmissions exist), and rare-earth magnets carry China supply-chain exposure.

Make BLDC + FOC your starting assumption for any powered robot joint. It handles continuous duty, runs above 90% efficiency, gives you torque from current for sensorless force estimation, and puts the heat in the stator where you can sink it. Then choose the transmission around it: none for transparency, a single-stage planetary for QDD limbs, a strain-wave gear for precision arms. Step away only at the edges. If the axis is worth under roughly $20 and never runs hot, use a brushed motor or a stepper. If open-loop positioning is good enough, use a stepper. If the spec is in nanometers rather than newton-meters, look at direct-drive, linear, or piezo hardware instead. For a mainstream joint, anything other than BLDC needs a specific reason.

Efficiency above 90% at rated load · continuous torque typically a third to a half of peak · ungeared specific torque usually under 10 Nm/kg on robot-joint outrunners · no-load speed 1,000–10,000 rpm depending on winding · frameless kits $50–150 at volume from Chinese suppliers, $500 and up from Western catalogs.

T-Motor and Maxon EC frameless kits in robot joints; every drone propulsion motor; Tesla Optimus and Unitree joint modules; ODrive/moteus/VESC open-source FOC drives that made hobby-grade force control possible.

T-Motor — high-torque outrunners and robot-joint motors · mjbots — open FOC drives and moteus servo modules

Frameless torque motors that cost $500+ from Western catalogs ship for $50–150 from Chinese suppliers at volume, and humanoid demand is pushing prices down further. The motor itself is becoming a commodity, so the margin is moving to the integrated module (motor + gear + encoder + drive) and to the FOC silicon. Magnet supply is the thing to watch.

16 minBrushless Motor - How they work BLDC ESC PWMThe Engineering Mindset · 5m+ views

9 minWhat is FOC? (Field Oriented Control) And why you should use it! || BLDC MotorGreatScott! · 500k+ views

10 minUnderstanding Field-Oriented Control | Motor Control, Part 4MATLAB · 100k+ views

AN885: Brushless DC Motor Fundamentals (Microchip) · Sensorless BLDC Position and Speed Control: A Technical Review (arXiv)

A stepper gives you position control without feedback. A toothed rotor snaps between discrete magnetic detents (typically 200 per revolution, microstepped finer), so counting pulses tells you where the rotor is. Open-loop positioning this cheap is why steppers run essentially every 3D printer, desktop CNC, camera stage, and lab automation axis in use. Closed-loop variants add an encoder and fix the classic lost-step failure for modest extra cost.

Steppers are very cheap, simple to control, hold full torque at standstill, and position repeatably with no sensors. Weaknesses: torque falls off quickly with speed, and the motor draws full current while holding still, which makes it hot and inefficient. Missed steps under overload produce position errors with no warning, resonance bands cause vibration, and torque density is poor. Steppers position light loads; they don't power limbs.

Pick a stepper when the load is light and predictable, speeds stay in the low hundreds of RPM, and the whole axis has to land under roughly $30. That covers instrument stages, syringe pumps, tool positioners, and printer-class motion where full holding torque at standstill is useful. If a missed step would matter but the budget still can't reach a servo, use a closed-loop variant with an encoder. Avoid steppers wherever the load can vary or collide, because lost steps fail silently, and wherever watts or heat matter, because the motor draws full current doing nothing. Above that line, use a BLDC servo axis. For powered limbs, a stepper was never a candidate.

200 full steps per revolution (1.8° step angle), microstepped to 1/16 or finer · holding torque around 0.4 Nm for a NEMA 17, several Nm for a NEMA 34 · usable speed in the low hundreds of rpm before torque falls off · positioning accuracy roughly ±5% of a step, non-cumulative · motor $5–20 plus about $2 for the driver.

Every FDM 3D printer axis (NEMA 17 as a de facto standard), syringe pumps and lab robots, pick-and-place machines, telescope mounts; hybrid closed-loop steppers (ClearPath, cheap Chinese servo-steppers) taking share from the low-end servo market.

STEPPERONLINE — commodity NEMA steppers at volume · Teknic — ClearPath closed-loop servo-steppers

Steppers are the cheapest precision positioning per axis available, at $5–20 for the motor and $2 for the driver, with a large commodity ecosystem behind them. In robotics proper they're confined to instrument axes and tooling, but they set the cost floor for everything above them: any servo axis has to justify its premium against a $15 stepper that already does open-loop positioning.

14 minHow Stepper Motors Work - Electric motorThe Engineering Mindset · 1m+ views

6 minStepper Motors Explained | Full Step, Half Step and MicrosteppingEngineering Technology Simulation Learning Videos · 100k+ views

11 minClosed loop stepper motors -- very impressiveMatthias random stuff · 100k+ views

Stepper Motor Basics (Oriental Motor) · Control of Stepping Motors: A Tutorial (University of Iowa)

An integrated servo is a motor sold as a finished solution: motor, encoder, drive, and tuning packaged as one closed-loop product. The range runs from the $3 hobby RC servo up to the industrial AC servo axes that power every CNC machine and industrial robot. What you're paying for is the integration, since you buy position (or velocity, or torque) control as a commodity instead of engineering it yourself. "Smart servos" (Dynamixel and its descendants) brought daisy-chained digital buses to hobby and research robotics.

Engineering time drops to configuration time. Industrial versions come with decades-proven reliability, safety functions, and fieldbus integration, and the surrounding ecosystem of cables, gearheads, and software is complete. The downside is that you inherit the vendor's choices, usually including a high-ratio gearbox with poor backdrivability. Hobby-grade servos strip gears and burn out under sustained load. The markup over loose components is what the integration is worth, and at humanoid volumes that markup is being competed away by in-house modules.

Buy an integrated servo when engineering time is scarcer than unit cost. That covers research platforms and low-volume machines (Dynamixel-class smart servos) and industrial axes that need certified safety functions, fieldbus integration, and decade-scale reliability (Yaskawa/Beckhoff-class). It's usually the right call up to a few hundred units a year, or wherever the axis isn't your product's differentiator. If the joint needs backdrivability or force transparency, skip it, because the bundled high-ratio gearbox rules that out; use a QDD module instead. At humanoid-scale volumes, the integration markup pays for an in-house motor-gear-encoder module within a product generation.

Rated torque roughly 0.3–30 Nm across 100 W to 5 kW industrial frames · rated speed usually 2,000–3,000 rpm, peak 4,000–6,000 · absolute encoders of 20 bits and up on modern industrial servos · bundled gearhead backlash typically 3–15 arcminutes on standard grades · unit cost $3 for a hobby RC servo, $30–100 for a smart servo, roughly $500–3,000 per industrial axis.

Yaskawa, Fanuc, Siemens, and Beckhoff AC servo axes across industrial automation; Dynamixel (Robotis) as the research-robot standard; Feetech/Waveshare clones powering the low-cost arm wave (SO-ARM100 class); KEBA-driven cobot joints.

ROBOTIS — Dynamixel smart servos, the research standard · Yaskawa — industrial AC servo axes

Industrial servo axes are a mature, oligopolistic $10B+ market with stable margins. The disruption is at the bottom: $30 Chinese smart servos now offer what cost $300 a decade ago, and the open-source robot arms built on them have pushed down the entry price of manipulation research. For robot builders the recurring question is whether to buy the module or build the joint, and volume is usually what decides it.

16 minWhat is a Servo Motor and How it Works?RealPars · 1m+ views

10 minWhat is an AC Servo Motor? from AutomationDirectAutomationDirect.com · 50k+ views

3 minTechnical animation: How a Servo Motor workslearnchannel · 1m+ views

Servo Motor Overview (Oriental Motor) · Dynamixel Smart Actuator Documentation (ROBOTIS)

Direct drive means no transmission at all: a large-diameter, many-pole motor drives the load directly. There's no backlash, no gear friction, and no reflected-inertia amplification, so it's about the cleanest torque source available, with position resolution limited only by the encoder. The cost is fundamental. Torque scales with motor volume and current, so direct drive means big, heavy, hot motors doing work a gearbox would do for free. Frameless kits (rotor plus stator, no housing) let you build the motor into the joint structure itself.

Backdrivability and force transparency are as good as they get, precision reaches sub-arc-second with a good encoder, operation is silent, there are no wear parts, and control bandwidth is the highest of any option here. Weaknesses: torque density is the worst of any electric option, and continuous torque at temperature is the binding limit. Holding a gravity load burns power the whole time. The motors are expensive per Nm because there's no gear multiplying anything.

Go direct drive when transparency or precision is the product and the axis carries no standing gravity load. Typical cases are haptic interfaces, gimbals, rotary tables, metrology and wafer stages, and force-sensitive test rigs, where the spec is sub-arc-second resolution, zero backlash, or kilohertz force bandwidth and the motor can be as big as it needs to be. Avoid it for gravity-loaded limbs and mobile robots, since holding torque burns continuous power and torque density is the worst on this sheet. That's the gap QDD closes with a single 6–10:1 stage at a fraction of the motor mass. Budget the encoder alongside the motor, because at this precision class it can cost as much as the motor.

Zero backlash by construction · continuous torque from under 1 Nm to several hundred Nm depending on diameter · specific torque roughly 2–10 Nm/kg continuous, the lowest of the electric options here · position resolution to sub-arc-second with a matched encoder · rated speed usually a few hundred rpm · encoder cost often as much as the motor.

Semiconductor wafer stages and rotary tables (the natural home), Kollmorgen/ETEL/Tecnotion frameless lines, the direct-drive SCARA arms of the 1980s (AdeptOne, which proved the approach worked and also showed why it stayed niche), haptic devices, gimbal motors on every camera drone.

Kollmorgen — frameless torque motor kits · Celera Motion — direct-drive motors and precision encoders

Direct drive owns the applications where precision or transparency is the product and mass is tolerable: stages, metrology, and haptics. As robot joints, pure direct drive lost out to QDD, but the same frameless motors are QDD's core component, so category volume still benefits from humanoid demand. Price per Nm is the metric to compare on, and encoder cost often rivals the motor's.

4 minWhat is a frameless motor?Design World · 10k+ views

9 minCustom Brushless Robotic Actuator using Frameless Motors! (CubeMars RO Series)Kevin Wood | Robotics & AI · 10k+ views

3 minDriving gimbal motors with moteusJosh Pieper · 10k+ views

Direct Drive Motors, Frameless Resolvers and Ring Encoders (Novanta Celera Motion)

The planetary is the workhorse transmission. A sun gear, planet gears, and a ring gear share the load across multiple tooth contacts, giving 3–10:1 per stage in a compact coaxial package that stacks to any ratio. That load sharing gives planetaries the best torque density of any conventional gearing, and at low ratios (a single stage) they stay efficient and backdrivable. That's the property that made them the transmission of choice for quasi-direct-drive designs.

Torque density is high, efficiency runs 90–97% per stage, the layout is coaxial, they take shock well, and they're manufactured everywhere at every quality grade. Weaknesses: backlash is on the order of arcminutes unless the gearbox is preloaded, and precision grades cost accordingly. Multi-stage stacks lose efficiency and backdrivability quickly, and they get noisy at high speed. At the ratios arms traditionally need (50–150:1), stacked planetaries end up bigger than strain-wave and cycloidal designs.

Default to a planetary for any ratio you can reach in one or two stages. Use a single 3–10:1 stage for backdrivable QDD limbs, and two stages up to roughly 100:1 for drivetrains, wheels, and general servo axes where arcminute backlash (or the cost of a preloaded grade) is acceptable. It's the cheapest torque density you can buy, at every quality point from hobby grade to ground-and-preloaded. If the spec is 50:1 or more in a pancake envelope with zero backlash, use a strain-wave gear instead. If it's a base joint taking heavy shock at high ratio, use a cycloidal, which survives loads that stacked planet gears won't.

Ratio 3–10:1 per stage, stacking to 100:1 and beyond · efficiency 90–97% per stage · backlash 6–15 arcminutes on economy grades, under 1 arcminute ground and preloaded · acceleration torque typically 1.5–2× the nominal rating · unit cost $10 for hobby gearheads to $1,000 and up for precision servo grades.

Every cordless drill; Neugart/Wittenstein precision gearheads on industrial servos; single-stage planetaries inside MIT Cheetah-descended QDD joints (Unitree, Tesla Optimus rotary joints); drone gimbal and rover wheel hubs; automotive transmissions by the hundred million.

Neugart — precision planetary gearheads · WITTENSTEIN — low-backlash servo planetaries

The category is fully industrialized, with a steep quality-price ladder running from $10 hobby gearheads to $1,000+ ground-and-preloaded servo grades. Robotics demand concentrates in the low-backlash middle, where Chinese suppliers are pushing prices down. Because the single-stage planetary is the enabling component of QDD modules, it's one of the highest-volume beneficiaries of the humanoid buildout.

8 minHow does a planetary gear work? | Design and operating principle simply explainedtec-science · 100k+ views

10 minWhat makes planetary gearboxes so amazing?3D Printer Academy · 5m+ views

14 minHow to Achieve Massive Speed Reductions with Planetary Gears (160:1 and Beyond!)Mentored Engineer · 10k+ views

Epicyclic Gearing Explained (Neugart) · Gear Technical Reference (KHK Gears)

The strain-wave drive is the transmission that made the modern robot arm practical. An elliptical wave generator flexes a thin-walled flexspline inside a rigid circular spline, and the two-tooth difference between them gives ratios of 50–160:1 in a single pancake-thin stage with essentially zero backlash. No other transmission packs that much ratio into that little space and mass, which is why six of them sit in nearly every industrial robot and cobot wrist built since the 1980s.

You get zero backlash, very high single-stage ratios, a thin coaxial form factor with a through-hole for cables, and excellent positional repeatability. Weaknesses: the flexing metal spline has finite fatigue life and strict shock-load ceilings, since impacts can ratchet or crack it. Efficiency is mediocre (60–85%) and friction depends strongly on load. Torsional wind-up and resonance make force control harder, and the drive is essentially non-backdrivable. Until recently, one company set the world price.

Choose strain-wave when you need 50–160:1 with zero backlash in minimum mass and axial depth. That means arm wrists and distal joints, cobots, positioning duty where trajectories are controlled and impacts are rare, and anywhere a cable pass-through hole is worth having. Design around its limits: derate for shock, since a single hard impact can ratchet the flexspline, and plan for force control through its wind-up and load-dependent friction to be a real controls project. Avoid it for impact-exposed proximal joints, where a cycloidal handles the hits better, and for contact-rich tasks that need backdrivability, where a QDD joint is the modern answer. If you're buying cheap Chinese units, qualify them on fatigue-life data rather than catalog torque.

Ratio 50–160:1 in a single stage · backlash under 1 arcminute, effectively zero · repeatability typically within a few tens of arcseconds · efficiency 60–85%, worse at low load and low temperature · wave generator life 7,000–10,000 hours L10 at catalog torque and 2,000 rpm input, 35,000–50,000 hours L50 average · unit cost $500–2,000 from incumbents, roughly a third of that from Chinese suppliers.

Harmonic Drive Systems (the namesake, ~half the world market), every UR cobot joint, most industrial-robot wrists, space mechanisms from lunar rovers to JWST; Chinese challengers (Leaderdrive, Zhongdadi) scaling with the humanoid boom; Tesla Optimus uses strain-wave stages in high-torque joints.

Harmonic Drive — the strain-wave originator · Leaderdrive — Chinese strain-wave supplier scaling with humanoid demand

This is a textbook precision-manufacturing moat. The know-how in flexspline tooth grinding and material fatigue supported a near-monopoly for decades at $500–2,000+ per unit. Humanoid demand (a dozen or more units per robot) has started a real supply race, with Chinese units now at about a third of incumbent prices and quality closing. Compare fatigue-life data rather than catalog specs, because that's where cheap units still differ.

16 minWhat is Strain Wave Gear a.k.a. Harmonic Drive? A Perfect Gear Set For Robotics Applications!?How To Mechatronics · 500k+ views

2 minHarmonic Drive® Strain Wave Gear: Functional Principle | The OriginalHarmonic Drive SE · 500k+ views

19 minWhy Harmonic Drives Are Awesome.Jeremy Fielding · 100k+ views

Strain Wave Gear Operating Principle (Harmonic Drive) · Harmonic Drive Gear Failures in Industrial Robots: An Overview (PHM Society)

The cycloidal is the heavy-duty precision transmission. An eccentric input rolls cycloidal discs against ring pins, engaging a large fraction of the lobes at the same time. That multi-tooth contact gives cycloidals the two things harmonic drives lack: very high shock-load capacity (they survive 500% momentary overloads) and high torsional stiffness, at ratios of 30–300:1. They dominate the high-torque base joints of industrial robots the same way harmonic drives dominate the wrists.

Overload tolerance and rigidity are both very high, backlash is near zero when preloaded, fatigue life is long, and single-stage ratios are high. Weaknesses: they're heavier and bulkier than strain-wave at equal ratio, and the eccentric motion needs balancing and produces a small cyclic ripple. Manufacturing tolerances are very tight, which is part of why the market is a duopoly. Efficiency is similar to harmonic, and the drive is effectively non-backdrivable.

Pick a cycloidal for high-torque proximal joints that will take shock loads: industrial-robot bases and shoulders, humanoid hips and knees, and anything that lifts hundreds of newton-meters or absorbs impacts as part of normal operation. Its 500% momentary overload rating and torsional stiffness are what keep those joints from becoming service calls. It suits ratios of 30–300:1 and a mass budget that can absorb the extra kilograms over strain-wave. Avoid it in wrists and distal joints where grams and axial depth dominate, since harmonic packages better there, and in force-controlled, contact-transparent joints, where a low-ratio QDD stage is the right architecture.

Ratio 30–300:1 in one or two stages · momentary overload capacity around 500% of rated torque · backlash under 1 arcminute when preloaded · efficiency roughly 80–90% at rated load · rated service life 6,000 hours at rated torque and 15 rpm output, which is not the same basis as a harmonic drive's rated hours.

Nabtesco RV series — the base and shoulder joints of the large industrial robots of Fanuc, ABB, KUKA, Yaskawa (Nabtesco holds ~60% of the world precision-reducer market); Sumitomo Fine Cyclo; Spinea; Chinese entrants (Shuanghuan, Zhongdadi) chasing the market; 3D-printed hobby cycloidals as a maker staple.

Nabtesco — RV cycloidal reducers, ~60% world share · Spinea — compact TwinSpin cycloidals

Cycloidals are the other half of the Japanese precision-reducer duopoly, and the story looks much like harmonic drives': decades of stable oligopoly pricing, now under pressure from Chinese suppliers as robot volumes grow. Their shock tolerance makes them good candidates for humanoid hips and knees, which take impacts constantly. Whoever industrializes a compact, cheap cycloidal for that duty gets a large share of the embodied-AI BOM.

16 minHow does a cycloidal gearbox work? | Structure and function simply explained | parametric equationtec-science · 100k+ views

17 minWhat is Cycloidal Drive? Designing, 3D Printing and TestingHow To Mechatronics · 1m+ views

3 minAnimation CYCLO Principle - Sumitomo Drive TechnologiesSumitomo Drive Technologies EMEIA · 100k+ views

Functional Principle of Cycloidal Gears (Nabtesco) · RV Two-Stage Reduction Engineering (Nabtesco Motion Control)

A screw converts rotary motion to linear motion with mechanical advantage in the thousands: the screw turns and the nut advances. Lead screws slide, which makes them cheap, self-locking, and lossy. Ball screws recirculate bearing balls between the threads, which makes them 90%+ efficient, precise, and expensive. Both give the highest force density of any electromechanical option, since a small motor plus a fine-pitch screw can press tonnes. That's why machine tools, presses, and now humanoid knees are built on them.

Force multiplication is very large, ground ball screws position to microns, lead screws self-lock so they hold load with no power, and the whole arrangement is stiff and simple. Weaknesses: the motion is linear only, and speed is capped by screw whip and nut dynamics. Lead screws lose 50–70% of the input to friction. Ball screws are sensitive to shock and contamination. Backdrivability runs from poor (ball) to none (lead), so force control has to be sensed and computed rather than felt through the mechanism.

Pick a screw when the joint is fundamentally a linear force problem: humanoid knees and ankles, presses, clamps, and lifts. You get kilonewtons from a modest BLDC, micron repeatability from ground ball screws, and power-off self-locking for free if you use a lead screw. If the same envelope has to carry 3–10× the load and shock, step up to a planetary-roller screw. Avoid screws where the motion is naturally rotary, where speed matters (whip caps travel rates), or where the joint needs to feel contact, since backdrivability runs from poor to none. If you need transparent linear force, use a linear motor or a cable drive. If you need a backdrivable rotary joint, use QDD.

Ball screw efficiency 90% and up; lead screw 30–50% · lead typically 2–25 mm per revolution · positioning accuracy about ±8 µm over 300 mm on C3-grade ground ball screws · axial load from a few hundred N to tens of kN, and 3–10× that from a planetary roller screw in the same envelope · lead screws self-lock below roughly 5° helix angle, holding load with no power.

Every CNC machine axis and injection-molding press; THK, NSK, Hiwin as the incumbent industry; electric vehicle brake and steering actuators; Tesla Optimus and Figure's linear knee/ankle actuators, reflecting the fact that leg joints are largely linear force problems; aircraft flap actuators.

THK — ball screws and linear motion at global scale · Rollvis — planetary roller screws for humanoid-class linear joints

Standard grades are a mature bearing-industry commodity, while precision ground screws still carry real margins. Humanoid designers have been choosing planetary-roller screws (the ball screw's high-force cousin, using rollers instead of balls for 3–10× the load capacity and price), which turned a quiet aerospace niche into a fast-growing one. Compact roller screws could end up being the humanoid era's supply chokepoint, in the way harmonic drives were for industrial arms.

6 minLead Screw, Ball Screw, and Roller Screw Cylinders: What's the Difference?NEFF Automation Experts · 100k+ views

10 minChasing Micrometres with the best Ball ScrewsMarco Reps · 1m+ views

5 minAdvantages of Roller Screws | Planetary vs Ball Screw Actuator TechnologyTolomatic · 50k+ views

2.72 Elements of Mechanical Design: Lecture Notes incl. Screw Drives (MIT OpenCourseWare)

These transmissions move torque over a distance. Synchronous belts, steel cables, and capstan windings carry torque from a motor mounted where mass is cheap (the base) to a joint where mass is expensive (the wrist). Cable and capstan drives also give near-zero backlash and very good backdrivability, because the friction is rolling rather than sliding. That's why they're used in haptics and surgical robotics, where feeling the load is the whole point of the device.

Remote actuation cuts limb inertia substantially. Tensioned cables run smooth, quiet, backdrivable, and backlash-free, and belts are cheap and tolerant of misalignment. The main weakness is compliance: belts stretch and cables creep, which costs stiffness and means tensioning maintenance. Routing through multiple joints couples those joints kinematically. Power density is modest, and cable fatigue and fraying are wear items that need inspection (part of why surgical robots replace instruments on a schedule).

Reach for belts and cables when moving mass is the problem. Putting the motors at the base drops limb inertia a lot, which is why tendon drives are standard in dexterous hands and cable capstans are standard in haptics and surgical instruments, where backdrivable, backlash-free feel is the product. Belts are the cheap default for planar stages (CoreXY, SCARA, deltas) and any long run that has to tolerate misalignment. Avoid them when you need stiffness under high load or maintenance-free life, since cables creep, fray, and need tensioning; for heavy sustained force, use a screw or a geared joint instead. Keep the routing simple, because every joint a cable crosses becomes a kinematic coupling you have to compensate for.

Synchronous belt efficiency 95–98% · capstan reduction typically 5–20:1 per stage · backlash effectively zero on a tensioned cable or capstan drive · belt pitch 2–5 mm on small robots (GT2 and GT3 profiles) · minimum pulley diameter roughly 20× the cable diameter for reasonable fatigue life.

The da Vinci surgical system's cable-driven instruments (the category's flagship), Phantom haptic devices' capstans, 3D printer CoreXY belt systems, SCARA and delta robot belt stages, tendon-driven robot hands (Shadow Hand), Agility Robotics' leg transmissions mixing cables and links.

Gates — synchronous belts and drives · SDP/SI — small-mechanism belts, cables, and pulleys

Belts are a commodity. Precision cable and capstan work is closer to craft engineering, with real know-how moats in tension management, termination, and routing, but there's little standalone industry around it; the expertise mostly lives inside products like surgical and haptic systems. For humanoid hands, tendon drives are the current compromise between dexterity and packaging, which keeps this old craft relevant.

20 minHigh Precision Speed Reducer Using RopeAaed Musa · 1m+ views

19 minTeardown of Surgical Robot InstrumentKyle Bartholomew · 100k+ views

13 minCoreXY explained: Comparison + strengths & weaknessesTeaching Tech · 500k+ views

Forte: A Strong, Accurate, Low-Cost Capstan-and-Belt Robot Arm (arXiv)

QDD is the architecture that took over legged robotics. You pair a large-diameter, high-pole-count BLDC with a single modest planetary stage, typically 6–10:1. The low ratio keeps reflected inertia and friction small, so the joint stays backdrivable enough to feel contact through motor current alone, which gives you proprioceptive force control with no force sensor. The gear stage still raises torque density enough to drive a real limb. The MIT Cheetah popularized the approach, and it's now the default recipe in quadrupeds and most humanoid rotary joints.

Transparency and impact tolerance are both good, since impacts backdrive the motor instead of breaking gear teeth. You get force estimation for free, high control bandwidth, and a mechanically simple joint. Weaknesses: torque density is below what high-ratio geared joints achieve, so QDD limbs run hot holding static loads. The big-diameter motors dominate joint mass and cost. Precision is limited by the motor-side encoder plus the gear's small backlash, which is fine for locomotion but short of what machining-grade arms need.

QDD is the default for contact-rich dynamic joints: legs, dynamic arms, and anything that makes hard contact and needs to sense it. Impacts backdrive the joint harmlessly, torque is estimated from current with no load cell, and $100–500 buys the whole module off the shelf. Size it for duty cycles dominated by motion rather than holding, because the torque density is real but the thermal headroom for standing under gravity all day isn't. If the joint mostly holds static load, use a 50:1+ strain-wave joint, which runs cooler and smaller. If the spec is machining-grade repeatability, a harmonic-drive servo axis is still the right tool, since QDD's motor-side encoder and residual gear lash can't get there.

Gear ratio 6–10:1 in a single planetary stage · continuous torque typically 5–20 Nm from a module under 1 kg, with peaks 2–3× higher · specific torque roughly 10–20 Nm/kg continuous · joint efficiency around 85–90%, since only one gear stage sits in the path · module cost $100–500 for motor, gear, encoder and FOC drive together.

MIT Cheetah/Mini Cheetah (the archetype), Unitree's entire quadruped and humanoid line, Tesla Optimus rotary joints, open-source actuators (MIT-inspired T-Motor AK series, ODrive-based builds), Ghost Robotics.

CubeMars — AK-series QDD modules, the open-robotics default · MyActuator — integrated QDD joint modules

The AK-series-class QDD module packages motor, planetary, encoder, and FOC drive into one puck for $100–500. It started as a research part and is turning into a commodity at Chinese volume pricing. Margins are moving to whoever owns module integration and reliability data. The open question is thermal: QDD humanoids that stand still all day may push designers back toward higher ratios.

17 minMIT Mini Cheetah ActuatorSkyentific · 100k+ views

16 minMIT Mini Cheetah Actuator TeardownRobert Robotics · 100k+ views

7 min3D Printed Planetary Gearbox Robotic Actuator BLDC with ODrive S1Kevin Wood | Robotics & AI · 10k+ views

Quasi-Direct Drive for Low-Cost Compliant Robotic Manipulation (arXiv) · Alternative Metrics to Select Motors for Quasi-Direct Drive Actuators (arXiv)

A series-elastic actuator puts a spring between the gearbox and the load deliberately. Because the deflection of a known spring is itself a force measurement, force control turns into position control of spring compression, which is robust, cheap to sense, and stable in contact. The spring also filters shock loads before they reach the gear teeth and stores energy during cyclic motion like walking. Pratt and Williamson published the idea in 1995, and it defined a generation of human-safe and legged robots until QDD offered transparency without a spring.

Force fidelity is very good and requires only cheap sensing. Impacts get absorbed rather than transmitted, energy storage improves locomotion efficiency, and the design tolerates the cheap high-ratio gearboxes it wraps around. Weaknesses: the spring caps force-control bandwidth, because fast force changes have to wait for the spring to deflect. It adds mass, volume, and a resonance you have to control around. Position precision degrades through the compliance, and the per-joint engineering complexity never really commoditized.

Choose an SEA when the spring pays for its mass twice over: cyclic joints with real energy exchange (ankles, hips, exoskeleton and prosthetic drives), and human-contact applications where compliance has to survive a controller fault rather than being simulated by the controller. It also makes cheap high-ratio gearboxes usable, since the spring supplies the force sensing and shock protection the gearbox lacks, which helps when the BOM can't afford QDD-class motors. Avoid it when force bandwidth matters, since fast force changes wait on the spring and closed-loop force response typically lands well below what a QDD joint's current loop delivers, and avoid it when position precision is the spec. For general contact-rich joints with no energy-storage story, QDD is the modern default and the design you have to beat.

Force-control bandwidth typically 10–50 Hz, set by spring stiffness and load inertia · spring deflection usually a few degrees at rated torque · gear ratio behind the spring commonly 50–200:1 · peak joint torque 50–150 Nm on legged and exoskeleton designs.

Baxter and Sawyer (Rethink Robotics, which built its safety case on SEAs), NASA's Valkyrie, Agility Robotics' Cassie/Digit lineage (leaf-spring elements), rehabilitation exoskeletons (Lokomat-class) where compliant interaction matters most, ANYmal's early ANYdrive joints.

HEBI Robotics — X-series series-elastic smart actuators

SEA lost the mainstream legged-robot argument to QDD, which gets its compliance from control rather than steel, but it survives where physical energy storage or guaranteed contact safety is worth paying for: exoskeletons, rehab, prosthetics, and ankle and hip joints with large cyclic energy. If you're evaluating a modern SEA pitch, the question to ask is why software impedance on a QDD joint isn't sufficient. Sometimes it genuinely isn't.

3 minSeries Elastic ActuatorsReNeu Robotics Lab · 10k+ views

12 minDIY Series Elastic Actuator for legged and humanoid robots: force control!HOX robotics · 5k+ views

8 minElastic actuators for a prosthetic ankle joint for walking and runningLauflabor · 10k+ views

Series Elastic Actuator: Design, Analysis and Comparison (IntechOpen) · Modeling and Application of Series Elastic Actuators for Force Control (arXiv)

In a variable stiffness actuator, mechanical stiffness is a controlled variable. Each joint uses two motors: one sets position, and the other adjusts the preload on a nonlinear spring. The joint can be rigid for precision one moment and soft for impact the next, and the compliance is physical rather than simulated by the controller. The biological argument for this is straightforward, since muscles co-contract to modulate stiffness all the time. The engineering cost is a second actuator's mass, cost, and complexity in every joint.

The compliance is physical, so it still works when control fails, which is a safety case software impedance can't fully match. Energy storage can be tuned to the gait or task, and impact response is smooth. Weaknesses: you roughly double actuator mass and cost per joint to get a property that QDD approximates in software. The nonlinear spring mechanisms are intricate and prone to wear, and controlling the coupled system is a research field of its own. After two decades of prominence in EU research projects, no volume application has adopted it.

Specify a full VSA only when physically guaranteed, control-independent compliance is a hard requirement. That usually means a safety case where software impedance on a failed controller isn't an acceptable answer, or research studying stiffness modulation itself. If you only need one tunable axis (gait-dependent stiffness in a prosthetic ankle, or a clutch that disengages on impact), use the single-motor, selective versions of the idea instead of two motors per joint. Otherwise avoid it. The second actuator's mass and roughly 2× joint cost buy a property a QDD joint approximates in software for free, and two decades of prototypes with no volume adoption is a base rate worth taking seriously.

Two motors per joint, one setting position and one setting spring preload · stiffness adjustable over one to two orders of magnitude, wider in some research designs · stiffness transition time typically a few hundred milliseconds · joint mass and cost roughly 2× a single-motor actuator of the same output torque.

DLR's Hand Arm System and FSJ joints (the field's flagship), IIT's AwAS and CompAct series, qbrobotics' commercial VSA modules (the rare productization), MACCEPA-based prosthetics research.

qbrobotics — commercial variable-stiffness actuators and soft hands

The category is scientifically rich but commercially stalled, because the extra safety and efficiency over software-compliant QDD hasn't justified 2× joint cost anywhere at scale. The ideas do show up in simpler forms that work: clutches, adjustable ankles in prosthetics, and single-motor nonlinear springs. A good rule of thumb is to treat full two-motor VSA pitches as research instruments, and to look for the commercial opportunity in selective, single-axis stiffness modulation.

4 minneoDavid - A humanoid robot with variable stiffness actuation and dexterous manipulation skillsDLR RM · 5k+ views

2 minVariable Stiffness in Robotic Arms Lab DemoMIT CSAIL Alliances · 1k+ views

Variable Impedance Actuators: A Review (Robotics and Autonomous Systems)

Servo valves meter pressurized oil at 200–350 bar into cylinders and rotary actuators. This gives the highest force density of any actuation technology, by an order of magnitude. A fist-sized cylinder exerts tonnes, and because power arrives through hoses, the joints themselves weigh almost nothing. Hydraulics built every excavator, press, and aircraft control surface, and they powered the dynamism of Boston Dynamics' hydraulic Atlas and BigDog before the company went electric.

Force and power density at the joint are unmatched. Relief valves give intrinsic overload tolerance, force transmission is stiff, and the hardware holds up in filth and heat. The weaknesses start with the system tax: pump, accumulator, valves, hoses, filters, and heat exchanger. At small scale that supporting hardware erases the density win. Servo valves cost thousands and demand immaculate fluid, efficiency is poor (throttling losses often put system-level efficiency at 10–30%), leaks are common, and valve nonlinearity makes force control hard to get right.

Use servo-hydraulics when forces reach tens of kilonewtons per axis, the environment is dirty, hot, or shock-laden, and one power plant can be amortized across many joints. That covers excavator-class machinery, forging and press automation, high-force test rigs, and autonomy retrofits of machines that are already hydraulic. Below roughly excavator scale the math flips: the pump, valves, hoses, and cooler cost mass and money, and 10–30% system efficiency erases the density win. That's why legged robots moved to electric QDD, which needs a tenth the maintenance. If you need hydraulic-class force without a central plant, look at an EHA instead. If a ball screw plus a BLDC can push it, stay electric.

Supply pressure 200–350 bar · cylinder force tens of kN per axis, hundreds of kN on press and forging equipment · servo-valve bandwidth roughly 50–250 Hz · system efficiency 10–30% after throttling losses · positioning typically 0.1–1 mm · $10k+ per axis once the pump, valves, and cooling are counted.

Hydraulic Atlas's backflips (retired 2024, effectively the end of hydraulic humanoids), every excavator and forestry machine (now targets for robotic retrofit: Gravis, Built Robotics), aircraft actuation, industrial testing rigs (MTS/Instron), die-casting and forging automation.

Moog — servo valves and hydraulic motion control · Bosch Rexroth — industrial hydraulics at scale

Hydraulics dominate where forces are enormous and the mass budget is generous: construction, mining, and presses. That's a large, mature industry, and it's being retrofitted with autonomy rather than replaced. In legged robotics the question is settled, because electric QDD matched the dynamism at a tenth the maintenance. What's left for hydraulics is electro-hydrostatic hybrids (the next entry) and autonomy retrofits of existing hydraulic fleets, where the machine has already been bought.

3 minProportional hydraulics, proportional valve, servo valve - how it works - Technical animationlearnchannel · 500k+ views

7 minWhat are Servo Control Valves? - Tutorials and ExplanationInstrumentation Tools · 100k+ views

2 minHow do Hydraulic Actuators work? - A Galco TV Tech Tip | GalcoGalcoTV · 50k+ views

Electro-Hydraulic Valves: A Technical Look (Moog) · Understanding Electrohydraulic Valve Types (Power & Motion)

An EHA is hydraulics with no hydraulic plant. It's a self-contained unit where an electric motor drives its own small pump, moving sealed fluid directly to its own cylinder, so there's no central pump, no servo valves, and no throttling. Force control comes from motor torque (efficient and reversible) rather than valve metering (lossy), which makes EHAs backdrivable in a way classic hydraulics never were. Aerospace adopted them for "more-electric" aircraft. Robotics uses them where electric density runs out but hydraulic infrastructure isn't acceptable.

You get hydraulic force density with electric-drive controllability and 60–80% efficiency. The fluid is sealed for life, each joint is a module with only wires running to it, and the compliance of the fluid gives inherent shock tolerance. The weaknesses: a pump per joint means many precision pumps, and the pump is the expensive, wear-prone part. Bandwidth trails valve-controlled hydraulics. Small EHAs lose a disproportionate share of their power in the pump. And the design work is engineering-intensive, with few off-the-shelf products at robot scale.

Consider an EHA when sustained joint force outgrows what a BLDC plus a roller screw can package (multi-kilonewton loads with shock exposure) and a central hydraulic plant isn't acceptable: aerospace surfaces, heavy exoskeletons, high-payload humanoid legs. You get hydraulic density with motor-side force control, 60–80% efficiency, sealed fluid, and only wires running to the joint. Avoid it while your requirements still fit electromechanics. Every joint carries its own precision pump, nothing is off the shelf at robot scale, and prices are aerospace-adjacent, so adopting one is a program decision rather than a catalog purchase. If you need the bandwidth or force of valve-controlled hydraulics, classic servo-hydraulics is still the better answer.

Force 1–10 kN at robot scale, tens of kN on aerospace control surfaces · internal pressure 200–350 bar · efficiency 60–80%, roughly double a throttled servo-hydraulic system · positioning typically 0.1–1 mm · $10k+ per axis, and most robot-scale units are semi-custom.

F-35 flight-control actuators (the flagship deployment), Moog and Parker aerospace lines, Apptronik's early EHA work and various humanoid knee prototypes, KNR and Kawasaki hydraulic-robot research, heavy-payload exoskeletons.

Domin — additively-manufactured direct-drive servo valves and compact EHA systems · Parker — aerospace and industrial EHA lines

On paper EHAs are the right endpoint for high-force robot joints. What holds them back is the absence of a commodity micro-EHA supply chain, so every unit is semi-custom at aerospace-adjacent prices. If humanoid load requirements outgrow ball-screw-plus-BLDC solutions, EHAs are the successor waiting to be used. A company that industrialized a cheap, sealed 1–10 kN robot EHA would be filling a real gap.

3 minEAS-EPU Electrohydrostatic Actuation SystemMoog · 10k+ views

3 minElectro Hydraulic Actuator (EHA) by Bosch RexrothBosch Rexroth US · 1k+ views

Electrohydrostatic Actuation Technology Overview (Moog) · Electro-Hydrostatic Actuation in Next-Generation Machines (Moog)

Solenoid valves switch compressed air at 6–8 bar into cylinders, which gives fast, simple, two-position actuation for pennies per cycle. Pneumatics is the automation most people never see: industry has vastly more pneumatic axes than servo axes. They do the clamping, ejecting, sorting, and pick-and-place that never needs a controlled trajectory, just "out" and "back" against mechanical stops.

Pneumatics is very cheap and fast (full stroke in tens of milliseconds), intrinsically compliant and overload-safe, clean enough for food and pharma, and explosion-proof by nature. The weaknesses follow from air's compressibility, which makes mid-stroke position control hard enough that it's rarely done in practice; pneumatic axes run to stops instead. System efficiency is poor, roughly 10–20% from compressor to work. Air prep (drying, filtering) and leaks are permanent costs in a factory. Force density is modest at safe pressures.

Use pneumatics for binary motion against mechanical stops at high cycle rates: clamp, eject, sort, gripper open and close. It fits wherever plant air already exists and the axis has to cost tens of dollars, survive washdown, or sit in an explosive atmosphere. Compressibility helps at the gripper, where it gives intrinsic compliance and overload safety, and hurts everywhere else. Avoid pneumatics for mid-stroke position or force control, for anything untethered (the compressor is part of the system), and on energy-audited lines where 10–20% wall-to-work efficiency draws scrutiny. If the motion needs a trajectory rather than two endpoints, a servo-electric gripper or axis is the upgrade.

Line pressure 6–8 bar · cylinder force roughly 200 N to 2 kN across common 20–63 mm bores · full stroke in tens of milliseconds · compressor-to-work efficiency 10–20% · repeatability mm-class, and only at the end stops · cylinder plus valve costs tens of dollars per axis.

SMC and Festo's entire catalogs (a $10B+ duopoly-adjacent industry), every injection-molding sprue picker and packaging line, gripper open/close on countless robot arms, soft-robotics research (air is the usual power source for soft actuators; see the McKibben entry).

SMC — the pneumatics volume leader · Festo — pneumatic and electric automation components

Pneumatics has the lowest cost per motion in automation, and the compressor's inefficiency is hidden in the plant utility bill. Robotics uses pneumatics at the periphery (grippers, tooling) rather than in joints. One trend worth watching: as factories electrify end-of-arm tooling for controllability, cheap servo grippers are taking share from pneumatic grippers, and both SMC and Festo now sell electric axes defensively.

4 minPneumatic Cylinder Working explained (Animation)TecknoMechanics · 1m+ views

6 minDirectional Control Valve Working Animation | 5/2 Solenoid Valve | Pneumatic Valve Symbols ExplainedUpmation · 1m+ views

8 minHow a Industrial Pneumatic Systems Works And The Five Most Common Elements UsedRG Group PA · 500k+ views

4 Basic Pneumatic Circuits (Power & Motion) · Industrial Pneumatics: Operational Essentials and Specification Tips (Power & Motion)

A linear motor is a rotary motor unrolled: the magnets lie flat as a track and the coils form a moving forcer, producing linear force directly with no screw, belt, or gear in between. Because nothing converts rotary motion into linear motion, nothing backlashes, wears, or limits speed. Linear motors hit 10+ m/s and nanometer-class precision at the same time, which is why semiconductor lithography stages, the most precise moving machines ever built, use nothing else. Iron-core versions maximize force; ironless versions eliminate cogging, which makes the motion very smooth.

Zero backlash, no transmission wear, extreme speed and acceleration, sub-micron precision with linear encoders, and travel you extend by adding track. The main weakness is that there's no mechanical advantage: every newton comes from amps and magnets, so continuous force per dollar is poor. Strong open magnet tracks collect ferrous debris and are awkward to handle. Holding force requires constant current or a brake. And the magnet track's cost scales with travel length.

Pick a linear motor when a linear axis has to be both fast and precise (meters per second and micron-or-better accuracy in the same move), or when throughput per track meter justifies the cost of the magnets: lithography stages, SMT gantries, programmable transport lines. Ironless variants give you cogging-free smoothness for scanning; iron-core variants give you force. Avoid linear motors when the axis mostly holds force rather than moves, because holding burns continuous current or needs a brake, while a ball screw holds for free. Also avoid them when travel is long but speed is modest (a belt or screw costs a fraction of the track), or around ferrous swarf, which the open magnet track will collect.

Continuous force roughly 50 N to 5 kN per forcer, with peak around 3× continuous · speeds above 10 m/s · acceleration typically 3–10 g on light stages · zero backlash, sub-micron positioning with a linear encoder and nanometer-class on lithography stages · $1k–10k per axis, and the magnet track's cost scales with travel length.

ASML wafer and reticle stages, pick-and-place machine gantries (the SMT industry runs on them), laser cutters and high-end CNC axes, maglev-style transport systems (Beckhoff XTS, Planar motors), Rockwell/B&R track systems replacing conveyor lines.

Tecnotion — iron-core and ironless linear motors · ETEL — precision direct-drive systems for semiconductor stages

Linear motors own the applications where precision times throughput justifies the cost of the magnets, semiconductors above all. In robotics proper they appear in gantry robots, and they're moving into "linear transport" factory lines that replace conveyors with programmable movers. Falling magnet-track prices from Chinese suppliers are widening the niche downward into general automation.

3 minHow does a linear motor work?Tecnotion - direct drive in motion · 10k+ views

6 minLinear Motors | How do they work?Sabin Civil Engineering · 500k+ views

8 minThe entire linear motor manufacturing process including the production of the moving stator.YHDFA · 1m+ views

What Are the Benefits of a Linear Motor? (Tecnotion) · Why Choose an Ironless Motor? (Tecnotion)

A voice coil actuator applies the loudspeaker principle to precision motion: a coil sits in a permanent-magnet gap, force is exactly proportional to current, and the coil moves without friction over short strokes (millimeters to a few centimeters). There's no cogging, no hysteresis, and no mechanical contact, which makes it the cleanest force source in engineering, with bandwidth into the kilohertz. Every hard drive head, camera autofocus, and fast-steering mirror is a voice coil, and haptics and micro-dosing borrowed them for the same reasons.

Force is perfectly linear with current, there's no friction or backlash, bandwidth reaches kilohertz, the actuator is silent, and it's intrinsically backdrivable. The weaknesses: stroke is short. Force density is poor, holding force burns continuous current, and there's no holding at all with the power off. Heat in the moving coil limits continuous duty. And past a few centimeters of travel, a linear motor is the better version of the same idea.

Use a voice coil when the stroke is millimeters to a couple of centimeters and you need fast, clean force: kilohertz bandwidth, force exactly proportional to current, zero friction. That fits fast-steering mirrors, precision dosing, haptic transducers, and the fine stage of a coarse-fine axis layered on a screw or linear motor. If travel goes beyond a few centimeters, use a linear motor instead (same physics, packaged for distance). Avoid voice coils wherever the load must be held without power, since there's no detent and no self-locking, so you'd have to add a brake or use a screw. Avoid them for sustained high force too, because the moving coil reaches its thermal limit quickly.

Stroke a few millimeters to a few centimeters · continuous force typically 1–100 N, with peak several times higher in short bursts · bandwidth into the kilohertz · sub-micron resolution with a position encoder · no holding force at zero current · cents at phone volume, hundreds of dollars for precision industrial units.

Hard-disk head actuators (billions made, the highest-volume example), smartphone camera OIS/AF modules, fast-steering mirrors in optics and laser comms, Apple's Taptic Engine, wafer-stage fine-positioning layers, ventilator and dosing valves.

H2W Technologies — voice coil and linear actuator specialist · Akribis — voice coils and precision stages

There are two economic worlds here. Consumer-electronics voice coils are stamped out for cents at phone volume, while precision industrial units (Akribis, ThorLabs, BEI) run hundreds of dollars. In robotics they appear wherever a joint needs a fast, fine force stage layered on a coarse actuator. Semiconductor equipment perfected that coarse-fine architecture, and humanoid hands may yet rediscover it.

3 minVoice Coil ActuatorCelera Motion, A Novanta Company · 50k+ views

6 minHard drive voice coil motor demonstrationmbbrutman · 5k+ views

16 minEEVblog 1400 - Hard Drive Micro Actuators are AMAZING!EEVblog · 100k+ views

What Is a Voice Coil Actuator? (H2W Technologies)

A solenoid is the bluntest form of electromagnetic actuation: energize a coil and an iron plunger snaps in. It's binary, fast, and nearly free, and it does the industrial world's twitch work. Solenoids open valves, throw latches, kick parts off conveyors, and fire pinball flippers. Force is highly nonlinear (weak at full extension, strong near closure), so solenoids are good for impacts and holds, not motion control.

Solenoids are the cheapest electric actuator per unit, respond in milliseconds, need no electronics beyond a switch, and are very reliable in on/off duty. The weaknesses: they're essentially uncontrollable mid-stroke, and the force-stroke curve is severely nonlinear. Continuous energization overheats the coil (latching variants fix this with magnets). Stroke is short, and efficiency in sustained duty is poor. A solenoid is a switch that moves, not an actuator that positions.

Use a solenoid when the requirement is genuinely binary and short-stroke: throw a latch, fire a valve, kick a part, engage a brake. You get millisecond response for cents to dollars, with nothing to tune. If the state has to be held for minutes or held with the power off, specify a latching (permanent-magnet) variant before the coil overheats. Avoid solenoids as soon as the spec mentions position, speed, or proportional force anywhere between the endpoints, because the force-stroke curve won't cooperate. If you need controlled short-stroke force, use a voice coil; if you need controlled travel, use a small gearmotor.

Stroke usually 2–25 mm · force a few newtons at full extension, rising to tens of newtons near closure · response roughly 5–30 ms · duty cycle typically 10–50% on non-latching coils, since continuous current overheats them · cents to a few dollars per unit.

Every fluid solenoid valve (the largest actuator population on earth), door locks and vending machines, part-reject flippers on sorting lines, pinball machines, camera shutters, automotive starter engagement.

TLX Technologies — custom solenoids and electromagnetic actuators · Johnson Electric — motors and solenoids by the hundred million

Solenoids are a fully mature commodity measured in cents to dollars, and they have no robotics story of their own. They are everywhere in robot peripherals, though: pneumatic valve manifolds, tool-changer locks, and safety brakes are all solenoid-actuated. They get an entry here mainly so the niche is recognized. When the requirement is truly binary, nothing cheaper or more reliable exists.

9 minSolenoid Basics Explained - Working PrincipleThe Engineering Mindset · 1m+ views

2 minHow does a Linear Solenoid work? (Tubular Solenoid)Geeplus · 50k+ views

8 minHow Solenoid Valves Work - Basics actuator control valve working principleThe Engineering Mindset · 1m+ views

How to Select a Linear Solenoid (Geeplus) · Shaping the Solenoid Force Curve (TLX Technologies)

Piezo ceramics are crystals that flex under voltage. They deform by only ~0.1% of their length, but they do it with sub-nanometer resolution, microsecond response, and enormous force. Stacks deliver tens of microns at kilonewtons, amplified flexures trade force for millimeter strokes, and ultrasonic or stick-slip motors chain microscopic steps into unlimited travel at surprising speed. Wherever positioning is measured in nanometers (microscopy, photonics, semiconductor metrology), piezo is the only option.

Resolution and bandwidth are unmatched. Piezo works in vacuum and at cryogenic temperatures, produces no magnetic fields, holds position unpowered (stacks), and is immensely stiff. The weaknesses: native stroke is microscopic, so any longer motion comes from amplification, and amplification has its own costs. Hysteresis and creep mean you need closed-loop sensing for accuracy. Drive voltages are high, 100–1000 V. The ceramics are brittle and don't tolerate tension or shock. And ultrasonic motors wear at their friction interfaces.

Reach for piezo when resolution is measured in nanometers, bandwidth in kilohertz, or the environment rules out electromagnetics (vacuum, cryogenic, MRI-adjacent, or field-sensitive metrology), and the stroke fits in microns for stacks or a few millimeters for amplified flexures. Ultrasonic and stick-slip motors extend travel further when speed can be modest. Two bonuses no coil matches: stacks hold position unpowered, and stiffness is extreme. Avoid piezo when the stroke is centimeters at high speed or the load is limb-scale, because amplification eats force and the ceramics don't tolerate shock or tension. Budget for the overhead up front, since closed-loop sensing for hysteresis and 100–1000 V drives are part of the bill. If you need clean millimeter-scale force at low voltage, a voice coil is the simpler tool.

Native strain about 0.1% of stack length, so stacks give tens of microns and amplified flexures a few millimeters · blocking force into the kilonewtons for a stack · sub-nanometer resolution and microsecond response · drive voltage 100–1,000 V · open-loop hysteresis roughly 10–15%, so accuracy needs closed-loop sensing · $1k–10k for a closed-loop stage.

Every AFM and confocal microscope stage (Physik Instrumente, npoint), camera autofocus ultrasonic ring motors (Canon USM lineage), fuel-injector piezo stacks (billions of firings), fiber-alignment stages, and inkjet printheads, which are arguably the highest-volume piezo actuator of all.

Physik Instrumente — piezo stages and stacks, the precision incumbent · PiezoMotor — piezo-stepping micro motors

Piezo is a profitable precision oligopoly (PI, Cedrat, Noliac/CTS) at instrument-industry margins. Automotive-scale stack production for injectors shows the ceramics can be cheap at volume. Robotics touches piezo only at the extremes: micro-manipulation, surgical micro-tools, and any claim of nanometer dexterity. It won't ever move a limb, and it doesn't need to.

3 minPiezoelectric motorApplied Science · 100k+ views

5 minKEMET Piezoelectric ActuatorsTOKIN · 50k+ views

10 minPiezo Actuators & Their Limitations in Digital MicroscopyDover Motion · 10k+ views

Piezo Basics: Fundamentals of Piezoelectricity and Piezo Actuators (PI) · Piezoelectrics in Positioning: Nanopositioning Tutorial (PI, PDF)

Below roughly a millimeter, Coulomb attraction between charged surfaces beats magnetics, and silicon micromachining builds actuators by the wafer-full. Comb drives and parallel plates steer mirrors, tune optics, and pump micro-fluids, with the die itself costing cents; thermal and piezoelectric micro-elements cover the same jobs where electrostatic force falls short. The same physics that is hopeless at robot scale, where the forces are micronewtons, is dominant at chip scale.

MEMS actuators are batch-fabricated on semiconductor economics, thousands per wafer, with nanometer precision, microsecond speeds, and near-zero power (electrostatic holding is capacitive). The weaknesses: forces and strokes are microscopic, so they're useful only where the load is also microscopic. Stiction and dielectric charging are chronic failure modes. Packaging often costs more than the die. And scaling up in force means ganging thousands of elements, which is rarely worth it.

Choose MEMS when the load itself is microscopic (a mirror facet, an optical element, a microliter of fluid) and volumes justify wafer economics: thousands of units per wafer, nanometer precision, near-zero holding power. It's a design-win business, so plan on packaging and fab NRE dominating the die cost. Avoid MEMS for any macroscopic force or stroke, because micronewtons don't gang up economically. A good rule of thumb is that one millimeter is the practical boundary: above it, piezo or voice-coil hardware takes over. No robot joint decision ever lands here.

Force in micronewtons and travel typically 1–100 µm · drive voltage usually 10–100 V · switching in microseconds · holding power near zero, because electrostatic actuation is capacitive · thousands of die per wafer · die cost in cents, with the package and its high-voltage driver, not the die, setting what a usable part costs.

TI's DLP micromirror arrays (millions of mirrors per chip), MEMS autofocus and OIS (a growing phone niche), MEMS scanning mirrors in LiDAR and AR displays, RF-MEMS switches, microfluidic pumps and valves, Fabry-Perot tunable filters.

Mirrorcle Technologies — MEMS mirror actuators · Texas Instruments — DLP micromirror arrays

MEMS is a semiconductor business rather than a motion-control one, and value tracks design wins in consumer and automotive optics. For robotics the relevance is components rather than motion: MEMS mirrors inside solid-state LiDAR, micro-optics in sensors. It's included here to mark the boundary. Below a millimeter, "actuator" means silicon, and the supply chain is a fab.

9 minHow Digital Light Processing (DLP) worksApplied Science · 100k+ views

18 minDigital Micromirror Devices - in-depth operationtesla500 · 100k+ views

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

Mirrorcle MEMS Mirrors: Technical Overview (Mirrorcle Technologies, PDF) · Electrostatically Actuated MEMS Resonators: A Review (Microsystems & Nanoengineering via PMC)

Nitinol wire, deformed while cool, snaps back to its trained shape when heated through its phase transition, contracting ~4–5% like a muscle fiber. It does this silently, with no motor, gear, or magnet. Work density per gram is very high, and a hair-thin wire lifts hundreds of grams. Heating is easy, since you run current through the wire itself, which makes SMA the simplest possible electric "muscle": two wires and physics.

Force per mass is extraordinary, motion is silent and vibration-free, the wire is intrinsically compliant and cheap, and there's no mechanism at all. The weaknesses follow from the fact that SMA is a thermal engine. Efficiency is ~1–3%. Bandwidth is limited by cooling (a hertz at best, less when the wire is insulated). Hysteresis makes precise control hard. Fatigue shortens life at high strain. And holding a position costs continuous heat. SMA is good at occasional twitches and bad at cycling.

Use SMA where the stroke is sub-millimeter, the duty is light, and silence, low mass, or mechanism-free simplicity is what matters: camera modules, micro-valves, deployables and latches that fire rarely, medical mechanisms where a motor won't fit. If you keep strain low (well under the 4–5% maximum) and cycles gentle, the wire lasts millions of actuations for pennies. Avoid SMA for anything continuous, fast, or efficient. 1–3% efficiency, cooling-limited bandwidth of a hertz or so, and having to heat the wire just to hold position make it wrong for limbs and wrong for sustained duty. In those cases a voice coil or micro gearmotor is the honest answer. Be skeptical of any pitch promising limb-scale SMA muscle, because it's working against thermodynamics.

Contraction 4–5% at maximum, and well under that for long life · a 0.1–0.15 mm nitinol wire pulls roughly 1.5–3 N · actuation temperature typically 70–90 °C · efficiency 1–3%, because it is a thermal engine · bandwidth around 1 Hz, set by cooling · millions of cycles at low strain, far fewer near the strain limit · wire costs pennies at consumer volume.

Smartphone camera OIS/autofocus (Cambridge Mechatronics, SMA's one high-volume application, billions of units), medical stents and guidewires (superelastic cousins), aerospace deployables and louvers (Boeing's morphing chevrons), micro-grippers, animatronic and soft-robot research.

Cambridge Mechatronics — SMA actuators at smartphone volume · Memetis — SMA micro-valves and miniature actuators

Where strokes are sub-millimeter and duty is light, SMA has already won at consumer volume. The camera-module industry proves the wire can cost pennies and stay reliable for millions of cycles at tiny strain. As a limb-scale muscle, thermodynamics is the hard limit. A good diligence rule: trust SMA pitches at millimeter stroke and hertz bandwidth, and question everything bigger and faster.

10 minNitinol: The Shape Memory Effect and Superelasticityengineerguy · 500k+ views

5 minMagical metals, how shape memory alloys work - Ainissa RamirezTED-Ed · 100k+ views

3 minNitinol Wire/Shape Memory Alloy - How to Use itRimstarOrg · 100k+ views

AI Control Methodologies for SMA Actuators: A Systematic Review (Micromachines via PMC) · A Physics-Based Model of Hysteresis in SMA Wire Transducers (arXiv)

A McKibben muscle is a rubber bladder in a braided sleeve. Inflate it and the braid geometry converts radial expansion into axial contraction, up to ~25% strain, with force-per-weight far above what electric motors manage. It was invented for prosthetics in the 1950s (McKibben). The pneumatic muscle contracts, softens, and springs like the biological original, and it costs almost nothing to make. It remains soft robotics' default power source and the usual "muscle" in biomimetic projects.

Force density and inherent compliance are both excellent, the muscle is naturally safe around humans, it tolerates misalignment (it's a rope that pulls), it's water- and dirt-proof, and it's extremely cheap. The weaknesses: it needs compressed air, so the compressor, valves, and tether are the real system. Force fades nonlinearly as the muscle contracts. You need antagonistic pairs to move in both directions. Hysteresis and air compressibility make precision very hard. And bladder fatigue is a genuine wear item.

Use McKibben muscles when compliance and force-per-gram dominate, precision doesn't, and compressed air is already on site or a tether is acceptable: compliant tensioning and special machines on factory air, wearable and rehabilitation prototypes, bio-inspired research where muscle-like springiness is the point. Design in antagonistic pairs and treat bladder fatigue as a scheduled consumable. Avoid them on untethered robots, where the compressor, valves, and hoses are the real system and cancel the mass advantage. Avoid them for anything needing repeatable position or force, too, since hysteresis and air compressibility get in the way. In that case use electric QDD joints or plain cylinders running to stops.

Contraction up to about 25% of length · operating pressure 4–6 bar on plant air · peak pull roughly 600 N at 10 mm muscle diameter to 6 kN at 40 mm · force falls toward zero as contraction approaches its limit · the muscle costs a few dollars, and the compressor and valves are the real system cost.

Festo's Fluidic Muscle (the industrial productization) and its Airic's-arm demos, Shadow Robot's original air-muscle hand, countless bio-inspired legged and wearable prototypes, exosuit research (Harvard's soft exosuit lineage uses related pneumatics), Bridgestone's 1980s "rubbertuators."

Festo — the Fluidic Muscle, the industrial McKibben productization

The muscle itself is nearly free. The pneumatic infrastructure is the entire cost, and it's the reason no untethered robot uses them at scale. Commercially they persist in the niches Festo serves (compliant tensioning, special machines) and in research. The investable adjacency is the valve and micro-compressor problem: whoever makes soft-robot pneumatics battery-portable would bring this whole class back.

2 minMcKibben Artificial Muscle - opensoftmachinesOpen Soft Machines · 10k+ views

3 minHow to Build a McKibben Air MuscleXYZAidan · 100k+ views

2 minActive Textile made of Thin McKibben MusclesSuzumori Endo Robotics Laboratory · 10k+ views

A Review of Pneumatic Artificial Muscle Actuators: Force Model and Application (Actuators via Virginia Tech)

These are electrostatic muscles. Dielectric elastomer actuators (DEAs) squeeze a rubbery capacitor between compliant electrodes; apply kilovolts and it flattens and expands. HASEL actuators (2018, Keplinger lab) improved on the recipe by filling flexible pouches with liquid dielectric. The field zips the pouch shut and hydraulically displaces the fluid, which combines electrostatic speed with muscle-like contraction, a self-healing dielectric, and built-in capacitive self-sensing. This family is the most credible fully-electric artificial muscle so far.

Strain is muscle-like (10–30%+) at tens to hundreds of hertz, efficiency is high (the mechanism is electrostatic, and energy recovery is possible), operation is silent, and the actuator senses itself. It's made from films and liquids using processes friendly to roll-to-roll production. The weaknesses: it needs kilovolt drive electronics, and miniaturizing a safe 5–10 kV supply is half the product. Forces per unit are modest, so you need stacks and arrays. Dielectric breakdown and electrode fatigue set the lifetime. And no application beyond demos has shipped at volume yet.

Consider HASEL/DEA when the application needs silent, muscle-like motion at modest force with built-in self-sensing, and can absorb a 5–10 kV drive stage: haptic surfaces, small valves and pumps, adaptive optics, soft-robot research. Those are jobs where tens of hertz and 10–30% strain in a film-thin package beat any motor. Design around stacks and arrays from the start, since single units push newtons, not tens of newtons. Avoid these for load-bearing joints and for any product that needs demonstrated lifetime today, because breakdown and electrode fatigue data are still thin and nothing has shipped at volume. If the muscle-like form factor isn't itself the requirement, use conventional electromagnetic actuation.

Strain 10–30%+, close to biological muscle · bandwidth tens to hundreds of hertz · drive voltage 5–10 kV, and miniaturizing that supply is half the product · force of order newtons per unit, so stacks and arrays are the norm · built-in capacitive self-sensing · roughly $100–1k per axis at today's low volumes.

Artimus Robotics (HASEL commercialization), Toyota Research and academic humanoid-muscle demos, historical SRI DEA work (the field's origin), haptic-display startups using DEA films, adaptive-optics and valve prototypes.

Artimus Robotics — HASEL actuator commercialization · ESTAT Actuation — electroadhesive clutches built on the same electrostatic principle

Venture capital keeps circling this category, because the materials are cheap, manufacturing is film converting, and the performance envelope genuinely resembles muscle. The unresolved questions are lifetime under load and the HV-electronics bill of materials. A reasonable posture is to watch for the first boring industrial design win (valves, haptics, pumps). Muscles for humanoids come after the boring win, not before.

2 minHASEL actuators with muscle-like performanceRobotic Materials Department at MPI-IS · 1m+ views

4 minTechnical Deep Dive: Expanding (E-Series) HASEL ActuatorsArtimus Robotics · 5k+ views

2 minDielectric Elastomer Actuator - opensoftmachinesOpen Soft Machines · 10k+ views

HASEL Actuator Technology (Artimus Robotics) · Review of Electrohydraulic Actuators Inspired by the HASEL Actuator (Biomimetics via PMC)

These are two muscles made from string. Twisted-string actuators (TSA) spin a pair of cords with a tiny high-speed motor. Twisting shortens the bundle, which converts cheap RPM into high linear force. The result is a transmission that weighs grams, costs cents, and has a ratio that grows as it twists. Supercoiled polymer (SCP) muscles go further: nylon fishing line, twisted until it coils, contracts several percent when heated. That 2014 discovery turned sewing thread into a muscle with 100× the work density of biological tissue.

Force per gram and cost are both remarkable. TSA turns any micro-motor into a tendon puller, which is ideal for prosthetic hands and wearables. The weaknesses: TSA strings fatigue, and the nonlinear ratio complicates control. Travel is limited and speed is modest. SCP muscles inherit SMA's thermal problems (single-digit efficiency, cooling-limited hertz bandwidth) plus polymer creep. They demo well and are rarely used in real applications.

Use a twisted-string actuator when a gram-scale, cent-scale mechanism has to pull like a tendon: prosthetic and robotic fingers, exo-gloves, cable tensioners. Anywhere a tiny high-RPM motor plus two cords can replace a gearbox, TSA is worth a look, and the naturally rising ratio suits grasps that start fast and end forceful. Budget for string replacement as a wear item and handle the nonlinear ratio in your controller. Avoid TSA for long-travel or high-cycle continuous axes. Avoid SCP muscles in products entirely, since single-digit thermal efficiency and hertz-class bandwidth confine them to research until a non-Joule heating path appears. For controlled linear force at scale, a screw or voice coil is still the tool.

TSA travel typically 10–30% of free string length · a gram-scale motor at thousands of rpm gives tens to hundreds of newtons of pull, with the ratio rising as the bundle twists · SCP contracts a few percent under Joule heating · SCP work density roughly 100× biological muscle · SCP efficiency single-digit, bandwidth around 1 Hz · string and thread cost cents, and both are wear items.

TSA: prosthetic hands and exo-gloves in research (and consumer products quietly — some e-bike and cable-tensioning mechanisms), NASA tendon experiments; SCP: University of Texas/Baughman lab's nylon muscles, haptic and micro-robotics demos, textile-integrated actuator research.

Allonic — braided/twisted-fiber muscle actuators

TSA is an underrated engineering trick with real product wins available today in hands, grippers, and wearables, anywhere a gram-scale motor has to pull like a tendon. SCP remains materials-science upside without a thermal exit, so watch for photonic or chemical (non-Joule) heating breakthroughs before believing product claims. Together they illustrate the pattern in this field: the transmission innovations ship, and the thermal muscles wait.

4 minDesigning Anthropomorphic Robot Hand with Active Dual-Mode Twisted String Actuation MechanismMSC kaist · 100k+ views

7 minMaking artificial muscles from fishing lineBionicMuscles · 50k+ views

2 minFishing Line Artificial MusclesIntelligent Polymer Research Institute · 50k+ views

New Twist on Artificial Muscles (PNAS via PMC) · Data-Driven Twisted String Actuation for Dexterous Hands (Biomimetics via PMC)