Menu

The Gearbox Is the Joint

The Gearbox Is the Joint: Planetary vs Harmonic vs Cycloidal

Notes · Motors and motionNovember 19, 202618 min readSenova Systems Research

Planetary, harmonic, and cycloidal gearboxes compared on the four numbers that decide a robot joint, backlash, stiffness, efficiency, and cost, with a worked demo on a real arm.

The motor gets the glory. The gearbox does the job.

In the last article we chose the motor: torque decides, not RPM, and the answer was a BLDC with field-oriented control at the shoulder and the elbow, a stepper at the wrist. But a motor on its own cannot turn a robot joint. A BLDC spins at thousands of RPM and makes a few newton-metres of torque.

The shoulder of our bench arm needs 126.8 newton-metres at 15 RPM. Something has to trade the speed for the strength, and that something is the gearbox, and the gearbox you pick changes the joint more than the motor you pick.

Here is the plain version. A gearbox is a lever made of circles. A small gear turning a big gear is the same trade as a long spanner turning a tight bolt: the small gear spins fast and weak, the big gear turns slow and strong.

The gear ratio is how many times the small gear has to turn to turn the big gear once. A 100:1 gearbox means the motor turns a hundred times for every single turn of the joint, and, roughly speaking, the joint gets a hundred times the motor's torque and one hundredth of its speed.

That trade is simple. Everything else about the gearbox is not, and the three families of gearbox make the trade differently. A planetary gearbox splits the load across several small gears. A harmonic drive bends a thin steel cup to make the gears mesh. A cycloidal drive rolls a disc around pins. Each one wins on a different number, and each one loses on a different number, and the number it loses on is the one that will bite you, because it is the number you did not check.

The words, in plain English

Gearbox (reducer). The set of gears between the motor and the joint. It trades speed for torque. "Reducer" is the same thing, it reduces the speed.

Gear ratio (N). How many motor turns make one joint turn. 100:1 means the motor turns a hundred times per joint turn, and the joint gets roughly a hundred times the motor's torque.

Backlash. The free play in the gearbox, the tiny amount the joint can wobble when nothing is driving it. Imagine a steering wheel with slack in it: you turn the wheel a little and nothing happens at the road wheels. That dead zone is backlash, measured in arcminutes. One arcminute is one sixtieth of a degree.

Arcminute (arcmin). One sixtieth of a degree. A full circle is 360 degrees, which is 21,600 arcminutes. Backlash is usually quoted in arcminutes, and sometimes in arcseconds, one arcsecond is one sixtieth of an arcminute.

Torsional stiffness. How much the gearbox twists when you load it. A stiff gearbox barely twists; a soft one winds up like a spring. Stiffness is measured in newton-metres per radian, how much torque it takes to twist the gearbox by one radian (about 57 degrees). Higher is stiffer.

Torque density. How much torque a gearbox delivers for its size or weight. A high-torque-density gearbox is small and light for the torque it carries.

Backdrivability. Whether an outside force can turn the joint backwards through the gearbox. A backdrivable joint yields when something pushes it, which is what you want when the arm works beside a person, and what you do not want when the arm has to hold a load with the power off.

Reflected inertia. The motor's own spinning mass, as the joint feels it, multiplied by the square of the gear ratio. A 100:1 gearbox multiplies the motor's inertia by 10,000. That is why a high-ratio joint feels heavy and sluggish when something pushes it, and why a low-ratio joint feels light.

Strain-wave drive (harmonic drive). The gearbox family that bends a thin steel cup, the flexspline, to make the gears mesh. "Harmonic drive" is the trade name; "strain-wave" is the generic name for the same idea.

Flexspline. The thin-walled steel cup inside a harmonic drive. An elliptical plug (the wave generator) deforms it into an oval so its teeth mesh with a rigid ring (the circular spline). The flexspline is the part that fatigues.

Cycloidal drive (RV reducer). The gearbox family that rolls a disc with lobes around fixed pins. "RV" is the common trade name for the industrial version, which is a planetary front stage plus a cycloidal rear stage.

Quasi-direct drive (QDD). Not a gearbox family but an architecture: a big, high-torque motor with a very low gear ratio (roughly 6:1 to 10:1), so the joint stays backdrivable. The motor does most of the work and the gearbox does very little.

Precision life. How long a gearbox holds its quoted backlash and accuracy before wear opens it up. Not the same as the time until it breaks, a gearbox can keep turning long after it has stopped being precise.

The three families, and how each one works

Planetary: the load is shared

A planetary gearbox is the oldest idea in the list and the most forgiving. A central sun gear drives three to five planet gears, which orbit inside a fixed ring gear. The input is the sun; the output is the carrier that holds the planets. Because three to five planets share every load pulse, no single tooth carries the whole torque, which is why a planetary gearbox survives shock loads and torque spikes that would kill the other two families.

One stage gives a ratio of 3:1 to 10:1. Stack two stages and you reach 100:1 or more, at 90–97% efficiency per stage (92–94% for a two-stage unit). The price is backlash: a standard planetary has 3–15 arcminutes of free play, because the sun and the planets need clearance to mesh. Precision-ground planetary units reach 1–3 arcminutes, at a higher price.

Input speeds run 5,000–10,000 RPM. The cost index is 1.0, the planetary is the baseline everything else is priced against, and a planetary of the same frame typically costs 30–50% less than a harmonic drive.

Harmonic (strain-wave): zero backlash, paid for in stiffness

A harmonic drive has three parts: an elliptical wave generator, a thin-walled flexible cup called the flexspline, and a rigid circular spline. The wave generator deforms the flexspline into an oval, so its teeth mesh with the circular spline at two opposite points. Because the flexspline has two fewer teeth than the circular spline, each turn of the wave generator advances the flexspline by two teeth, which is how a single stage reaches 50:1 to 160:1.

The genius of the design is that many teeth are in mesh at once, and the mesh is preloaded, there is no clearance to take up, so there is almost no backlash. A typical harmonic drive reaches 1 arcminute or better, many models reach 0.5 arcminutes, and some are quoted within 10 arcseconds. That is why harmonic drives are the default for precision light-load joints: cobot arms, surgical robots, anything that needs sub-millimetre positioning.

The price is paid in three places. First, stiffness: the flexspline is a thin steel cup, and it twists under load. A harmonic drive has roughly a third of the torsional stiffness of a planetary gearbox of the same size, and a fifth of an RV reducer's.

Second, efficiency: 65–85%, because the flexspline is constantly being deformed. Third, life: the flexspline fatigues. Precision life is 8,000–10,000 hours, and the precision decays over time, a harmonic drive that starts at 0.5 arcminutes does not stay there.

And the input speed is limited to 3,000–4,000 RPM, because the flexspline fatigues faster at high speed. Cost index: 1.5.

Cycloidal / RV

A cycloidal drive rolls a disc with lobes around fixed pins. The disc is mounted on an eccentric shaft, so as the shaft turns, the disc wobbles and its lobes engage the pins one after another. Because many lobes are in contact at once, the load is shared across a large area, which is why a cycloidal drive has the highest torque density of the three families (2–4× a harmonic drive) and the highest torsional stiffness (roughly five times a harmonic drive's, and about 1.7× a planetary's).

The RV reducer, the industrial version, is a planetary front stage plus a cycloidal rear stage, which is how it reaches ratios of 30:1 to 190:1 with backlash under 1 arcminute (1–2 arcminutes per one source). Efficiency is 85–90%. Precision life is 20,000–30,000 hours, and, unlike the harmonic drive, the precision does not decay significantly over time.

The price is mass and cost: cost index 2.0, twice a planetary of the same frame, and the heaviest of the three. That is why RV reducers live in the base and shoulder joints of industrial robots, where the torque is largest and the arm's own weight is the enemy, and not in the wrist, where every gram counts.

The four numbers that decide

Every gearbox decision is a trade between four numbers, and you cannot have all four at once:

  1. Backlash, how much free play the joint has when unloaded. Decides the positioning accuracy when nothing is pushing on the arm.
  2. Torsional stiffness, how much the joint winds up under load. Decides the positioning accuracy when something is pushing on the arm, which is to say, always, because gravity is always pushing.
  3. Efficiency, how much of the motor's power reaches the joint, and how much becomes heat in the gearbox. Decides the motor size, the battery life, and whether the gearbox cooks itself.
  4. Cost and life, what the gearbox costs, and how long it holds its numbers. Decides whether the design is affordable and whether the accuracy survives the first year.

The table is the whole article in one place:

PlanetaryHarmonic (strain-wave)Cycloidal / RV
Backlash3–15 arcmin (precision 1–3)≤1 arcmin (many 0.5)<1 arcmin (1–2)
Torsional stiffness~3×10⁴ N·m/rad~1×10⁴ N·m/rad~5×10⁴ N·m/rad
Efficiency90–97% per stage65–85%85–90%
Ratio, single stage3–10:150–160:130–190:1
Input speed5,000–10,000 RPM3,000–4,000 RPM(see note)
Precision life15,000–20,000 h8,000–10,000 h20,000–30,000 h
Shock loadHighLowHighest
Cost index1.01.52.0

The stiffness figures are order-of-magnitude values from one comparison guide, not datasheet numbers, treat them as the shape of the trade, not as a specification. The exact stiffness of a real gearbox comes from its datasheet, and it varies with size.

Backlash is not the whole accuracy story

Here is the mistake the table invites: pick the harmonic drive, because its backlash is the smallest. Backlash is the error you can see on a datasheet, and it is the error that disappears when the joint is loaded in one direction. But a robot joint is never unloaded.

Gravity pulls on the arm, the payload pulls on the arm, and the arm's own weight pulls on the arm, and under that load, the gearbox twists. The twist is the wind-up, and for a harmonic drive the wind-up is much larger than the backlash.

The demo below runs the numbers on our bench arm, and the result is the article's central point: the harmonic drive's backlash is under 0.25 mm at the end of the arm, but its wind-up under the shoulder load is 10.8 mm, forty-three times larger. The planetary gearbox, with three to fifteen times the backlash, winds up only 3.6 mm. The RV reducer, with roughly the same backlash as the harmonic, winds up 2.2 mm.

That does not mean the harmonic drive is wrong. It means the question is not "which gearbox has the least backlash" but "which error dominates in my joint, and can I correct it?" Backlash is a dead zone, the controller cannot see it and cannot correct it, because the motor turns and the joint does not move.

Wind-up is a spring, the controller can correct it, if it knows the load and the stiffness, because the deflection is proportional to the torque. A joint with a known stiffness and a torque sensor can be more accurate than a joint with zero backlash and unknown wind-up. This is the difference between precision and accuracy: the harmonic drive is precise (it returns to the same place every time) but, under varying load, it is not as accurate as its backlash figure suggests.

Reflected inertia, and the fourth option

There is a second trade the table hides, and it is the one that decides whether the arm can work beside a person. When a motor spins behind a gearbox, the joint feels the motor's spinning mass multiplied by the square of the gear ratio:

J_reflected=J_motor×N²\text{J\_reflected} = \text{J\_motor} \times \text{N²}

A typical frameless BLDC rotor has an inertia of about 0.0008 kg·m². Behind a 6:1 gearbox, the joint feels 0.029 kg·m². Behind a 100:1 harmonic drive, it feels 8.0 kg·m², 276 times more. That is why a high-ratio joint feels heavy and sluggish when something pushes it, and why it barely yields to external force: the friction and inertia reflected through the gearing scale steeply with the ratio, so the joint is effectively rigid, which is dangerous if the robot contacts a person.

The fourth option is to refuse the trade. Quasi-direct drive (QDD) pairs a large-diameter, high-torque-density BLDC motor with a very low gear ratio, roughly 6:1 to 10:1, or no gearbox at all. The low ratio keeps the reflected inertia small, so the joint stays backdrivable: the controller can sense contact through the motor current and yield, which is why QDD is the default inside modern humanoids, Unitree's G1 and H1 ship with QDD-style joints, and the architecture comes from academic labs (MIT's Cheetah) that needed legs to survive foot strikes.

The price of QDD is that the motor has to do the work the gearbox used to do. A low ratio means the motor needs far more torque, which means a bigger motor, more current, and more heat, and it needs a high-resolution encoder (typically 17-bit to 23-bit) and fast control loops (1–4 kHz) to hold position without the gearbox's help. Backdrive torque with no power is 0.3–1.5 N·m for a QDD joint against 5–25 N·m for a high-ratio harmonic joint.

QDD is the right answer when backdrivability, dynamic control, and impact survival matter more than holding position with minimum current. It is the wrong answer when the joint has to hold a heavy load still for a long time on a small battery.

The selection framework

The choice is per joint, not per robot, the same lesson as the motor article, and for the same reason. The joints do different jobs:

  • Shoulder and base (30–120 N·m): the torque is largest, the arm's own weight is the enemy, and the load is mostly gravity, not precision. RV/cycloidal, or a two-stage planetary. The RV wins on stiffness and life; the two-stage planetary wins on cost and survives shock. A harmonic drive at the shoulder is the wrong call, its stiffness is the limiting factor, and its flexspline fatigues under the repeated peak loads.
  • Elbow and wrist (8–30 N·m): the torque is moderate and the precision matters. Harmonic drive, for the near-zero backlash and the single-stage ratio. This is where the harmonic earns its price.
  • Gripper and small axes (2–10 N·m): the torque is small and the cost matters. Planetary, cheap, fast, and the backlash does not matter at this scale, or QDD where the gripper has to feel what it holds.
  • Legs and impact joints: QDD, or RV. The leg has to survive a foot strike, and a harmonic flexspline does not like shock loads.

And one more lever, which is cheaper than any gearbox: where the encoder sits. A joint with an output-side encoder, the encoder on the joint side of the gearbox, not the motor side, tolerates 2–3× more gearbox backlash for the same absolute accuracy, because the controller sees the joint's real position instead of inferring it through the gearbox. That single choice can move a design from harmonic to planetary economics: the planetary's backlash stops mattering, because the loop measures past it.

What the output encoder cannot fix is the wind-up, the joint really does deflect under load, so the stiffness question stays.

Example Labs is a made-up company we use for worked examples.

the Example Labs bench arm, through three gearboxes

The arm is the same one from the motor article: two links, each 0.425 m long, Link 1 weighing 4 kg, Link 2 weighing 3 kg, a 5 kg payload at 0.85 m reach, horizontal worst case. The shoulder needs 126.8 N·m with the 1.5× safety margin; the elbow needs 40.7 N·m; the wrist needs 11.0 N·m. Joint speed at the shoulder is 15 RPM, 90 degrees in one second.

Step 1: the motor torque each gearbox needs

The motor torque is the joint torque divided by the ratio and the efficiency: τ_motor = τ_joint ÷ (N · η).

Gearbox (100:1, shoulder)EfficiencyMotor torqueMotor speed
Planetary, two-stage93%1.36 N·m1,500 RPM
Harmonic80%1.59 N·m1,500 RPM
RV / cycloidal87%1.46 N·m1,500 RPM

The harmonic costs the motor 17% more torque than the planetary, for the same joint, the missing 13% of efficiency becomes heat in the flexspline. At the elbow (40.7 N·m through 50:1), the planetary needs 0.88 N·m from the motor and the harmonic needs 1.02 N·m. At the wrist (11.0 N·m through 10:1, single-stage planetary at 96%), the motor needs 1.15 N·m.

Step 2: the backlash, at the end of the arm

One arcminute at 0.85 m reach is 0.25 mm (the arm's reach times the tangent of one arcminute). So:

GearboxBacklashEndpoint error at 0.85 m
Planetary, standard3–15 arcmin0.74–3.71 mm
Planetary, precision1–3 arcmin0.25–0.74 mm
Harmonic<1 arcmin<0.25 mm
RV / cycloidal1–2 arcmin0.25–0.49 mm

The standard planetary's worst case, 3.71 mm, is the number that scares people into buying harmonic drives. But it is the unloaded number, and the joint is never unloaded.

Step 3: the wind-up, under the shoulder load

The twist is the torque divided by the stiffness, and the endpoint error is the twist times the reach:

GearboxStiffnessTwist at 126.8 N·mEndpoint error at 0.85 m
Harmonic~1×10⁴ N·m/rad0.727°10.8 mm
Planetary~3×10⁴ N·m/rad0.242°3.6 mm
RV / cycloidal~5×10⁴ N·m/rad0.145°2.2 mm

The harmonic drive's wind-up is 10.8 mm, forty-three times its backlash, and three times the planetary's wind-up. The gearbox with the smallest backlash has the largest error under load. This is the point of the article: backlash is the error you can read off a datasheet, and wind-up is the error that actually moves the end of the arm.

The stiffness figures are order-of-magnitude values from one comparison guide, so treat the exact millimetres as illustrative, but the ranking is the point, and it holds across every source checked: the harmonic is the softest of the three, the RV the stiffest, the planetary in between.

Step 4: the cost of four joints

Gearbox cost, anchored on shipping hardware: an industrial-grade harmonic drive runs about $400 per joint (¥3,000+ each). A planetary of the same frame costs 30–50% less, $260 at the midpoint. An RV reducer costs twice a harmonic, $800, from the cost index of 2.0.

Build (four joints)Gearbox costWith motors + drivers ($140/joint)
All planetary$1,040$1,600
All harmonic$1,600$2,160
All RV$3,200$3,760
Mixed (RV shoulder, harmonic elbow + wrist, planetary gripper)$1,860$2,420

The mixed build costs $1,340 less than the all-RV build and puts the expensive gearbox only where its stiffness earns it, at the shoulder, where the load is largest. The all-harmonic build is the trap: it spends the money on precision at every joint, including the gripper, where the backlash never mattered.

Step 5: the fourth option at the wrist

What if the wrist used quasi-direct drive instead, a big motor with a 10:1 planetary? The wrist needs 11.0 N·m, so the motor needs 11.0 ÷ (10 × 0.93) = 1.18 N·m, well within a frameless BLDC, and the joint becomes backdrivable, so the gripper can feel what it holds. The reflected inertia is 100× the rotor inertia instead of 10,000×.

The shoulder cannot do the same trick: 126.8 N·m through 10:1 needs 13.6 N·m from the motor, continuously, and a frameless motor of the size that fits this arm sustains 4–6 N·m and bursts to 15–20 N·m for a few seconds. QDD at the shoulder would need a motor three times bigger, which would not fit the joint. The ratio is the decision: QDD where the torque fits the motor, a gearbox where it does not.

The decision test

Four questions, in order. If you cannot answer the first, the rest do not matter.

  1. What is the joint's worst-case torque, with the margin? Size from the full load, payload plus the arm's own weight, arm horizontal, with a 1.5–2× safety factor. If you have not done this arithmetic, you are choosing a gearbox by feel.
  2. Which error dominates, backlash or wind-up? If the joint is lightly loaded and the precision matters, backlash dominates and the harmonic wins. If the joint carries a heavy load, wind-up dominates and the harmonic loses. Run both numbers, as in the demo.
  3. Does the joint have to yield? If a person can touch it, or it can hit something, backdrivability is a safety requirement, not a preference, and that points to QDD or a low-ratio planetary, not a 100:1 harmonic.
  4. Can an output-side encoder move you down a price class? If the controller can see the joint's real position, a planetary's backlash stops mattering. Check the encoder before you check the gearbox catalogue.

Objections

"But the harmonic drive has zero backlash." It has near-zero backlash, and the smallest backlash of the three. It also has the lowest stiffness, and under load the wind-up is forty-three times the backlash. Zero backlash is a true statement about an unloaded gearbox. The joint is never unloaded.

"But the planetary is too sloppy for a robot." A standard planetary is too sloppy for a precision joint, yes, 3.71 mm at the end of the arm. A precision planetary (1–3 arcminutes) is 0.25–0.74 mm, and with an output-side encoder the backlash stops being the limiting error at all. The question is which planetary, and where the encoder sits.

"But the RV reducer is the best, so just use it everywhere." It is the stiffest and the longest-lived, and it is twice the price of a harmonic and three times the price of a planetary, and the heaviest of the three. At the wrist, its stiffness earns nothing, the load is small, and its mass costs you payload. The best gearbox is the one that earns its price at that joint.

"But QDD is the future, so why use gearboxes at all?" QDD is the future for joints where the torque fits the motor. At the shoulder of this arm it does not, the motor would need 13.6 N·m continuously, and the motor that delivers it does not fit the joint. The gearbox is not going away; it is the thing that lets a small motor do a big joint's job.

Sources

  1. Beitto Driver, "Harmonic Drive vs Planetary Gearbox: Choosing the Right Robotic Joint Reducer"
    What we used

    Backlash 1 arcmin or better (many models 0.5), planetary 3–15 arcmin, torque density up to 3×, cost 30–50% lower for planetary, input speed 3,000–4,000 vs 5,000–10,000 RPM, efficiency 70–85% vs 90–97%, harmonic trades stiffness for zero backlash

  2. Repunite, "Robot Joint Modules: The Complete Guide to Selection and Integration"
    What we used

    Harmonic <1 arcmin, planetary 5–15 arcmin, ratios 50:1–160:1 vs 3:1–100:1, efficiency 65–85% vs 90–97%, base joint 50–150 N·m, wrist 10–30 N·m, output speed 20–100 RPM harmonic vs 50–300 RPM planetary

  3. Repunite, "Harmonic vs Cycloidal vs Planetary: Engineering Guide to Gearbox Selection for Robot Joints"
    What we used

    Backlash harmonic 0–20 arcsec, RV <1 arcmin, planetary 1–8 arcmin; torque density RV highest (2–4× harmonic); stiffness RV very high, harmonic moderate; efficiency 70–90% / 85–95% / 90–97%; joint classes hip/knee/shoulder 30–120 N·m → RV or two-stage planetary, elbow/wrist 8–30 N·m → harmonic; output-side encoder tolerates 2–3× more backlash

  4. RobotWorld, "The Complete Robot Joint Module Selection Guide"
    What we used

    Stiffness ~1×10⁴ / ~3×10⁴ / ~5×10⁴ N·m/rad (harmonic / planetary / RV); precision life 8k–10k h / 15k–20k h / 20k–30k h; cost index 1.5 / 1.0 / 2.0; RV precision does not decay over time; harmonic ¥3,000+ per joint, industry target ≤¥1,500 per joint; standards GB/T 30819-2014, GB/T 37718-2019, GB/T 41400-2022

  5. Beitto Driver, "Harmonic vs Planetary Gearboxes for Humanoid Robots: Hip vs Wrist Joints"
    What we used

    Harmonic single stage 30:1–160:1, planetary 3:1–10:1 per stage, standard backlash 3–10 arcmin with precision versions approaching 1 arcmin, peak hip torque 1.5–2.5 N·m per kg body mass

  6. QDD Actuator, "What Is a QDD Actuator?"
    What we used

    Reflected inertia J_reflected = J_motor × N²; 8318 frameless motor rotor inertia ~0.0008 kg·m²; 6:1 → 0.029 kg·m², 100:1 → 8.0 kg·m²; backdrive torque 0.3–1.5 N·m (QDD) vs 5–25 N·m (high-ratio harmonic); 8318 sustains 4–6 N·m continuous, bursts 15–20 N·m

  7. RobotWale, "Quasi-Direct-Drive Motors: The Backdrivable Joint Architecture"
    What we used

    QDD ratio 5:1–10:1, needs 17–23-bit encoders and 1–4 kHz control loops, Unitree G1/H1 ship QDD-style joints, MIT Cheetah origin

  8. Custom Humanoid Actuator, "QDD Actuator Design for Bipedal Humanoid Legs"
    What we used

    100:1 multiplies rotor inertia by 10,000, 10:1 by only 100; low-ratio planetary backlash often 3–10 arcmin; QDD most useful where backdrivability, dynamic control, and impact survival matter more than holding position with minimum current

  9. Article #20, "Choosing the Motor: Torque Decides, Not RPM" (Senova, 2026-11-19)
    What we used

    Shoulder 126.8 N·m, elbow 40.7 N·m, wrist 11.0 N·m with 1.5× margin; 1 arcmin at 1 m ≈ 0.29 mm; BLDC motor + driver $140 per joint; planetary 92–94% two-stage, harmonic ~80% at rated load, cycloidal/RV 80–90%

More in motion and sensing

The Motor That Doesn't HumYour motor hums because the drive is switching in blocks, and the torque dips at every 60° boundary. Field-oriented control holds the current exactly 90° ahead of the rotor flux, so the ripple falls from about 14% to under 1%, and the harmonics stop becoming heat. Here is the arithmetic, and the one place trapezoidal still wins.27 minThe Shape of the MoveYour arm arrives on time and then shakes for two seconds. The move was never the problem, the shape of the move was. Trapezoidal is faster on paper; S-curve is faster on the machine, and here is the arithmetic that proves it.19 minLiDAR vs Time-of-FlightLiDAR and time-of-flight both measure distance by timing light, but they are not the same sensor. Direct ToF times each laser pulse; indirect ToF measures a phase shift. One reaches 200 metres, the other reads a room at 240 Hz. This is how to choose by range, accuracy, sunlight, and surface, with Example Labs' warehouse robot as the worked example.23 minSensor Fusion and the Kalman FilterOne sensor is never enough: GPS is honest but slow, an IMU is fast but drifts 15 metres in 30 seconds. A Kalman filter is a running best guess that weighs each reading by how much it trusts it, an idea from 1960 that flew Apollo to the Moon on 2k of RAM.11 minChoosing the MotorPicking a motor for a robot joint or a drive? Torque decides, not RPM. A gearbox trades speed for torque, so size from the load first, then choose between stepper, BLDC with FOC, servo and brushed DC, joint by joint.24 min