Grab any marker to take over. Arrow keys nudge the focused joint by 3°.
Guided tourSeven steps down the page, in order. Each one says what that part is for and what you are meant to do with it.
Eleven actuator classes, twelve decision axes, and the arithmetic that tells you which one your joint actually needs. Start from the load case — not from the parts catalog.
Every selection error traces back to a missing load case. Work these four numbers before you look at a single datasheet: peak torque, RMS torque, output speed, and reflected inertia.
The gravity term dominates static arms; the inertial term dominates legs. For a 6 kg forearm with its CoM at 0.22 m, gravity alone asks 12.9 Nm at the elbow — before you accelerate anything. Add 30–50% margin for the worst-case pose you forgot to simulate.
Copper loss goes as I², so the duty cycle must be RMS-weighted, not averaged. Spec continuous rating ≥ τrms and peak rating ≥ τpeak. A joint that passes on peak and fails on RMS de-rates itself in the field — silently, via thermal foldback.
Ratio is the whole design decision in one number. A knee swinging 400°/s (66 rpm) on a 4000 rpm frameless rotor wants N ≈ 60 — harmonic territory. Cut the same joint to N = 6 and you need 10× the motor torque constant, i.e. a much larger diameter stator. Diameter buys torque; ratio buys it back at the cost of transparency.
The N² term is why high-ratio joints hit like a brick. A 12 g·cm² rotor behind N = 100 reflects 120 000 g·cm² to the output — often more than the limb itself, which makes impact energy go straight into the gear teeth. Keep Jreflected ≤ Jlink if the joint will ever hit the ground unplanned.
Two takes on the same ranking engine. Guided walks you through the load case in five questions. Workbench hands you the sliders and ranks live. Same scoring, different posture.
Sets the weighting: legs privilege impact and bandwidth, cobot joints privilege backlash and safety.
From step 1 of the sizing math: J·α + m·g·r·cos θ, plus margin. Rotary-equivalent for linear actuators at a 0.15 m arm.
Max joint velocity at the output flange. This is what kills high-ratio candidates.
Pick the one axis you would defend in a design review. It re-weights the ranking hard.
Unit cost at your build quantity, actuator + drive electronics. Support infrastructure (HPU, compressor) is called out separately in the results.
Fit score is a weighted sum over the twelve axes; hard flags mark physical or commercial infeasibility.
Specs are class envelopes across commercially available frame sizes, not single-part numbers. Named modules are representative, current-production examples. Classes carrying a vendor-verified badge have at least one figure checked against the manufacturer's own published datasheet, linked below the modules; the rest are engineering estimates until you check them.
{{ c.oneLiner }}
Scored 1–5 where 5 is always "better for you" — cheap, quiet, short-lead, tolerant. Read the rows, not the totals: no class wins on aggregate, and the one that wins your two critical rows is the answer.
The trade that defines legged robotics. Nothing occupies the top-right corner — high torque density and high transparency are structurally opposed, because both gearing and fluid power buy force by giving up backdrivability.
Pick up to three classes. Rows highlight the winner where the difference is decision-grade.
The shortcut path, for when you already know the load case and just want the conventional answer. Click through; the trail stays visible.
{{ treeWhy }}
{{ treeResultWhy }}
Pick a ratio and you have implicitly picked your bandwidth, your impact tolerance, your efficiency and your control architecture. Three derivations that make that explicit.
Reverse efficiency collapses non-linearly. At ηfwd = 0.90 you keep 0.89 backwards; at 0.70 you keep 0.57; below 0.50 the term goes negative and the joint is self-locking — it cannot be backdriven at all.
The fraction of impulse absorbed by the gearbox rises with N². This is the mechanism behind flexspline ratcheting on a dropped harmonic joint, and the reason a 6:1 QDD leg survives a fall that destroys a 100:1 one.
Mitigations, in order of cost: lower N, a series spring, a torque-limiting coupling, or a cycloidal set rated to 500% momentary.
Your achievable torque bandwidth is capped at roughly ωn/3. A stiff QDD joint puts its first resonance beyond 500 Hz, so current control is the only limit. Add a 300 Nm/rad series spring and ωn drops to tens of hertz — which is exactly the trade an SEA makes on purpose.
If your controller needs 200 Hz of impedance bandwidth, that requirement alone eliminates SEA, pneumatic, and SMA before any other consideration.
Continuous torque is not a mechanical rating — it is the torque at which winding temperature stabilises below the insulation class. Change the cooling and the same actuator gets a different datasheet.
Note the square. Running 1.4× the rated torque doubles the heat. Note also that Rth is where the design leverage lives: bolting the stator to an aluminium structural member instead of a plastic housing can move Rth,h-a by 3–4× and buy back 60–80% more continuous torque from a part you already own.
Torque constant also drifts: NdFeB magnets lose about 0.11%/K of remanence, so a 100 K rise costs you roughly 11% of Kt — and your current-mode torque estimate silently over-reads by the same amount unless you compensate.
{{ duty.verdictNote }}
Five joint archetypes, the conventional answer for each, and the credible dissent.
Dynamic, ground-contacting joints converge on low ratios and accept poor absolute torque. Positioning joints converge on high ratios and accept poor transparency. Almost every hard selection problem in robotics is a joint that is asked to do both.
The catalogue above ends at modules you buy whole. This is the layer under it: real parts, from vendors' own datasheets, for the five things a QDD joint is made of. Quasi-direct-drive first because it is what the matcher recommends most often; the other ten classes follow if this shape is right.
CubeMars RO100 KV55 behind a Neugart PLE080 at 7:1 gives 28 N·m continuous and 84 N·m peak at the output before efficiency, inside a gearbox rated for 65 and 104. Put an AksIM-2 MB080 on the output and an Everest NET behind it and you have a knee joint with load-side sensing. The soft spot is backlash: PLE080 is under 10 arcmin, which is honest for the class and loose for force control.
Nothing here has been built or measured by us. These are catalogue numbers, read on the date above, and vendors revise them. Pull the datasheet before you cut a PO — that is the same thing this guide says about every envelope on the page.
Run the matcher, then send us the load case. We put it in front of the people who actually make these joints and come back with who has it on the shelf, who could build it, and who says no.
Every envelope here is class-typical. Pull two real datasheets per shortlisted class, re-run the sizing math against their published curves, and only then cut a PO.
Run the matcher