Does "20 kg payload" ever mean 20 kg per arm while walking?

Almost never. Treat it as a dual-arm, static, peak, close-to-body figure until the supplier confirms otherwise in writing. The working single-arm number is usually a small fraction of the headline.

Why do two robots both say "12 DOF" but cost very different amounts?

Because DOF counts total joints, not driven ones. A 12-DOF fully-actuated hand has 12 motors and independent control; a 12-DOF hand with 6 actuators has coupled, underactuated fingers. Ask for actuated DOF, not total.

Is a robot with no IP rating waterproof at all?

No. An absent IP rating means the unit is unrated — treat it as indoor, clean, dry use only. Don't infer protection from a rugged appearance.

The demo video shows autonomous manipulation. Is that what I get?

Not necessarily. Much manipulation footage is teleoperated or scripted. Ask explicitly which, and for autonomous claims request a success rate over repeated trials in a described environment.

What's the single most important thing to confirm before paying?

For a platform build, that your purchased tier includes SDK and low-level control with English documentation. For a task deployment, the single-arm continuous payload and working runtime under load. Both belong on the proforma invoice — see the import playbook.

Are the numbers on this page verified?

No — figures marked† are manufacturer-reported and shown to illustrate how specs are published, not as endorsements. Confirm current configuration and pricing per unit via RFQ before you rely on any of them.

Reading a Humanoid Robot Spec Sheet: What "20 kg Payload" Really Means

Payload: the most-abused number on the sheet

"Payload" on a humanoid spec sheet is rarely qualified, and the qualifiers are where the truth lives. Four questions decompose it:

Continuous vs peak. Peak is a momentary rating; continuous is what the actuators can hold thermally over a shift. They can differ 2–3×.
Single-arm vs dual-arm. A "20 kg" headline is often the sum of both arms cooperating on one object held close to the torso. Split it and each arm may carry 5 kg or less at reach.
Static hold vs dynamic. Holding a mass with the robot standing still is not the same as carrying it while walking, where gait dynamics and balance margins eat into what the joints can spare.
At what reach. Payload collapses as the arm extends. A number quoted at the shoulder is meaningless for a pick 600 mm out.

For scale, published single-arm figures on research-class humanoids are modest: the Fourier GR-2 lists 3 kg single-arm payload†. When a maker claims a much larger whole-robot number, it is almost certainly a dual-arm, static, close-in peak.

What this means when you buy: the only payload that matters is single-arm, continuous, at your working reach, while the robot does what your task requires (standing, walking, or bimanual). Get that one number, in that framing, on the proforma invoice.

DOF counts: total vs actuated, and where they are

Degrees of freedom is a headline race, and the trick is total vs actuated. A joint can move (a DOF) without being driven by its own motor. Dexterous hands make this vivid:

Model	Published DOF	Actuated
Inspire RH56	12^†	6^†
Unitree Dex5-1	20^†	16^†
Robot Era XHAND1	12^†	12 (fully-actuated)^†

A 12-DOF hand with 6 actuators is a very different device from a 12-DOF fully-actuated hand — the first has coupled, underactuated fingers, the second controls each joint independently. The same gap appears on full humanoids: a "40+ DOF" headline may count both hands' passive linkages.

Just as important is where the DOF sit. A waist DOF, an actuated ankle, or wrist roll changes what tasks are reachable far more than raw count. The EngineAI PM01, for instance, publishes a 320° waist rotation† — one well-placed DOF that matters more than several in a gripper you won't use.

What this means when you buy: ask for actuated DOF, broken down by limb (legs / waist / arms / wrists / hands). Total DOF is a spec-sheet flex; actuated-and-located DOF is your kinematic envelope. For hands specifically, our dexterous hands guide breaks this down further.

Walk speed: demoed vs sustained

Top speed is almost always a demonstrated peak on flat, clean floor, tuned, often tethered for safety, for a few seconds. The Unitree H1 lists 3.3+ m/s demonstrated†; that is a highlight-reel number, not a duty cycle. Quadrupeds tell the same story — the Go2 publishes ≈3.5 m/s (EDU tunes higher)† — impressive, and unrelated to how fast it moves reliably across your floor for an hour.

What this means when you buy: ask for sustained speed on your surface (concrete, grating, ramp, whatever you have) carrying your payload, and the speed the shipped firmware allows out of the box. The demoed number and the number you'll operate at are different specs.

Runtime: idle vs working, and can you swap the battery

Runtime headlines are frequently idle or standby figures — powered, standing, computing little. Working runtime under motion and manipulation load can be a fraction of it. The AgiBot Yuanzheng A2 lists ≈2 h continuous†; quadruped endurance figures like the B2's 4–5 h / 15–20 km† are similarly best-case, walking lightly loaded.

The bigger operational question is hot-swap. A robot with swappable batteries runs indefinitely with a spare on charge; one with an internal pack forces a downtime block every cycle. That single fact often decides whether a humanoid fits a real shift.

What this means when you buy: ask for working runtime under representative load, whether batteries are hot-swappable, swap time, and price per spare pack. Then budget spares — and remember spare lithium packs ship under stricter rules than installed ones (see the import playbook).

Height and weight: why they gate your cell

Height and weight read like brochure trivia; they're actually integration constraints. Weight sets floor loading, whether a person can lift or catch it, and what fixturing and lifting gear you need. Height sets reach envelope, transport crate size, and doorway/ceiling clearance in your facility. Published humanoids span a wide band:

Model	Height	Weight	Actuated/total DOF (published)
Unitree R1	1.21 m^†	≈25 kg^†	26^†
Unitree G1	1.32 m^†	≈35 kg^†	23 (EDU to 43)^†
EngineAI PM01	1.38 m^†	40 kg^†	24^†
Fourier GR-2	1.75 m^†	63 kg^†	53 incl. 12-DOF hands^†
AgiBot Yuanzheng A2	1.69 m^†	≈69 kg^†	40+^†

A 69 kg robot is not a benchtop item — it needs a lifting plan, a rated stand, and safeguarding a 25 kg unit doesn't.

What this means when you buy: check height/weight against your bench, door, crate, and lifting gear before the deposit. A robot you can't safely move or fixture is a robot you can't use.

Autonomy vs teleoperation: read the caveat

This is the biggest gap between demo video and delivered capability. Manipulation footage that looks autonomous is frequently teleoperated or scripted — a human puppeteering, or a fixed trajectory replayed in a controlled scene. That is legitimate for showing hardware dexterity, but it is not autonomy, and the spec sheet rarely says which.

What this means when you buy: ask flatly — is this task autonomous, teleoperated, or scripted? For anything claimed autonomous, ask for the success rate over N trials, in what environment, with what perception. "Runs your task 9 times in 10 on a cluttered bench" is a spec; a flawless single take is not. Assume teleop until proven otherwise.

IP rating: is it even rated

Most research humanoids carry no IP rating at all — they are lab machines, not for dust or water. When a rating appears it's usually on ruggedised quadrupeds: the Deep Robotics X30 lists IP67†. Don't assume; the absence of an IP figure means unrated, and IP54 (splash) is very different from IP67 (immersion).

What this means when you buy: if your environment has dust, moisture, or washdown, require the IP rating in writing and confirm whether it covers the whole robot or just the torso. No stated rating = indoor, clean, dry only.

Research/EDU vs commercial: the SDK is the real spec

The same-looking robot ships in tiers, and the difference is access, not chassis. EDU/research configurations typically unlock the SDK, low-level joint control, and onboard compute; base tiers lock you to canned behaviours. This is priced accordingly — the Unitree G1 runs from $16,000 US list with EDU configs to ≈$74k†, and Go2 tiers explicitly decide SDK and compute access†. Same silhouette, very different developer surface.

What this means when you buy: if you're building on the platform, confirm the exact tier gives you low-level API access, torque/position control, and documented interfaces — in English — before payment. A cheaper unit that won't let you close a control loop is not the same product.

Spec term → common overstatement → question to ask

Spec term	Common overstatement	Question to ask the supplier
Payload	Dual-arm, static, peak, close-in	"Single-arm, continuous, at ___ mm reach, while walking?"
DOF	Total count including passive/coupled joints	"How many actuated DOF, and where are they?"
Walk speed	A few-second tuned demo peak	"Sustained speed on my surface, with payload, on shipped firmware?"
Runtime	Idle/standby figure	"Working runtime under load? Hot-swap? Swap time and spare price?"
Height/weight	Presented as trivia	"Confirmed crated dimensions and lifting/fixturing requirements?"
Autonomy	Teleoperated or scripted demo	"Autonomous, teleop, or scripted? Success rate over N trials, where?"
IP rating	Implied, or unrated	"Stated IP rating for the whole robot, or none?"
Configuration	EDU features shown, base tier sold	"Does my tier include SDK, low-level control, compute — documented in English?"

The pattern is the same every line: the sheet quotes a best case; you buy a worst case; the gap is closed only by a specific, written question. A supplier who answers these precisely is one worth quoting. Send the list with your RFQ.