How to read a humanoid robot spec sheet in 2026: DOF, payload, runtime, TOPS and safety

This guide explains every headline number on a humanoid robot spec sheet: degrees of freedom, payload, walking speed, runtime, battery, torque, sensors, compute, IP rating and safety standards. For each one it shows the typical range across the 87 humanoid records in the RankedRobot database, and how makers stretch the figure or leave it out. Use it before you compare two robots side by side or repeat a maker's claim.

Last checked 1–9 October 2026. Spec-based analysis from maker-published figures; we have not tested these robots hands-on.

Start with the configuration, not the number

Most humanoids are sold in several editions, and the spec sheet often mixes them. One robot can be listed with a different weight, joint count, processor and battery for each edition. In our database, 13 humanoids state more than one total degrees-of-freedom (DOF) figure depending on the edition, and several more list two or more weights or compute options.

So the first question for any spec is: which build does this number describe, and is that the build I would receive at the quoted price? The Unitree G1 is a good example. Its product page lists 23 DOF for the base robot, and 23 to 43 for the EDU version once the optional three-finger hands, extra waist joints and wrist joints are added. Arm payload is about 2 kg on one and about 3 kg on the other. You will see both numbers quoted as "the G1".

Degrees of freedom: subtract the hands first

A degree of freedom is one independently driven joint axis. A human-style leg usually has six (three at the hip, one at the knee, two at the ankle). The headline DOF total is the most quoted humanoid spec, and it is also the easiest to inflate, because a dexterous hand can add 6 to 22 joints per side.

Makers rarely say whether their total includes the hands. The quickest check is arithmetic. Take the total, subtract twice the per-hand figure, and see what is left for the body.

The result is striking. Once the hands come out, 13 maker-stated full-size biped builds land between 29 and 34 body DOF. The bodies are much more alike than the headlines suggest. 1X confirms this on its own NEO page, which breaks the 75 down into 44 hand joints plus 31 for arms, neck, spine and legs. That is the same body count as LimX Oli's headline 31, which leaves the hands out.

Edition ladders show the same thing. Booster's T1 page lists 23 DOF for the base robot, 31 with grippers and 41 with dexterous hands. Moving from 23 to 31 also takes each arm from 4 to 7 joints, so the step is not only about the hands. EngineAI's T800 goes from 25 DOF (Basic and Open Source, no hands) to 43 and 46 (Pro and Max, 7-DOF hands).

What per-hand DOF tells you. Across the 34 records that state a per-hand figure, the values fall into rough tiers:

• 1: a simple parallel gripper (Pollen Reachy 2, Aldebaran NAO).
• 6 to 7: a basic dexterous hand, usually with coupled fingers (Booster, Pudu, LimX, Unitree Dex3-1).
• 11 to 13: a more capable hand (Atlas, Fourier, Kepler).
• 16 to 22: research-grade dexterity (Sanctuary Phoenix 21, Sharpa D01 22, 1X NEO 22).

More hand joints do not mean a better grasp. Ask what grip force and fingertip sensing the hand has. Very few makers publish grip force. Reachy 2's gripper is rated at 10 N, and it is the only grip-force figure in our records.

Payload: four different numbers that share one word

"Payload" on a humanoid sheet can mean any of these:

1. Per-arm payload: what one arm can hold, usually at a favourable pose.
2. Carry payload: what the robot can carry in both arms while walking.
3. Peak or instant lift: a short burst, not something it can sustain.
4. Repeated payload: what it can handle cycle after cycle in a real task. This is rare and the most useful.

Of our 87 records, 55 state some payload figure: 38 give a per-arm number, 28 a carry number, and only 11 give both. Just three publish a separate peak figure next to a sustained one.

Three things stand out.

First, the gap between figures for the same robot can be very large. NEO's peak lift is about 8.6 times its per-arm payload. AgiBot's X2 page gives 3 kg "in specific postures" and 1 kg or less across the arm's full range, both measured without the end effector. Unitree states the problem plainly in a footnote on its G1 page: "The maximum load of the arm varies greatly under different arm extension postures."

Second, peak figures end up quoted as capacity. Unitree's H1 page lists the H1-2 arm at about 21 kg peak and about 7 kg rated. Our record currently holds the 21 kg figure under per-arm payload, so read it as a peak, not a working load. We have flagged it for correction.

Third, a repeated figure is worth more than a bigger one-off figure. Agility's Digit 5 announcement describes 22.7 kg as a load the robot can lift again and again, and ties it to single-person lift tasks. That is a claim you can test in a pilot.

Typical ranges. The median maker-stated per-arm payload is 3 kg for bipeds (22 records) and 5 kg for wheeled humanoids (9 records). The median maker-stated carry payload is 20 kg (19 records). One reported carry figure of 160 kg for the RobotEra STAR1 comes from media coverage rather than the maker, and is far outside every other number we hold.

Walking speed: the lab figure and the working figure

Across 26 bipeds with a maker-stated top walking speed, the median is 1.68 m/s. The range runs from 0.4 m/s (Booster K1, the speed used for its battery test) to 4 m/s (DEEP Robotics DR02). Unitree labels the H1's 3.3 m/s a world record.

Top speed is a short-burst figure. AgiBot's X2 store page is a rare case of a maker printing all three figures: up to 2 m/s measured in the lab, 1.8 m/s as the general maximum, and typically 0.8 m/s or less. For planning a task, the typical figure is the one that matters. If a sheet gives only one speed, assume it is the best case.

Wheeled humanoids list a drive speed instead, and it does not compare directly with walking speed. In our records, the Galbot G1 and the Rainbow RB-Y1 each give 1.5 m/s.

Runtime and battery: ask "doing what?" and "how quickly back to work?"

A runtime without a task attached tells you little. A few makers do attach one:

• Booster's T1 gives about 2 hours walking and about 4 hours standing.
• The K1 gives 30 minutes (2 Ah pack) or 80 minutes (5 Ah pack), both walking at 0.4 m/s.
• AgiBot rates the X2 at about 2 hours walking at 0.5 m/s, and adds that real runtime depends on the task.

Standing is cheaper than walking, so a runtime with no task attached is likely a light-duty figure.

Body type makes the biggest difference. The median maker-stated runtime is 120 minutes for bipeds (30 records) and 480 minutes for wheeled humanoids (9 records). A wheeled base does not spend energy balancing, and it has room for a bigger pack.

You can check a runtime claim with the battery size. Divide watt-hours by hours of runtime to get the average power the claim implies:

Most bipeds land between about 200 and 480 W. The AgiBot A3's ten-hour claim implies about half the draw of the next-lowest biped. That may be possible, but the maker does not say what the robot is doing for those ten hours, so ask before you plan a shift around it.

Then look at the time to get back to work. Agility gives Digit 5 a 90-minute runtime but a 9-minute charge, which it calls a 10:1 run-to-charge ratio. 1X quotes NEO's quick charge as 6 minutes per hour of runtime. Other robots swap batteries instead: 22 records say the battery is hot-swappable. Figure states that the Figure 03 battery is not swappable, but that it is certified to UN38.3 for transport. A robot that runs 90 minutes and recharges in 9 can do more useful work than one that runs 4 hours and needs 2 to recharge.

Height and weight

The median maker-stated biped is 1.71 m tall (38 records) and weighs 57.5 kg (36 records). Heights run from 0.95 m (Booster K1) to 1.9 m (Atlas), and weights from 19.5 kg (Booster K1) to 129 kg (Digit 5). Wheeled humanoids are heavier, at a median of 92.5 kg across 11 records, because the base and the bigger battery add mass.

Two caveats apply. Weight often changes by edition (the T800 is listed at 75–85 kg, the Rainbow RB-Y1 at 131 or 170 kg), and sheets rarely say whether the hands and battery are included. Several wheeled designs also have height-adjustable torsos and give a range. The AgiBot G2, for example, is listed at 1.225–1.795 m. For door frames, shelving and stairs, the range matters more than the headline height.

Actuators and joint torque: the spec most sheets leave out

None of our 87 humanoid records holds a joint-torque figure. Where makers do publish torque, it is usually one peak number:

• Unitree gives the G1's maximum knee torque as 90 N·m (120 N·m for EDU), and notes that this is the largest joint motor, not every joint.
• The H1 page lists about 360 N·m at the knee, 220 N·m at the hip and 59 N·m at the ankle. It also lists a peak torque density of 189 N·m/kg and names the motor as an internal-rotor permanent-magnet synchronous motor.
• Booster lists a "max peak torque" of 130 N·m for the T1 and 60 N·m for the K1, without saying which joint.

A peak torque figure does not tell you how long the joint can hold it. Continuous (rated) torque, which is limited by motor heating, decides whether a robot can hold a box at chest height for a minute or squat repeatedly. If a supplier gives only peak torque, ask for continuous torque per joint, the speed at which it was measured, and how long the joint can run before it overheats. The same peak-versus-rated gap that inflates arm payloads also applies here.

Sensors: count the cameras, then check for touch

Cameras are almost always listed (59 of 87 records), and LiDAR in 25. Force/torque sensing (6 records) and tactile sensing (7 records) are much rarer. These matter most for contact-heavy work, such as inserting parts or handling soft items.

The best sheets itemise everything. Galbot lists four torso RGB cameras, two wrist depth cameras, a 3D LiDAR, two IMUs, eight microphones and two six-axis wrist force/torque sensors on the G1. The weakest only list sensor types: Pudu's D9 announcement mentions visual, tactile, force and auditory sensing with no counts. For touch, look for a measurable figure. Figure says Figure 03's fingertips detect forces as small as 3 g. Sharpa's D01 e-skin is specified at 0.1–20 N per sensing point with 0.2 N resolution.

Compute: TOPS describes the chip, not the robot

Twenty-three records quote an AI-performance figure. They range from 6 TOPS (Fourier GR-2) to 2,250 TOPS (XPENG IRON, three in-house chips). Ten of the 23 match the headline rating of an off-the-shelf NVIDIA module. NVIDIA's Jetson module page lists AGX Orin at up to 275 TOPS, Orin NX at up to 157 TOPS, and the Thor series at up to 2,070 FP4 TFLOPS. That last number explains why 1X quotes NEO at "2070 TFLOPS" while Unitree's H2 and MagicLab's MagicBot X1 quote "2070 TOPS". It is the same chip-level figure under two different labels.

Three cautions apply:

• The figure is the module's maximum, at low numeric precision, and usually at its highest power setting. NVIDIA rates the Thor series at 40–130 W, a large share of the 210–480 W total draws in the battery table.
• The same module can appear with different numbers. PNDbotics quotes 100 TOPS for its Orin NX 16GB build, while AgiBot and LimX quote 157 TOPS for Orin NX.
• TOPS says nothing about the software. A robot with a small processor and a well-tuned policy can beat one with a large processor and an early-stage policy.

IP rating: check which part of the robot it covers

IP codes come from IEC 60529. The first digit rates protection against solids and dust, the second against water. Only 6 of our 87 humanoids publish one: Atlas IP67, DR02 IP66, Galbot G1 IP54, AgiBot G2 IP42, Spirit AI Moz1 IP40, and 1X NEO.

NEO shows why the detail matters. Its page lists IP68 for the hands (submersible) and IP44 for the body (splash-proof). Quoting "IP68" for the robot would be wrong. Always ask which part a rating covers. Also ask whether it was tested by a third party or self-declared.

Safety standards: what exists in October 2026

• ISO 10218-1:2025 covers safety requirements for industrial robots, with integration in Part 2. It is written for conventional industrial robots and does not address machines that must actively balance.
• ISO 13482:2014 covers personal care robots: mobile servants, physical assistants and person carriers. ISO marks it for revision, and a replacement draft (ISO/FDIS 13482) is in progress.
• ISO/CD 25785-1 is the work item aimed at humanoid-type machines. Its title covers dynamically stable industrial mobile robots, whether legged, wheeled or otherwise. It applies to robots that need active control to stay upright, in workplaces where the public is excluded or restricted. It is still a committee draft: its comment period closed on 8 July 2026. A Part 2 on integration will follow. Agility calls it the first international safety standard for the humanoid category and says it contributes to it. No robot can claim conformity to it yet.
• ISO 13849-1:2023 defines performance levels (PL a to e) for safety-related control parts. A PL applies to a specific safety function, so "PLd-level safety protection", as on the AgiBot A2 page, should prompt the question: which function?

What makers actually state is thin. Seven records carry any certification entry:

• CE marking under the Machinery and Radio Equipment Directives plus EN 18031 for the AiMOGA Mornine M1, per the maker's press release.
• IEC 62368-1 for NAO.
• UN38.3 for Figure 03's battery.
• An unnamed "Certificate of Conformity" for the AgiBot X2.
• Safe torque off for the Spirit AI Moz1.
• IP42 listed as a certification for the AgiBot G2.
• For Digit 5, Agility claims an independent field evaluation against OSHA standards on a customer line, and a CE mark that is expected, not granted.

Choosing by use case

• Research and teaching. Compare the exact edition you can afford. DOF, compute and SDK access often change by edition (the Unitree G1's SDK is EDU-only). Look at body DOF after the hands, published joint torque, and open-source tooling.
• Factory or warehouse pilot. Prioritise repeated payload, run-to-charge ratio or hot swap, an IP rating for the whole body, and written safety evidence tied to named standards. Ask how the robot will be assessed once ISO 25785-1 is published.
• Home. Look at body IP rating, noise, hand safety and how much remote human help the robot relies on. Eighteen of our records include some statement about teleoperation, either for training data or for live assistance.

To compare robots field by field, use /compare/ or narrow the list with /finder/. For ranked shortlists, see best humanoid robots and our humanoid buying guide. For a ranking of advanced robots across categories, see Robot500.

What the numbers say (Analysis)

• Hands drive the DOF headline. Once hands are subtracted, 13 maker-stated full-size biped builds fall between 29 and 34 body DOF. Their headline totals range from 41 to 76.
• Payload reporting is patchy. 55 of 87 records give a payload figure, but only 11 give both per-arm and carry figures, and only 3 separate peak from sustained.
• Legs cost runtime. The median maker-stated runtime is 120 minutes for bipeds (n = 30) against 480 minutes for wheeled humanoids (n = 9).
• Battery and runtime figures together imply 115–475 W of average draw across the 9 bipeds that publish both. The 115 W outlier (AgiBot A3) is the claim most worth checking.
• Compute claims cluster on NVIDIA module ratings. 10 of the 23 AI-performance figures equal NVIDIA's 275, 157 or 2,070 headline numbers.
• The safety and durability fields are the emptiest. 6 of 87 records give an IP rating, 7 give any certification, and none give a joint-torque figure. Across the whole humanoid dataset, 976 of 1,193 recorded values (82%) are maker-stated. The rest are reported by media, estimated or unconfirmed, and are labelled as such on each robot page.

Frequently asked questions

Does a higher DOF count mean a more capable humanoid?

Not on its own. Most of the difference between headline totals comes from the hands. Subtract twice the per-hand figure, then compare the body DOF and the hand quality, such as tactile sensing and grip force.

Is per-arm payload the same as how much the robot can carry?

No. Per-arm payload is usually measured at one arm's best pose. Carry payload is for both arms while walking. Peak lift is a short burst. 1X NEO's figures, for example, are 8.16 kg, 24.95 kg and 69.853 kg.

Why do two robots with the same NVIDIA chip quote different TOPS?

Makers may quote different power modes, precisions or ratings for the same module, and some label FP4 TFLOPS as TOPS. Ask which module and which setting was used for the figure.

How long do humanoid batteries really last?

Maker-stated biped runtimes cluster around 2 hours, and wheeled models around 8. Few makers say what task the figure assumes. Booster and AgiBot are exceptions and give walking speeds or states. Charge time and hot swap matter as much as runtime.

Is any humanoid certified to a humanoid safety standard?

Not yet. ISO 25785-1 was still a committee draft as of October 2026. Some makers cite related certifications, such as CE marking, battery transport certification or performance levels for specific functions.

What does IP67 on a humanoid mean in practice?

Under IEC 60529, it means a high level of protection against dust and water. Check whether the rating covers the whole robot or one part. 1X NEO's hands are rated IP68, but its body is IP44.

Methodology

Figures come from the 87 humanoid records in the RankedRobot database. Each value is tied to a source URL and labelled maker-stated, reported, estimated or unconfirmed on its robot page. Medians and counts use maker-stated values only unless we say otherwise. Where a maker gives a range, we use the upper end for runtime and the single value elsewhere. "Biped" and "wheeled" groups follow each record's form-factor field. Records with no stated form factor are left out of those splits. Body DOF and implied power draw are our own arithmetic from published figures. Between 1 and 9 October 2026 we re-checked key manufacturer pages (Unitree, Booster, AgiBot, 1X, Boston Dynamics, Agility) and the ISO, IEC and NVIDIA reference pages linked above. We have not tested these robots hands-on. See our methodology and the humanoids category for the full dataset.

Spotted an error or a newer figure? Email corrections@rankedrobot.com and we'll check it and log the correction.