How to Read a Humanoid Robot Spec Sheet: Turning DoF, Payload and Runtime into Task Numbers

Humanoid spec sheets are usually short tables, but what each value was measured against tends to hide in the footnotes. The same “43 DoF” or “3 kg payload” becomes useless for task design if the counting scope or the arm posture differs. This article uses the official specification pages for Unitree G1, Boston Dynamics Atlas and Agility Digit 5 to show what to check for each line item, and then walks through three short calculations that turn spec-sheet values into numbers for your own task.

Design philosophy and AI stacks for these platforms are compared in the humanoid platform comparison. Here the focus is on reading the sheet and doing the math.

What the official pages actually list

Start by putting what each manufacturer does and does not publish into a single table. Blank cells mean the page did not list the item as of the verification date.

Sources: Unitree G1, Boston Dynamics Atlas, Agility Digit 5. The Digit 5 page states that its specifications are preliminary estimates based on internal testing and that the product is in development. Figure 03 is left out because the official announcement gives relative values, such as 9% less mass than Figure 02 and 2 kW wireless charging in the feet, but no absolute height, DoF or payload.

One thing stands out immediately: none of the four manufacturers lists walking speed on its official spec page. If your task involves long walking distances, that number has to be requested separately.

Calculation 1: add degrees of freedom over the same scope

Degrees of freedom (DoF) count the independent joint axes a robot can move. The usual comparison problem is that some models include optional configurations and others do not. Taking the Unitree G1 specification as an example, the base model has 6×2 leg, 1 waist and 5×2 arm DoF for a total of 23. The EDU’s maximum of 43 adds the options on top:

So 18 of the 43 come from optional extra waist axes, the force-controlled three-finger hands and the wrist axes. To compare against another robot:

1. Split each model into legs, waist, arms, wrists, fingers and head.
2. Fill in each segment from the spec page and leave unpublished segments blank.
3. Add up only the segments your task needs. For a carrying task that never uses the fingers, hand DoF can drop out of the comparison.

The Atlas page gives 56 DoF without a per-segment breakdown, so aligning it with other robots means asking Boston Dynamics for the split.

Calculation 2: payload shrinks as the arm extends

A spec-sheet payload is usually a single value under one posture and one condition. The Unitree G1 page notes that arm payload varies greatly with arm extension, and Atlas publishes 50 kg instantaneous and 30 kg sustained separately. Sustained capacity is $30/50 = 0.6$ of the instantaneous figure, so any task that carries an object while walking or repeats a lift should be designed around the sustained value.

You can estimate the posture effect from the torque the shoulder has to hold. With object mass $m$ and horizontal distance $r$ from the shoulder to the object, the torque from the object alone is:

Holding a 3 kg object close to the body at $r = 0.2\,\text{m}$ needs about 5.9 N·m, while holding it at arm’s length at $r = 0.5\,\text{m}$ needs about 14.7 N·m, two and a half times as much. Those distances are illustrative assumptions, and in practice the arm’s own weight adds to the torque. For a task that reaches deep into a shelf, ask the manufacturer how much the robot can lift in that working posture rather than relying on “maximum payload.” The reducers and actuators that produce joint torque are compared in harmonic, cycloidal and planetary reducers.

Calculation 3: convert runtime into shift availability

Runtime alone does not tell you how many hours a day the robot can work; you also need the time spent charging or swapping batteries. With runtime $t_{run}$ and charge time $t_{charge}$, availability is:

Digit 5 publishes a 10:1 run-to-charge ratio, which gives $A = 10/11 \approx 0.91$. Multiplied by 24 hours that is about 21.8 hours, while the same page lists 20 working hours per day, which suggests the remaining time goes to travel, queuing or other overhead. To keep one work station covered around the clock you need $1/A \approx 1.1$ robots, which in practice means two robots rotating.

By contrast, Unitree G1 lists only about 2 hours of runtime and a quick-release battery, and Atlas lists 4 hours and self-swapping. Neither page gives the swap time, so availability cannot be computed without asking. Battery capacity has the same gap: the G1’s 9,000 mAh can only be converted to energy (Wh = Ah × V) with the pack voltage, and the page does not list it.

Footnotes worth copying into your notes

Dropping footnotes when you transcribe specs is how comparisons go wrong later. The main notes on each page as of the verification date were:

• Unitree G1: maximum torque refers to the largest joint motor; parameters may vary by scenario and configuration; some functions are still in development and testing (Unitree G1).
• Digit 5: the product is in development, specifications are preliminary estimates from internal testing, and some safety features are still in development (Agility Digit 5).
• Figure 03: camera frame-rate, latency and field-of-view improvements are all relative to the previous generation (Figure 03 announcement).

Questions to send before requesting a quote

Every blank in the spec sheet becomes a question. Collecting the items the three calculations need but the pages do not publish gives this list:

Clearing these blanks first narrows a pilot’s scope much more than judging from demo videos. How the walking approach affects speed and stability continues in humanoid bipedal locomotion.