The obvious worry about a humanoid robot is that it runs out of power. It does, in about two hours. That turns out not to be the constraint, because charging the pack takes less time than draining it, so a second battery buys the whole shift back. What a second battery does not buy is the person who walks over three times a shift to change it. Two manufacturers publish enough of the battery numbers to check. Neither publishes that one.

The numbers they do publish

Unitree’s parameter table for the G1 gives “Battery Life: About 2h” and, in its accessories block, names the part “Smart Battery (Quick Release)” at 9000mAh. The charger is its own line, 54V 5A, and a third row puts the power supply at a 13 string lithium battery. I read that 5 A as the charger’s output into the pack rather than its draw from the wall, because that is the reading under which 54 V and a thirteen-cell series string describe the same battery. The arithmetic below rests on it.

The table gives no charging time. It gives you enough to work one out.

9,000 mAh ÷ 5 A = 1.80 hours.

That applies the charger’s rated current to the pack’s rated capacity and ignores the taper at the top of a lithium charge, so the true figure is longer than 1.80 hours. Longer helps my argument, so read 1.80 as a floor rather than an estimate. Nothing below depends on the exact value. It depends on one inequality: charging is faster than running.

One input below is mine, not theirs. Neither company publishes a shift length. I use eight hours, and 250 of them a year, which is one shift a day, five days a week. A site running two or three shifts gets proportionally more of everything that follows, so this is the low end of the answer rather than the middle.

One pack is the wrong picture

Run a shift on a single battery and you get the number you would expect. Two hours of work, 1.8 hours standing still, repeat. That is 4.4 useful hours out of eight, or 55% availability, meaning the fraction of the hours you wanted the machine working that it could actually work.

It is also the wrong picture, and the vendor hints at why in the name of the part. A battery called Quick Release is one designed to come out by hand and go back in fast, which only matters if there is a second one waiting.

Add a second pack and the arithmetic inverts. The robot runs two hours and hands over to a charged battery. The one it just emptied is ready 1.8 hours later, twelve minutes before it is wanted again. Eight hours of work in eight hours, no waiting. A third pack adds nothing.

So energy is not the ceiling. I assumed it was when I started, and the arithmetic says otherwise. The ceiling moved somewhere less convenient: buying those eight hours took three battery changes.

Three a shift, 750 a year

Twelve minutes of headroom is not much of a margin. The 1.80 hour figure ignores the taper, and if real charging crosses two hours then two packs stop being sufficient and a third starts earning its place. That possibility cuts in my favour, so treat two packs and three swaps as the floor rather than the expected case.

Three swaps a shift over 250 shifts is 750 battery changes per robot per year. Two back-to-back shifts make it 1,750 rather than 1,500, because the battery clock does not pause for the shift change and a fourth swap lands inside the handover.

I cannot tell you what one costs, because neither company publishes how long it takes. So invert the question. At 750 swaps a year, every single minute a swap takes is 12.5 hours of labour per robot per year. Put your own number on the minute. The multiplier is fixed by the battery.

It also scales the wrong way. Interventions per robot per shift was the metric I used for teleoperation, and swaps behave the same: a line item that grows linearly with the fleet, inside a business whose whole pitch is that labour does not.

The thing neither page says

Agility Robotics describes Digit as having “a 35 pound carrying capacity, 4 hour battery life, and the capability to work continuous shifts.”

Four hours is not a shift. For the third clause to hold, something has to happen at the four-hour mark, and the page does not say what. I searched both pages for dock, recharge, autonomous charging and charging station. As they read on 6 September 2026, none of those words appears on either. The detailed spec sheet sits behind a contact form, so I am not quoting it and not counting it.

That is not a claim that the answer is bad. It is an observation that the answer is missing from the page where duration and continuity are both asserted. Unitree leaves the same hole from the other side: it publishes a duration, names the battery for how fast it comes out, and never says whose hands take it.

What it costs

The capital line per robot is not one battery. It is two packs and a charger, on a machine Unitree lists from US $13.5K, before tax and shipping. The battery and the charger sit in the accessories block of that table and carry no price of their own, so I cannot tell you what the second pack adds. It is a one-time number either way, and against $13.5K a small one.

The operating line is the one to watch, because it is labour and it recurs 750 times a year per robot. If a payback model in front of you scales output with hours worked, check whether anything in it scales with swaps. If a model gets eight hours of work out of a two-hour battery and carries no labour line for the changeover, it is assuming a person who is not in the budget.

Where I could be wrong

The load behind “About 2h” is not published. Walking continuously and standing idle are not the same two hours, and every figure above inherits that.

Unitree disclaims its own table. Footnote 5 reads “The above parameters may vary in different scenarios and configurations, please subject to actual situation.” That covers the two hours and the 9000mAh alike, and it is the manufacturer telling you these are not guarantees.

The charge time is derived, not published. Unitree gives a capacity and a charger current and no charging time. If the true figure is well under 1.80 hours the margin is comfortable rather than twelve minutes. If it is over two hours, you need a third pack.

The G1 is a research platform, not a warehouse robot. At $13.5K with an 8 month warranty, my read is that it is priced and covered for a lab bench rather than a production line. I used it because it publishes the numbers. Digit, the machine actually sold for shift work, publishes fewer.

Autonomous docking would delete the labour line. If self-charging or self-swapping becomes standard and gets documented, most of this argument goes with it. That is the disclosure to watch for, and it is a cheap one to make if it is true.

Not every deployment needs a pack. Stationary work on a tether has no battery and no swap.

The energy problem is solved, and it was solved by adding a person. That person is not on the parameter table, and the parameter table is what the payback model gets built from.