I wrote a piece two weeks ago asking whether robots need to be humanoid at all. If you haven’t read it, do so:

The shape of every robot that makes money was calculated. The answer is rarely humanoid.

A robot has to be general ONLY when these two exits are both closed:

As long as one is open, a function-specific robot wins.

Trillions of dollars are betting that the situations where both routes are blocked are growing.

The logic is quite simple: as long as humans can do it, the robots should be able to. And the world was built for humans. If you want a robot to buy your groceries, it has to walk into a store, pick things off shelves at different heights, wait in line, and carry bags home.

You’re not going to rebuild the supermarket for that. You’d build the robot like a person.

One thing is clear: no one is claiming humanoids will replace the robotic arms already in factories. They are betting on the future. Maybe the near future.

Let’s look at house cleaning, which is what most humanoid robot demo videos love to show you.

Exit one is closed here. You won’t rebuild your house like a training ground to buy a robot.

Actually, Figure AI should sell their robots with a renovation package for new houses. That way they wouldn’t have to worry about a general robot adapting to different homes. If the robot has data from being trained in 1,000 different kitchens, buyers could pick from 1,000 kitchen designs. Right?

If you have to adapt your home to a robot, you’ve just built a factory. A general-purpose robot is supposed to fit into the world as it already is without anyone changing anything.

So exit one is closed. What about exit two: shrinking the task?

House cleaning means wiping surfaces, washing dishes, picking up whatever the kids left on the floor. But if you shrink it down to just the floor, a vacuum robot handles that fine. No arms needed. It’s disc-shaped and sticks to the ground. It doesn’t have to open a drawer or fold laundry.

If you have one at home, you’ve probably noticed: it gives up the moment there’s a staircase (just like my living room).

About 32.7 million vacuum robots shipped in 2025. That number is exit two working. Someone carved out the easiest layer of housework and sold thirty million units of it.

The problem is what’s left after the floor. Washing dishes, laundry, the thing on the stairs. Nobody has found a way to shrink those into a job a disc can do. Exit two peeled off one layer and got stuck.

For everything still waiting above that layer, a general-purpose robot is the shape that makes sense.

A warehouse is a little different from a factory. A factory's space is designed around a certain engineering workflow. For example, the width of a conveyor belt would be dependent on the parts to be delivered. Humans were playing a role in that workflow, so there's almost no change required when robotic arms were introduced. They don't have to think about rebuilding in this case; they might not even need to redesign. So that's exit one. Closed. How about exit two? The factory pipeline was designed to break tasks into smaller pieces from the start, e.g., putting two parts together, which can be easily done by a robotic arm built for that single purpose.

Every single dimension of a warehouse is designed around humans. The height of a rack is set to be within reach of a standard human. The passageway is wide enough for humans to pass through when pushing a trolley. Routes through the building are planned based on humans' walking speed.

In a factory, we only have to replace a human with a robotic arm (maybe not 1-to-1, but you get the idea). But for a warehouse, they have to change the whole layout from "humans walk to the goods" to "the goods move on their own".

Amazon chose Exit One. Tens of thousands of wheeled robots crawl under the shelves, lift the entire rack, and move it to a fixed workstation for humans to do the picking. It turns the scenario around: from "humans walking towards the goods" to "goods walking towards humans". Basically, it rebuilds the entire space based on the robots' logic. The robot itself doesn't have to be too complicated, as long as it can move smoothly and lift heavy things.

There's another company, GXO (a huge logistics company). They skipped exit one. They use a humanoid robot called Digit (from Agility Robotics) to work in the same space humans were already using. In that case, they don't have to redesign the space. But the cost? The robots they use are, in general, more complicated and expensive and need a longer time to stabilize. These are also the general pain points for humanoid robots (more in the next section).

GXO chose the humanoid figure not because it was better, but because the initial investment in renovating the entire warehouse was too large.

Agility Robotics has signed a multi-year Raas (Robot-as-a-service) contract with GXO Logistics to deploy the Digit robots into their warehouses. Although the amount of the contract is not publicly disclosed, this is indeed great evidence of real cash transactions on humanoid robots (not just for demo purposes!).

These robots have moved a cumulative total of over 100,000 totes since 2024. I know, it’s not like the robots have replaced humans completely in the warehouse; they’re moving totes! Exit two is not blocked in this case since it’s just a narrow task. Mainstream warehouse automation is still based on wheeled robots. But I do see this as proof that bipedal robots can survive well in a space designed for humans. For more than 2 years now.

The next example might be better-known: Figure AI. Figure 02 has been piloted for more than 1,250 hours on the assembly line at BMW Spartanburg, contributing to X3 production. It was responsible for loading sheet metal parts (more than 90,000 after 11 months) into specific fixtures.

Not gonna lie, I was amazed at first. Then I had second thoughts: neither exit was closed. The robots were deployed in a factory (designed to be a pipeline), and their tasks only involved pick-and-place operations on a fixed fixture. Don’t get me wrong, it’s not as “simple” as it sounds, but I still feel this is a bit different from what I expected a humanoid robot to be able to do. The pilot of Figure 02 was probably not meant to prove that a humanoid robot is better than a robotic arm for that job, but just to prove that bipedal robots could survive and accomplish tasks for a long period. From this perspective, it’s a legit result. Still, I want to see how humanoid robots can perform more complex and unstructured tasks in a human environment. If you have the same thought as I do, the next pilot might be exciting for you.

Figure 03 was deployed to the same plant (BMW Spartanburg) last month (June 2026). This time, it’s been assigned a more complicated task: advanced sequencing. This is how it will work:

This is not a simple pick-and-place scenario, with a fixed movement trajectory. It requires loco-manipulation (simplified explanation: moving its whole body while using its hands at the same time) to adjust its footing and posture to grab the parts, while also pulling/pushing the trolley. I’ll also focus on the performance of Helix 02 — the pixels-to-actions VLA (vision-language-action) model — for real-time motion correction by detecting changes around the robot (another core challenge for humanoid robots in a fully open environment). It’ll be a massive breakthrough if the Figure 03 pilot is successful. I’ll definitely provide an update when there’s one.

The gap between humans and humanoid robots is getting smaller, especially on structured tasks. The examples we mentioned above are proof of that. For reference, Digit robots handling totes in GXO’s warehouses have achieved approximately 60% of human throughput.

You might have seen this video recently (I mean two months ago): Figure 03 robot vs human intern in a package sorting challenge. In this 10-hour challenge, the human beat the robot. But it was a close one. The human did 12,924 package sorts, and the robot did 12,732 (the difference is just 1.3%). Think about it. A human needs a longer rest after 10 hours of work, but a robot could keep going as long as the battery lasts. The robot will definitely evolve, and like what Figure AI CEO Brett Adcock said, this will be the last time a human will ever win. And I do agree with that.

But don’t worry, we won’t lose our jobs over this, right away. In the kinda chaotic, unstructured settings where most of us actually work, the real deployment of humanoid robots still has a long way to go. In these scenarios, humanoid robots are much slower than us, only at about 10–40% of human speed. UBTech, one of the major players in humanoid robotics, admitted that their Walker S2 robots were only 30–50% as productive as humans (and only in certain tasks such as stacking boxes and quality control).

Figure 03 can last about 5 hours on a full charge. That’s actually one of the better ones. What’s cool is that it can walk to the wireless charging station on its own, and another fully charged robot takes over. But 2–4 hours is the industry norm right now. To cover a full shift, they have to deploy more robots to fill the endurance gap.

More robots in the same space means more things can go wrong. And right now, there’s no standard for what “safe” even means for a walking robot. Stationary robotic arms have well-established safety standards (ISO 10218). Dynamically stable robots don’t, and it’s not easy to draft one. If a robotic arm malfunctions, the range of motion is known, and a safety fence around it would be enough. A walking robot can go almost anywhere. Where it falls is unpredictable too. So how would you define a “safe” distance around a walking robot? There’s no standard answer as of now.

You might have also seen a lot of news headlines featuring the number of shipments or orders for a humanoid robot company. Because it’s the most eye-catching number. It’s not fake, but we have to understand the difference between shipped out and truly deployed into a working environment. The flow usually goes like: order → shipment → pilot → deployment. We should pay closer attention to the second half of the process. In Unitree’s annual report for 2025, they shipped over 5,500 humanoid robots. You might have the same feeling as I did when I first saw the number: does that mean more than 5,500 households now have humanoid robots to do the chores? Then I looked into where these robots actually went. Apparently they were bought by universities, research labs, corporate R&D departments, or government projects. So the purchase was not for deployment into a working environment, but more towards experimentation. I don’t mean these shipments have no value. They do contribute majorly towards the robotics space. It’s brought down the barrier to entry for humanoid robots (Unitree G1’s starting price is $16k) by a significant margin. But we have to be aware that these are not quite what the news headlines make them sound like. Humanoid robots becoming mainstream? Not yet.

Btw, I wrote a piece about Unitree robots a few months ago, worth a read:

Imagine a factory that’s set to produce one product this year, and a different one next year. They’d have to redesign the fixtures, recalibrate the arm’s motion, and retest everything. The whole process normally takes weeks, even months, and costs a bomb.

But what if you have a general-purpose robot? Technically, you just have to update the instructions: what does the new part look like, where to grab it, where to place it. Not much else needs to change.

This gap gets even more obvious in industries with short product lifecycles. Consumer electronics might change every single year (or even quarter). Same goes for fast fashion and the electric vehicle industry (if you’ve noticed, the pace of iteration for EVs is accelerating). Imagine having to retool whenever a product line shifts. So the economic logic of general-purpose robots isn’t about being “cheaper today”. The more frequently you retool, the more cost-effective they become. Specialized robots can do one thing better than humans could. By 10x or maybe 100x. But if the market or task changes, they turn into a heap of scrap metal.

I’ll give you a software analogy: specialized robots are like hard-coded scripts. They do one specific thing extremely well. But general-purpose robots are the AI agents. They’re slower and more problematic at this point, but no one’s ever questioned the long-term direction.

Hardware is just part of the equation. The real cost of versatility lies in software and training. If you want a robot that can do “everything”, you need:

In future issues of Robot Economy, we’re going to get into each of them. But for now, just know that the decision between specialized and general-purpose comes down to task switching frequency. If you’re confident the task at a workstation will never change, a specialized robot is the better investment.

The cost of renovating spaces is rising too. Land and construction costs keep going up. Many buildings have zoning restrictions that prevent arbitrary modifications. This is especially true for warehouses in cities. Not everyone can revamp their warehouses the way Amazon does.

So I do have great faith that general-purpose robots (hopefully humanoid) will be the way forward. We’re getting closer, but nowhere near ready.

Try asking these two questions whenever you see a humanoid robot headline:

That’s it for the week. I’ll see you in the next one.Cheers.