Humanoid robots at €20,000: China now controls 44% of the market

The domestic robot is already selling for €20,000 and asks you to remove the box

The first gesture of a humanoid fresh out of its packaging is not to hold a glass or recognize its owner: it is to ask for the cardboard to be removed. "Please, could you remove the box so I can start working?" asks the machine. The scene — a robot deploying its instruction manual before its skills — sums up better than any industry report the exact point domestic robotics has reached. The product has arrived. The packaging still weighs more than the promises.

The image is circulating these days: a domestic robot that goes from cinematic science fiction to a package that knocks on your door. The economic question is not whether the gadget is likeable, but at what price it enters homes and who is getting the business.

€20,000 per device and a Chinese manufacturer in charge

The price being discussed is around €20,000 per unit. For an average family, that is equivalent to a car — and a good chunk of a mortgage down payment — for a contraption that still needs constant supervision. With that price tag, the sales pitch does not hold up on domestic utility: it holds up for what comes next.

The industry figures are more eloquent than any demonstration. The world's number one manufacturer is no longer Unitree. AgiBot has increased its shipments by 562% and has taken 44% of the global humanoid market. Deliveries are soaring driven by industrial demand — factories, shops, logistics — and China sets the pace of commercialization. Translated: domestic robotics is, for now, a byproduct. The volume is in the grind, not in your living room.

It is no coincidence that a listed robotics ETF appears in the conversation as a hype thermometer. When a device that cannot yet empty a dishwasher moves index funds, the market has already decided that the question is not if it arrives, but when.

AI was not born with ChatGPT, and that changes the timeline

Among the buzz there is a confusion of dates that should be cleared up, because it determines how much leeway you think you have. The turning point of deep learning was not 2023, but 2012: AlexNet winning ImageNet with convolutional networks on GPUs. In 2013, word2vec turns words into vectors with semantics. In 2017 the Transformer arrives; in 2018, Bengio, LeCun and Hinton receive the Turing Award for this line; in 2020, scaling laws confirm that more data and more compute improve results. ChatGPT, at the end of 2022, only made tangible what had been cooking for a decade.

That distinction matters for the economics of automation. If the takeoff is placed in 2012 and not 2023, the industrial adoption curve is more advanced than the general feeling suggests, and the reaction margin for labor-intensive sectors is narrower than it appears.

One minute for a cup, two hours for the dishwasher

Skepticism is not technophobia. Domestic videos show a humanoid that takes a minute to place a cup in the dishwasher and would need two hours to empty it completely. Everything happens in controlled scenarios, with the environment prepared for the camera. The gap between a base model — a general brain that understands rooms and objects — and the execution of a specific task in a messy world remains abysmal.

In the background there is real division. One camp holds that the technology is already mature enough for mass distribution and that laughing at it is short-sighted. The opposite responds that, as long as a device does not walk the dog or do the laundry, it is plastic with servos and marketing. And it recalls an uncomfortable detail: companion robots have been sold since the 1980s without almost anyone having seen one working in a real home.

The self-driving car as a dress rehearsal

Where physical robotics is already being seriously tested is in driving. Tesla offers appointments to test its system in Spain — Alicante, Bilbao and Sabadell already active, with Madrid and Valencia coming —, and in Europe several countries are accumulating kilometers with supervised driving, including Nordic stretches with snow and ice. The target being handled is not the car capable of everything, but level 4: full autonomy on pre-mapped routes.

Not even that step is resolved, and some recall that the big leap — recognizing an agent's hand gestures, dodging a vehicle on a narrow road — relies on millions of kilometers of real data, not a laboratory trick. On the other side of the scale weighs the underlying distrust: nobody wants their life to depend on a system that, if a wire crosses, cannot be braked by hand.

The anomaly is this: the robot enters the home at the price of a car, without yet knowing how to empty a dishwasher, and the market does not doubt the timeline. Anyone with doubts should look at the 44% share of the manufacturer that three years ago almost no one could name. The packaging weighs. The business does, too.