Why I Am Not Bullish on Humanoid Robots

AI translation notice: This English version was translated from the original Chinese article with AI assistance. Some nuances may differ from the original.

TLDR: I do not deny the technical value of humanoid robots in control, hardware, and artificial intelligence, and I do not think the field will fail outright. What I am not bullish on is something else: “human-like” does not automatically mean “general-purpose,” and “general-purpose” does not automatically mean “useful.” Until the problems of reliability, cost, energy, data, and maintenance are solved, humanoid robots are more likely to become research platforms, industrial testbeds, performance props, and geek toys first — rather than spreading the way personal computers or smartphones did, much less becoming a “new quality workforce” that replaces blue-collar workers.

The Future Is Already Here?

Starting with “YangBOT” at the 2025 Spring Festival Gala, humanoid robots walked into the public eye once again. The nimble moves of Unitree's H1 robots made plenty of people feel that the future had already arrived, and the media of the moment collectively lost its mind on the spot, leaving everyone with the impression that the direction of future technology finally had a clear signpost again.

Since then, on the one hand I have been amazed by how agile and how much fun humanoid robots are; on the other, I have been doubtful: is it really true, as these outlets claim, that humanoid robots are the future? Can embodied AI really be treated as roughly synonymous with humanoid robots? And the recent smooth running of the humanoid robot games brought the topic back to the center of public discussion.

I think I should state my judgment and my position clearly right now, and put it on the record so I can look back later. If it succeeds, I want to see what I am overlooking today; if it fails, I want to see whether the reasons match my judgment.

Defining the Scope

First, let me set out the scope of the discussion.

I am not rejecting humanoid robots on technical grounds. It of course brings together some of the most advanced work in control, mechanics, materials, sensors, and artificial intelligence. What I am not bullish on is this: that humanoid robots can shoulder what some people call “the next generation of general-purpose productivity.” The usual chain of reasoning — “human-like” = “general-purpose” = “practical” — is especially full of logical holes, and also of the stories capital likes to tell.

Let me now walk through this step by step.

The Technical Hurdles

I want to start by trying to explain why domestic robots have progressed so fast in recent years.

The Evolutionary Path

The first robot to hold the public's attention online for any length of time was probably Boston Dynamics'. As early as 2004, when the internet was just taking off, the BigDog project was already underway. As YouTube grew, everyone marveled at what this mechanical dog could do. The control technology, whole-body dynamics modeling, trajectory optimization, and contact-force solving behind it — those names are a moat built over decades by the smartest engineers and expensive hardware together. Boston Dynamics' robots amazed people again and again, and the technology and algorithms behind them were no less astonishing.

That story did not change until 2020.

The RSL group at ETH Zurich published that famous ANYmal reinforcement-learning off-road paper in Science Robotics, showing a path of training policies in simulation and then transferring them to real robots to handle complex terrain.

Training continuously in an RL simulation environment produced policies that ran more stably on complex terrain than Boston Dynamics' hand-tuned MPC. No hand-crafted models at all, and the code was open-sourced too.

Later tools — rsl_rl, legged_gym, plus NVIDIA Isaac Gym and others — kept lowering the bar for training and reproduction, while manufacturers like Unitree drove down the hardware price of electric-drive robots. The wall MPC spent decades building was easily walked around with RL. Capabilities that once only a handful of top teams could touch began to be reproduced by more labs and companies.

This was in fact a remarkable technical breakthrough. RL flipped the table, using the most cutting-edge techniques to finish in a short time work that was previously beyond imagination: many motions that are hard to model fully by hand can first be learned through massive trial and error in simulation, then combined with conventional control, state estimation, and safety constraints. The once towering technical barrier was, in part, simply bypassed.

After that, domestic companies had a cheaper, faster-iterating route to robot motion control, and all kinds of robot dogs and humanoid robots gradually began to appear on the market.

The Challenges That Remain

Let us go from the easier problems to the harder ones.

First, energy. There are of course battery-swapping solutions already, and they are passable, but that also means a robot's range and working hours are still limited. For a robot that performs for ten minutes this is not fatal, but for a production tool that has to work eight hours straight while carrying a load, it goes straight to cost.

Second, structure. Many humanoid robots today use motor-driven joints, and those joints have to handle torque, heat dissipation, weight, impact, and service life all at once, each pulling against the others. I do not know much about this area and do not dare go into specific designs, but I do know that demonstrating a motion once and repeating it thousands of times a day are held to completely different engineering standards. Whether it can get up on its own after a fall, whether precision drifts after continuous operation, how long it takes to swap out a broken joint — these small problems add up, and eventually they all show up on the bill.

Finally, the brain. Reinforcement learning has advanced a robot's locomotion and balance a great deal, but that is closer to the “cerebellum”: how to place a foot, how to stay stable, how to execute an already defined motion. What really determines whether a robot can be widely used is whether it can understand its environment, plan tasks, handle surprises, and know when it cannot do something.

That is also where several mainstream research directions come in — world models (here is a good, simple introduction — though not a short one!), long-horizon task planning, and semantic understanding and exception handling in open environments. (In fact, there is a fair amount of skepticism about how well these directions will ultimately pan out.)

Many companies have started collecting data, hoping to score a breakthrough through the scaling law the way large language models did. But real-world data, unlike text, cannot simply be downloaded from the internet. A segment of robot data requires real hardware to execute it, needs the actions and outcomes annotated, and may come with wear, collisions, and human takeovers.

In other words, a language model can be trained “for free” on text from the internet, while every round of “learning” for a robot model wears down hardware and consumes electricity and human labor. That cost structure makes it hard to simply replicate the explosive trajectory of large language models.

More troublesome still is whether data collected from different embodiments, different camera positions, and different joint parameters can transfer to one another. This resembles the problems I know from medical imaging — and may be even messier. Medical imaging still has no clean answer the way language models do.

So the “brain” is not just a question of whether the model is smart enough; it is also a question of where the data comes from, who bears the cost of mistakes, and whether capabilities can be reproduced reliably.

Who Is Actually Buying the Robots — Market and Application Space

Some robot companies already have decent revenue, and Unitree's numbers show it turned a profit long ago, but that does not directly prove that humanoid robots have a large end market. We first need to look at who is buying these robots and what they are using them for.

Judging from news reports and public cases, humanoid robots genuinely and steadily deployed in production are still very few; most buyers are universities, research institutes, showrooms, and companies that need to build on top of them. I often see universities publicizing that their lab has a Unitree robot, and I often see robots at product launches, shopping malls, and sports meets, but I rarely see a factory publicly show how long one has run continuously, how much cost it has replaced, or how much stable output it has generated.

Morgan Stanley's data says that 65% go to display, reception, and research, with only 35% actually reaching a production line. The materials Unitree filed with the Shanghai Stock Exchange are even more blunt: in the first three quarters of 2025, 73.6% of humanoid robot revenue came from research and education. And of the ones that do reach a production line, how many are there to tell a story? That is not something we can verify for now.

The robot here is more a paper-production tool, an object of capital narrative, and a direction for hype.

I prefer to think of it as an expensive Lego model, or more aptly, as something like DJI's RoboMaster: an excellent platform for teaching, research, and development that can also give rise to competitions, performances, and enthusiast culture — but that is a different thing from becoming a general production tool.

Of course, pilot projects are a necessary stage before mass production. I am not saying “there is none now, so there will be none later.” I only think that when we discuss the market, we cannot count every robot sent into a lab, a showroom, or a pilot line as already-validated productivity in advance.

At this point someone will say: no market today does not mean the potential market is small. Every device in society today is designed around human ergonomics, so the only truly general-purpose robot in the future can be a humanoid one.

But that logic does not close the loop.

Human-Like ≠ General-Purpose ≠ Useful

There are two faulty inferences here: first, human-like ≠ general-purpose; second, general-purpose ≠ useful.

The human hand is certainly very dexterous, but it was not designed for the modern industrial environment. There are many explanations for how the hand's shape evolved — grasping, throwing, and tool use may all have played a part; see this paper on the relationship between throwing and the evolution of the human hand. But society today has developed far faster than biological evolution. We invent different tools to do different things precisely to make up for what the hand does poorly.

This is itself an example of “general-purpose ≠ useful.” You can just about turn a screw with your fingers, but a screwdriver tightens it more; a ratcheting screwdriver tightens it fast and tight; and on a production line people go straight to a torque-controlled power tool. The human hand is general-purpose, but what actually raises productivity tends to be the tools that look not general-purpose at all.

The evolution of these tools is all accumulated industrial knowledge — why can't it carry over to robots? A wheeled base moves more efficiently, a robotic arm positions more stably, a fixture can fit a part directly, and a conveyor belt does not even need to “understand” what it is transporting. When a scenario is fixed enough, companies usually prefer to modify the environment and the process rather than buy a complex machine with two legs, ten fingers, and whole-body balance.

Even when a system genuinely needs to handle many tasks, it does not have to imitate a human completely. Modular end effectors, quickly swappable fixtures, a mobile base plus a robotic arm, and then letting humans — who are already smart enough — handle the few exceptions: that may be a cheaper and more reliable transition.

The greatest paradox of the humanoid robot is that it tries to use the most general-purpose body to solve the least general-purpose tasks.

The real advantage of a humanoid form is that it can walk straight into environments designed for people without first changing the doors, the stairs, and the tools. But that advantage carries a hidden precondition: the cost of remodeling the environment must be higher than the cost of maintaining the complexity of a humanoid form. In homes, disaster zones, or old buildings, that precondition may hold; in factories and warehouses that can be replanned, it may not.

So Where Is the Application Space?

In my view, the applications where a humanoid form is irreplaceable may not be the assembly line that gets advertised most, but the scenarios where being human-like is itself valuable. If we really can build a robot human-like enough in the future — one that feels, looks, and talks like a person — then of course it can become an NPC in a virtual theme park or a companion robot. At that point the humanoid form is not there to barely accommodate a wrench; it is the product experience itself.

Ordered from easy to hard, I think the progression looks more like this:

• The boxing robots in Real Steel — sports events
• The home-care robot in A.I. — emotional companion robots
• The virtual-park NPCs in Westworld — entertainment facilities
• The general-purpose robots doing all kinds of basic work in Detroit: Become Human — replacements for blue-collar workers

And we are still a long way from the state of things in science fiction. So the current market and level of technology only prove that “we can already build increasingly agile humanoid robots”; they do not directly prove that “the more a robot looks like a human, the closer it is to a correct business path.”

Why Is Everyone Bullish on Humanoid Robots?

Next I want to talk about a few points that come up in today's discussion of humanoid robots.

How Does the Emotion Arise?

First, movies and other science fiction have made people unconsciously hold a lot of rose-colored fantasies about humanoid robots. Compared with a robotic arm, a humanoid robot is easier to film, to spread, and to understand. It only needs to take a couple of steps, wave, and show an expression on its head, and the audience can empathize with it immediately.

As with the application spaces listed above, it is not that humanoid robots have no future; it is that we often mistake the imagining that science fiction has already done on their behalf for capabilities that real products have already achieved.

The economy is sluggish right now, and everyone is eager to find a breakthrough that can pull investment along. Our materialist education keeps stressing the importance of productivity and technological progress, so people are naturally willing to support advanced productivity they can see and touch, hoping that one technological leap will improve their current situation. The humanoid robot happens to satisfy exactly this expectation: it offers the imaginative promise of artificial intelligence, the visual impact of a physical machine, and a link to manufacturing and supply chains.

Capital, local governments, the media, and startups all have incentives to keep telling this story. Capital needs a big enough potential market, localities need a new industrial calling card, the media needs eye-catching footage, and startups must first convince others that the future is big enough before they can get the money to keep developing today's immature products. These incentives do not necessarily mean a scam, but they naturally tilt the discussion toward its most optimistic version.

Rebutting the Most Common Argument

The most famous argument, and the easiest to rebut: no technology was smooth sailing at the start.

Some cite the automobile as an example, others the airplane, the mobile phone, and the mobile internet. All of them met with skepticism at first and later genuinely transformed human life.

But this is not the first time humanoid robots have appeared in the public eye.

In the 1980s, Heathkit brought out the HERO 1, and Tomy released “home robots” like the Omnibot. They could move and make sounds, and some could pick things up with a simple arm. They carried the same dream of home automation at the time, yet later never became part of the household appliance lineup.

People remember only the products that succeeded. Plenty of technologies were once favored and later failed to achieve mass adoption; people simply skip over them when arguing, landing squarely in survivorship bias.

Examples include hydrogen airships, the Concorde supersonic airliner, 3D televisions, Google Glass, and the metaverse of a few years ago. Not every one of them was a total technical failure — some even have niche uses to this day — but none delivered on the promise, made at the peak of their hype, that they were about to change everyone's life.

These examples each match some of the current features of humanoid robots: the demos are stunning, the technology is genuinely advanced, and real demand can be found, but cost, supporting infrastructure, reliability, or social acceptance kept them from expanding as originally imagined.

So “people once misjudged the automobile” does not prove that questioning humanoid robots today is wrong. To complete that argument, at least a few more specific questions need answers: What problem does it solve that nothing else can? How much cheaper is it than a human or specialized equipment? How many tasks can it still complete outside a carefully staged demo environment? And who bears the cost of failures, maintenance, and safety?

How It Might Develop

But I am not pessimistic enough to think humanoid robots will amount to nothing. I suspect it will end up closer to blockchain, which was equally hot a few years ago.

Blockchain is a classic case of “over-promised, under-delivered, but leaving behind some real capability,” and it broadly matches the development path I have in mind for robots, so I will use it as the example.

At the peak of the ICO boom in 2017, many projects could raise large sums with nothing but a white paper, as if every industry should be on-chain; in 2018–2019, after the bubble burst and a large number of projects went to zero, blockchain was in turn denounced wholesale as a scam. Today the speculators have left, but the underlying technology really is still developing in a few areas, such as stablecoins and cross-border settlement.

In other words, a bubble and a technology are not an either-or choice. A field can contain real technological progress, inflated commercial promises, and a flood of trend-chasing projects all at once. When the bubble recedes, what remains may not be the product from the original story; it may be training frameworks, low-cost joints, sensors, control algorithms, supply chains, and a cohort of engineers who genuinely understand robot systems. A body of genuinely useful technology accumulates, and then gets taken apart and installed elsewhere.

My guess is that humanoid robots will go through something similar:

1. During the hype, every company has to make a humanoid robot, and every product launch has to have one walk a few steps.
2. Once capital starts demanding revenue and repeat purchases, a batch of companies that cannot find a stable scenario exits, and the industry enters a disillusionment phase.
3. The companies that remain stop emphasizing “general-purpose” and move into a few specific scenarios, honestly solving for one motion, one process step, or one hazardous environment.
4. Parts of humanoid technology get spun out into robotic arms, exoskeletons, wheeled robots, and other specialized equipment.
5. Parts of the humanoid form and its sensing technology can assist people with disabilities, much like the relationship between rockets and MRI.

If that is how it ends up, I will not consider humanoid robots a failure. What fails is the overinflated narrative that “it will soon be a general-purpose laborer in every home and every factory,” not the technology accumulated over these years.

If I Am Not Bullish on Humanoid Robots, What Am I Bullish On?

What I am genuinely bullish on is the AI models, perception, and reasoning behind humanoid robots.

A model that can understand the real world, recognize the relationships between objects, grasp human intent, and predict what an action will lead to — that is what could become the truly general part of embodied AI. It does not have to be installed in a humanoid body; it can control robotic arms, drones, cars, and wheeled bases, or even exist purely as an assistive system for people.

Compared with replicating the human body, I am more bullish on these directions:

• Specialized robots designed for clearly defined tasks, such as warehouse, cleaning, inspection, agricultural, and surgical robots
• Modular systems made of a mobile base, a robotic arm, and interchangeable end effectors
• Semi-autonomous systems where humans take over remotely and the model gradually learns, rather than pursuing full autonomy from the start
• Vision-language-action models that can transfer across embodiments, world models, and robot data infrastructure

None of these is as spectacular as a complete humanoid robot doing flips on stage, but they are more likely to enter the real world first. Because what users are buying is not a “feeling of intelligence,” but having a task completed more cheaply, more reliably, and more safely.

What Do I Think Will Happen to Humanoid Robots?

I think humanoid robots will not spread especially widely, but they will become a new kind of toy and a social culture. As prices fall, more schools, labs, and hobbyists will develop their own motions, appearances, and AI models, and bring their robots to fighting, soccer, racing, and obstacle-course competitions. Like RoboMaster, drones, and the early personal computer, they may first form a creative enthusiast ecosystem.

At the same time, they will also enter a small number of scenarios where the environment cannot be remodeled but a human form is needed. In those places, even an expensive, slow robot already has value if it keeps humans out of hazardous environments. Some robots will exist in the form of rentals, performances, marketing, and companion products. They may be a decent business — just not the business of “replacing all labor.”

Of course, I will also leave my future self a few criteria by which I can be proven wrong. If future humanoid robots can:

1. Work for hours on end in a home or factory without pre-scripting, and rarely need human takeover;
2. Learn new tasks after moving to a new site without collecting a large amount of new data;
3. Still be more cost-effective than a human or specialized equipment once purchase, energy, maintenance, downtime, and safety costs are all included;
4. Go from a demo at a launch event to a standard product that can be delivered at volume, maintained reliably, and has clear accountability;

then I will admit that my current judgment underestimated the transferability of general models and the speed at which mass production lowers the cost of complex hardware. By then, being “human-like” may truly be more than the easiest form factor to market; it may be the most economical way to fit into the human world.

But until those things happen, I am still unwilling to treat a smooth dance routine, a robot soccer match, or a slickly edited factory video as evidence that the next generation of productivity has arrived.

Humanoid robots are certainly part of the future, but the future never takes only the shape most familiar from science fiction. The technologies that truly change life are sometimes precisely the machines that no longer look human, no longer belong on stage, and have already been working quietly for years.

---

Update, September 19, 2026:

This article received two comments on Zhihu, both arguing that it commits the straw man fallacy — that I set up an easy target that cannot hit back, and then attacked it. Looking at it now, there is indeed some of that.

zhuanlan.zhihu.comZhihu link↗Jump over to Zhihu to see what everyone is saying in the comments

You can click this card to jump over and take a look.