48 Observations from the Robot Conference: Nobody Is Making Money, Yet 18 Companies Made 36 Investments
48 observations from the 2026 World Robot Conference: robots turning into assembly kits, real demos finally showing up, modifier inflation. Nobody in the industry is making money, yet 18 companies made 36 investments, and the cost of a robot body has fallen tenfold in three years.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Robot dogs ship as home assembly kits; humanoids are still swapping skins
Two companies this year launched a robot dog and a humanoid that the buyer assembles. The dog arrives as a single suitcase and the user puts it together like Lego. The humanoid starts as a thin skeleton, and then you swap on the face and the body. By comparison, the genuinely mature product is the robot dog; humanoids are still at the stage of being assembled in a factory and having their head and face changed on site. A centaur form also appeared, combining a humanoid upper body with the stability of a quadruped lower half, aimed at firefighting and disaster relief. The gap shows that product forms are maturing at very different rates: robot dogs can already be assembled at home, while humanoids are still ‘swapping skins’.
The laundry-folding demo the crowd is sick of is the real hard problem
Folding clothes was the most common demo this year. Staff ran it over and over, and visitors were plainly tired of it. Consumers think young people with all-in-one wardrobes never fold clothes at all, while enterprise buyers and engineers see it as ‘the best thing we can give you’ — clothing is a deformable object, so manipulating it is genuinely hard. On the floor, Galaxea (星海图) folded a T-shirt using two grippers; when one gripper closed on nothing, it kept trying until it got a grip. It is only a T-shirt — it cannot fold a down jacket or a duvet yet — but this is exactly the problem the industry wants to crack within one to two years. It shows robots are starting to have a little decision-making ability.
The humanoid's edge is not capability but not being purpose-built
Figure ran a 7x24 livestream in which a robot flipped parcels on a conveyor so that the shipping labels faced up, and it even hired a human to race against it — and won. Industry experts called the stunt out, arguing that large logistics centres have long handled this with suction-cup equipment and have no need for a humanoid. The opposing view is that the humanoid's advantage is precisely that it is ‘not purpose-built’: the machine costs a bit over a hundred thousand yuan, against several million for dedicated equipment, and it can be updated over the air and put on other jobs later. For a small or mid-sized logistics centre, buying one to test the water is a reasonable piece of decision logic. The dispute reflects how far apart the industry is on what counts as a ‘real problem’.
The stutters are what prove these demos are not fake
There were no fake demos this year. Several companies were willing to let visitors interfere with the robot, and the machines showed real inference lag — stopping to think partway from point A to point B. Why are real demos possible now? Because the two technical routes, VLA and world models, are starting to mature. You used to be able to spot a fake robot by things like whether its left and right foot lifted at the same moment; that kind of mechanical movement is now rarely seen. Demos have become more honest, and the industry is shifting from selling concepts to showing its actual state.
Anyone who needs a modifier to be number one isn't first tier
We photographed the booths and counted roughly 10 companies shouting ‘number one’: fastest growth, most units shipped, most quadrupeds shipped, top service robot, top quadruped service robot. Keep piling on modifiers and you can always find a ‘first’ of your own. Behind that mentality is anxiety, because a genuinely confident company does not stress any of it. Unitree, for example, had sold 18,000 units by the end of July, three times its full-year 2025 shipments, and needs no modifier. Companies that stack on modifiers are mostly not the first tier.
Most still aren't profitable, yet the money is already spilling into suppliers
Between January and 19 August 2026, 18 robotics companies made roughly 36 investments, internal and external. Most of them are not yet making money, and what they buy into is mainly upstream and downstream: foundation models, chips, core components, data fusion. In essence they are wiring a supply chain together. Direct competitors are also investing in the same company as each other. This is history repeating itself — in consumer electronics, many ODM plants stopped earning their revenue from contract manufacturing long ago and earn it from investing instead. Compared with the three previous humanoid waves, the fact that capital has spilled over far enough to invest outward is what makes this fourth wave distinctive, and Chinese robotics is starting to form the concept of an ecosystem.
Costs have fallen far enough to rent, but standards remain fragmented
Zhao Tongyang (赵彤阳), founder of EngineAI (众擎), says the body of a general-purpose humanoid has come down from 1 million yuan to 100,000 over the past three years, with room to fall further. In the supply chain, precision ball screws have dropped from 10,000-20,000 yuan to 1,000-2,000 yuan, and reducer prices will halve again. Falling costs are producing new business models: a company called Weilai Buyuan (未来不远) rents robots out at 3,000 yuan a month to take home. At the same time the industry still has no unified standard — different vendors' models and hardware do not interoperate — so the road to standardisation remains long.
In their own words · checked verbatim
If you keep stacking modifiers in front of it, you can always find some way to be number one — so they keep using this method to inflate their own title, to inflate their own label.
你在前面不断地加定语 你总会能找到一个第一 所以他不断地用这种方式 去膨胀自己的title 去膨胀自己的那个标签
This year nobody was running fake demos any more. ‘Fake’ here is in scare quotes. I ran into at least several companies on the floor, and the demos they gave me were all very real.
今年大家都没有做假demo了 所谓的假是带双一号的 我至少现场碰到了几家公司 给我的demo都是很真实的
He said: which scenarios can robots handle today, and what can you measure that by? Take those big padded work gloves from the Northeast — he said whatever jobs you can do wearing those big Northeastern gloves, those are the scenarios robots can gradually solve right now.
他说现在机器人 可以做的场景有哪些 你能用什么去衡量呢 比如说东北的 那个大手套 他说如果哪些活 是这个东北大手套 可以做的 那就是现在机器人 可以慢慢解决的场景
What I mean is a number one with no modifier in front of it, but so far I haven't seen one — or maybe you should call it a one-dimensional modifier.
我觉得就是你不加定语的第一 但是目前没看到 或者应该叫做一维定语
That money has nothing whatsoever to do with you, but it feels like: I witnessed history, and yet I didn't make a killing out of this wave. You haven't actually lost anything, but you always feel as if you've lost a lot of money.
这个钱跟你没有半毛钱关系 但是就感觉我见证了历史 但是我却没有从这一波里边赚到一把 你其实没有任何损失 但是你总感觉我好像损失了很多钱
This is a plain, unadorned episode. We just talk about our own experience — there really isn't much reaching for grand significance, or forcing analysis on top of the evidence.
我们也是一期很朴素的一个节目 我们就讲我们底的体验 其实没有特别多的 上价值和所谓的 根据强行进行分析
Figures
| Precision ball screw price | 10,000-20,000 yuan → 1,000-2,000 yuan | 43:23 |
| Robotics company outbound investment, January-August 2026 | 18 companies, 36 deals | 33:18 |
| Time to pick up a bottle of cola (Ant-backed) | 107 seconds | 21:09 |
| Robot rental price | 3,000 yuan a month | 43:23 |
Glossary
- VLA (vision-language-action model)
- A model that takes visual input and natural-language instructions and generates physical robot actions directly.
- ODM (original design manufacturer)
- A firm that designs and builds products for brand owners; robotics now has ODM plants whose output can be rebadged.
- UWB (ultra-wideband)
- High-precision indoor wireless positioning; in the episode it serves as a hard beacon placed on a duvet.
- Zero-shot
- A model completing a task it has never seen before; Generalist 1.5 was called a GPT-3 moment by some independent media.
How to listen
Embodied-AI founders, investors, engineers following where robots actually get deployed, and anyone curious about the ‘knockoff-phone era’ framing.
You can skip the opening walk-through of the venue and the book plug at the end.