Twenty Thousand Robots Shipped, Nine in Ten Idle: Embodied AI Enters Its ROI Era
Beneath the heat of WRC, there may be only one or two startups genuinely doing pre-training, real deployment demand may be only 10% of shipments, and the industry has moved on from demo theater into a hands-on phase where ROI has to be calculated.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
A demo that only runs on the hour is hiding its success rate
Ye Yangsheng taught a set of trade-show detection methods. First, check whether the demo runs on a fixed schedule — a timed performance may mean the team is not confident about its success rate and needs someone standing right there watching over it, or that the hardware cannot hold up under long periods of operation. Second, watch whether the two arms perform independent motions at the same time: with teleoperation the operator usually finishes one hand before moving to the other, so two hands on different trajectories is probably autonomous planning. Also look at whether the robot can recover by itself after a failure, and whether it still recognizes an object when you change the lighting or put a pure white object in front of it. He suggests going into the venue during the build-out days — many companies are in fact collecting additional data on site and retraining, and only Unitree, which takes a pure-simulation route, does not need that step.
— Ye YangshengThe real headline is not the new robots, it is procurement day
The biggest difference between this edition of WRC and previous ones is that it set up a procurement day: central state-owned enterprises and large corporates came in with actual purchasing projects in hand, and that is the real headline. What Ye Yangsheng exhibited was a delivery robot plus a robotic arm, redoing the hotel food-delivery scenario: the robot picks up the takeout order itself, presses the elevator button, and brings it to the guest's door. Why not the previous generation's approach of retrofitting the elevator? Because overseas that basically does not work — safety regulations may be involved, and retrofit costs are several times what they are in China; in China elevators are classified as special equipment, so filing and construction take time, and a no-retrofit solution can speed up delivery. His definition of a humanoid robot is a machine whose form and behavior are close to a human's and that can use facilities prepared for humans — but it does not necessarily have to have two arms.
— Ye YangshengStrip out the dancing shows and one in ten shipments being real would be good
The host read out a set of shipment figures: in the first half of 2026 global humanoid robot shipments passed 22,000 units, with Chinese manufacturers accounting for roughly 97%; AgiBot ranked first at 9,700 units, Unitree over 7,000, Galbot 1,100, and the top five together took 86% of the global total. Ye Yangsheng believes China holds a dominant position in the full machine, on the back of its manufacturing base — foreign models are strong, but on the body itself they simply cannot compete. But once you strip out the shipments that go into dancing and performance, he thinks getting real deployment demand to 10% would already be very, very good. The ROI does not work out for something like space-capsule retail: the environment is uniform and the motions are fixed, so it also does not generate the kind of data that improves generalization.
— Shi Jie, Ye YangshengEveryone claims a full-stack in-house world model, and the padding is heavy
The capital markets' heat is on world models, but very few exhibitors were actually showing world model demos. Of the three directions Fei-Fei Li described, generation and simulation-based evaluation can still be demonstrated; the third is planning-style world models in latent space, predicting how an object's latent vector changes over some future interval, which is very hard to visualize — the videos that do get released look no different from a VLA demo. Asked whether there is padding in every company claiming VLA plus world model plus full-stack in-house development, Ye Yangsheng said the padding is very heavy: the number of startups that have actually bought a ten-thousand-GPU cluster and keep buying data to do pre-training may be only one or two. Many companies are sitting on billions of yuan in the bank, waiting for foreign players to blaze the trail before they spend, so as to avoid doing the work that benefits someone else.
— Ye YangshengA robot shipped to a factory but not running is not revenue
Unitree's IPO was the biggest off-floor topic of this edition. Ye Yangsheng said a benchmark market capitalization is good for the industry — it proves embodied companies can get listed, and the higher the valuation the better for everyone. But the twenty-odd companies queuing up to list have to get through strict audits: a robot shipped to a factory but not in use cannot be recognized as revenue; an acceptance form signed but the machine not actually running does not count; sold to a distributor but never reaching the end user does not count; selling to a data company and then buying the data back amounts to circulating inside your own body, and that does not count either. Ye Yangsheng said his own company has been through this kind of audit — the auditors go to the factory in person to see whether the robots are running.
— Ye YangshengThis year's real increment is in the nervous system, not the brain
Ye Yangsheng splits the technology stack into three layers — brain, cerebellum, and nerves. At the brain layer, very few companies are genuinely doing pre-training and the models are highly homogeneous, basically following the foreign frontier; the new direction is using an agent to orchestrate multiple expert models to complete long-horizon tasks. At the cerebellum layer, everyone has made real progress on motion control, but motion speed is still an order of magnitude away from a human's. The nerves are where this year's real increment is: robots have more and more sensors and actuators, which calls for a bus protocol that is high-speed, stable, and supports hardware time synchronization — but there is still no standard definition of one, and the players actually doing this work are chip companies and low-level software companies.
— Ye YangshengData standards will not be defined into existence, they have to be fought out
The data pain points break into four layers. Data formats have no industry specification — every company defines its own set, so anyone taking someone else's data has to convert it first, which gets expensive at volume. Timing requires time synchronization at the tens-of-nanoseconds level, and if a sensor only supports software-based synchronization the precision is not enough, while different vendors' hardware synchronization protocols differ from one another. Labeling specifications have no consensus either: beyond labeling the cup, whether to label transparency or speed is all still being tried out. Above that, the collection criteria depend on the design of the body and how the brain is trained, so the whole chain cannot be normalized. Ye Yangsheng said the thinking behind their data subsidiary is to run the same underlying logic that once made their controller business number one in the world — adapt to as many other companies as possible, and unify things with a de facto standard.
— Ye YangshengThis year's progress came from post-training, not from world models
Asked what drove the industry's progress this year, Ye Yangsheng said it was not world models but post-training: reinforcement learning is itself a form of post-training, on top of higher data quality and stronger engineering capability brought by AI coding. But he acknowledged the industry as a whole is anxious, and that no product has yet appeared that can truly be deployed. On ROI, from what he observed on site he did not see the gap with industrial robots narrowing noticeably, though he agrees it will converge over time. He expects data and world models to produce surprises next year, because this year everyone has invested very heavily in high-quality low-cost data acquisition and pipeline infrastructure. Next year he will watch only two indicators: whether the motions are fast, and whether the robot runs continuously from the show's opening to its close rather than performing on a schedule.
— Ye YangshengIn their own words · checked verbatim
A lot of the other exhibitors, their demos may have to run on a schedule. Like, on the hour they'll do a demonstration. That makes me curious — as an insider, let me analyze it: why does it have to be on a schedule? One possibility is they're not too confident about the success rate, so they need someone standing next to it the whole time, watching over it.
很多其他展商，他们的demo可能是要定时的。比如说整点可以去演示一下。这我很好奇，作为行家来分析一下，就是为什么要定时，可能一个是对成功率不太自信，他得一直在旁边有人盯着守着
Ye Yangsheng3:18
Is this autonomous, or is there someone behind it operating, with the operating station hidden somewhere? Because when I go see a show, at every booth I still watch carefully, um, the more subtle motions of the integrated hand.
这是自主的，还是后面有人要操要操台藏在哪的。因为我去看展啊，就每个展位我还是会仔细观察，嗯他这个集成人的一些。比较细微的动作。
Shi Jie, Ye Yangsheng9:44
Well, if you set aside the dancing-and-performance kind, if you count those out, count those out, I think maybe having 10% would already be very, very good.
那你不说那种跳舞表演，算把那些把那些排掉排掉，我认为可能有有10%已经非常非常好了。
Ye Yangsheng20:19
The world models Fei-Fei Li talked about are actually three directions. The first one is — the first one is generation, the second one is simulation-based evaluation. And the third one is planning, I guess.
李飞菲说的世界模型其实三个方向，第一个是。第一个是生成，第二个就是仿真评测。第三个就是规划吧。
Ye Yangsheng26:55
I can reveal some inside-the-industry content, which is that right now the companies that may really be doing pre-training, there may be only 1 to 2 of them, 1 to 2, yes, 1 to 2. And then everyone else is maybe sitting on billions in the bank, all frozen there.
我可以透露一些行业内容，就是就是目前可能真正的在做预训练的公司，可能只有1到2家1到2家对，1到2家。然后大家可能拿着几十亿在账上都固化
Ye Yangsheng29:02
And it's the same for the machine side — actually, uh, there still isn't a fairly standard definition for this, uh, bus or protocol for robots, one that would let everyone plug in quickly and conveniently, plug into your cerebellum, or into your brain.
那对于机成也是的，其实呃还没有一个对于机器人这个呃总线或者说协议这样的一套比较标准的定义，可以去快速方便大家去接入这接入到你的小脑，或者说你的大脑中去。
Ye Yangsheng41:37
Actually the scenarios everyone defines, if you really talk them through, they're not all that complicated — they're still constrained scenarios, and for many of the scenarios we can see, the range of movement is fairly small. So VLA is entirely good enough.
其实大家定义的这个场景，你真的要去讲，其实也没有太复杂，还是有限的场景下，并且很多我们能看到这些场景，它的移动范围都比较小。所以VLA完全是够用的。
Ye Yangsheng57:54
Figures
| Global humanoid robot shipments, first half of 2026 | Passed 22,000 units | 18:10 |
| Chinese manufacturers' share of global shipments | About 97% | 18:10 |
| AgiBot shipments (first in the world) | 9,700 units | 18:10 |
| Unitree shipments | Over 7,000 units | 18:10 |
| Top five manufacturers' combined global share | 86% | 18:10 |
| Real deployment demand as a share of shipments | About 10% (Ye Yangsheng's estimate) | 20:19 |
| Number of startups genuinely doing pre-training | 1-2 companies | 29:02 |
| Number of embodied AI companies queuing for an IPO | Over 20 | 30:07 |
| Cash and interest at leading embodied companies (the ‘perpetual company’ formulation) | Billions of yuan on the balance sheet, a low few hundred million yuan a year in interest, enough to cover the salaries of several hundred R&D staff | 31:09 |
| Cost of a complete humanoid robot | Already below RMB 100,000 | 59:58 |
Glossary
- VLA / Vision-Language-Action model
- An end-to-end model that maps visual and language input directly into robot actions; the mainstream technology stack behind this edition's demos.
- World Model
- Lets a model simulate internally how the physical world evolves, used for generating scenes, simulation-based evaluation, or planning; comes in visualizable and latent-space forms.
- Teleoperation
- A human remotely controls the robot in real time, commonly used to collect data or to run remote demos; detectable from the continuity of the motions.
- EMG / Electromyography signals
- A wristband captures the bioelectric signals of muscle discharge, cross-validating hand pose against vision and touch; a new sensor trend this year.
- Data Pipeline
- The chain of processing from collection and labeling through format conversion to feeding the model; the industry still has no unified standard for it.
- Post-training
- The stage after pre-training where the model is further adjusted with reinforcement learning and high-quality data; the industry credits it for most of this year's progress.
How to listen
Investors trying to judge what Chinese humanoid robot companies are actually worth, and technical leads at manufacturers currently choosing a robot solution who want to get past the PR script.
The first two and a half minutes are the show intro and the guest's company background — skippable; everything after that is dense.