# WAIC 人形机器人同质化严重，缺"大脑"难落地

- 来源：X.PIN (@thexpin)
- 发布时间：2026-07-23 00:40
- AIHOT 分数：65
- AIHOT 链接：https://aihot.virxact.com/items/cmrwbnaaa0066roj05jmq5tv9
- 原文链接：https://x.com/thexpin/status/2079969850865414432

## AI 摘要

中国四足机器人占全球销量近 70%，已生产超 400 款人形机器人，但 WAIC 展会上多数机器人演示重复、动作缓慢且易失败。问题在于机器人硬件已成熟，但缺乏可靠的感知、决策与执行能力，企业需先获取订单和客户才能推动复杂工作落地。

## 正文

http://x.com/i/article/2079858359289995264

# China Built the Robots. Now It Has to Put Them to Work.

At WAIC， China flexed its manufacturing muscle， but the show was dominated by repetitive demos with little practical use. What these machines need now is real work.

On a rain-soaked day in Shanghai， I spent a one-day pass at the World Artificial Intelligence Conference trying to see as many robotics booths as possible.

The embodied-AI section was the loudest， hottest， and most crowded part of the show.

China has the numbers to justify the hype： official data says Chinese-developed quadruped robots account for nearly 70% of global sales， while the country has produced more than 400 humanoid robot models.

But the section was also deeply repetitive. Most robots wave， grip， dance， falter， and then， occasionally fail to complete a move. What China's humanoid industry has built ahead of schedule is the body. What it still lacks， in most cases， is the brain.

That creates an uncomfortable problem. These companies need sales channels， government relationships， and enterprise orders before their machines can reliably perform complex work. In other words， China's robots may require employment more urgently than humans do.

For customers and investors， the question is no longer whether China can build robots. It can. The harder question is whether robots can be put to work.

## Why Do Chinese Robots Look So Similar？

The answer is simple. Most companies are not building robots directly for consumer needs. They are building them to showcase their ability to produce fully fledged robotic hardware， helping them secure enterprise sales orders and pave the way for an IPO.

At WAIC， many humanoid booths showed the same scene： an upper-body robot trying to pick up objects from a tabletop.

The structure is familiar： robotic arms， a torso， a head， grippers or dexterous hands， and a model.

Robotic arms are usually six-axis or seven-axis systems. Some companies buy from suppliers such as RealMan Robotics or Elephant Robotics； more ambitious humanoid companies develop their own.

The torso contains computing hardware such as NVIDIA Jetson AGX Orin and Rockchip RK3588.

Lidar and camera suppliers include RoboSense， Hesai， Livox， Orbbec， Mech-Mind Robotics， and Intel RealSense.

Dexterous hands can come from Linkerbot's Linker Hand， INSPIRE Robots， or PaXini， though many humanoid companies still prefer self-developed parts.

Almost everyone claims to have built an in-house model. Most follow a vision-language-action， or VLA， route. The differences lie mainly in motion decisions and training pipelines.

Robot integration is complex. But the direction is clear： robots are starting to look like smartphones. Every component has suppliers. Integration companies can assemble the pieces. In the near future， robots will likely have their own ODM solutions.

Yet， conformity isn't the only issue.

Apart from looking quite alike， the robots rarely perform tasks smoothly and reliably. The grasping demos were slow， hesitant， and prone to small failures. In front of me， a Galaxea robot failed to lift a piece of bread as requested， even after two tries. The process took roughly 30 seconds. I would not ask it to make me breakfast.

After testing several demos myself， I identified several fractions of the problem： latency， visual error， and the absence of depth and touch.

Many robots， if you could picture it， don't really have wrists. A dexterous hand or gripper is mounted directly to the robotic arm， ignoring the fact that human wrist is crucial to the grasping movement which presented in motion-capture data.

A lot of robots also lack a reliable sense of depth. For example， camera failures and small errors can cause the robot to stop a few millimeters short of the object.

The perception-action loop creates another problem. The robot needs to understand the instruction， identify the object， select a subtask， generate a motion trajectory， execute it， and repeat. With imitation-learning strategies similar to Diffusion Policy， a task becomes： capture the image， run inference， generate an action， execute， repeat. That loop makes physical work feel like a loading bar. The first 60% can be smooth. The last 10% pauses， recalculates， and then stumbles through the remaining 30%.

The sheer volume of these humanoids is unnerving to witness. If they actually needed jobs to go with their "resumes，" I'm afraid the robot unemployment rate would easily dwarf ours.

There were also plenty of so-called realistic humanoid robots covered in silicone skin which could make facial expressions， but they're not designed for actual， useful work.

## Can Robot Companies Fix This？

Robotics companies have two problems to solve： how to make robots move more smoothly， and who to sell them to.

The first problem is painfully difficult. It requires a new training paradigm， and Chinese robotics has not yet had its DeepSeek moment.

From what I saw at these booths， VLA still does not match the demands of real-world work.

Language models are good at organizing intent， but they cannot fully describe a messy physical environment. Vision models can recognize objects， but they often lack reliable depth perception and still make basic mistakes. In many VLA systems， vision and language remain separate layers. They help robots understand a task， but they do not yet give robots the kind of instant， embodied decision-making that physical work requires.

Robotics still lacks its own equivalent of a general-purpose LLM： a model trained on physical operation， failure recovery， object handling， and real environmental constraints. That is why lessons from autonomous driving only go so far. A self-driving car mainly needs to read roads， lanes， vehicles， and pedestrians. A humanoid robot has to deal with soft packaging， awkward angles， slippery surfaces， unclear instructions， and objects that may fall， bend， or move in ways the model did not expect.

Anyone who has watched an assembly line closely， or helped install factory equipment， understands the gap.

Production work is never fully standardized. There are always irregular tasks， unexpected obstacles， and minor accidents that conventional robots cannot handle. Humanoids were supposed to solve these problems from the beginning， but today's robotics foundation models are not ready for that reality.

For now， VLA can make a robot look intelligent on stage. It cannot yet make one reliable on the factory floor.

That is why many robotics companies avoid the model question altogether. Solving these problems may require overturning the existing training paradigm， and most companies do not yet have the technical resources or practical experience to do that.

So they have chosen another direction： data.

At WAIC， data-collection gloves were everywhere. These devices usually combine tactile sensors with a head-mounted camera that records movement from a first-person perspective. Collecting full-body data requires a more elaborate exoskeleton-style system. A pair of gloves typically costs less than $1，500， while a full setup costs between $7，400 and $11，800.

The logic resembles the early days of LLM training： collect the raw material first and work out what to do with it later. For robots， the most valuable raw material is video of human movement that can be converted into training data for physical actions.

But most of that data is not particularly useful. One GenRobot employee told me that under strict cleaning standards， as little as 10% may survive.

I remain skeptical that data collection alone will solve the problem. Data may be infinite， but training-grade data is not.

Companies such as AgiBot are already selling robot-training data in China. But unless better models can translate that data into smooth， complex physical movements， collecting more of it will not necessarily make robots better at work.

The problem is that these companies cannot afford to wait. They have already built the robots. Now they desperately need customers.

At WAIC， larger companies were searching for practical use cases. Robbyant， Ant Group's embodied-AI unit， and Galaxea were both working with Guoda Drugstores on the idea of a 24-hour robotic pharmacist. Users would place orders through Alibaba's apps， while robots retrieve and dispense the medication inside the store.

It is a reasonable use case because medicine boxes are relatively standardized. Galaxea has also experimented with robotic receptionists， but retail is much messier. Packages vary too widely， and humanoid robots still cannot handle them efficiently enough.

Another route is education.

Beijing-based Booster Robotics demonstrated its Booster K2 soccer robot and has been building a sales and distribution network across mainland China. NOETIX Robotics， meanwhile， is marketing its Bumi robot to K-12 students as a tool for learning visual programming.

I have doubts about this direction. China's children's coding industry went through a franchise bubble not long ago. Coding centers expanded rapidly， then began closing in large numbers as the boom faded. With China's birth rate falling， robotics education could easily run into the same problem.

The least exciting but most reliable business remains the robot dog. Its production process is already standardized， and prices continue to fall. UNIUBI AI has pushed its price below $1，480， while ZSIBOT says it has shipped in large volumes. Unitree Robotics and DEEP Robotics continue to see steady demand across consumer， industrial， and inspection use cases.

Continue Reading