Exclusive Interview with ZHANG Wei of Zhuji Dynamics: Failing to Solve the Last Mile Will "Kill" the Deployment of Robots

Text by | Qiu Xiaofen

Edited by | Su Jianxun

If you've paid attention to the presence of "Zhujidongli" at various exhibitions in recent years, you might have a deep impression of this scene - the staff and visitors suddenly lift their legs and kick a small bipedal robot again and again. The robot staggers away and then regains its balance.

Zhujidongli's bipedal robot

When this robot became their traffic magnet, many people might think that "Zhujidongli" is just a company that makes robot legs. That's wrong. Zhang Wei, the founder of "Zhujidongli", believes that what they really want to do is "the NVIDIA in the field of embodied intelligence".

Indeed, in today's hot embodied intelligence track, "Zhujidongli" has a unique positioning. While many manufacturers are telling stories about the "brain" and even using VLA as a promotional slogan, Zhang Wei believes that the performance of the current specific model is not the most important thing.

In his opinion, there is no unified technical solution for the "brain" at present. It's useless to burn money to collect real - machine data. The ability to extract useful information for operational tasks from different data is the key - Zhang Wei calls it the "machine tool for producing embodied models".

Based on this judgment, "Zhujidongli" hopes to help downstream customers develop the "brain" by providing the body, the "cerebellum", and the model development toolchain. Zhang Wei said frankly, "We don't have the know - how to enter specific scenarios." Therefore, in terms of business choices, "Zhujidongli" seems to have "taken a step back" compared with most robot companies.

Zhang Wei abstracted their target customer profile with three letters - "IDS": Innovator, including scientific research institutions and technology companies; Developer, those who mainly develop new functions based on existing technologies; System Integrator, who integrate existing technologies and functions into implementation solutions.

Through these three types of users, Zhang Wei hopes that "Zhujidongli" can serve the innovation process of service robots, including technological innovation, development innovation, and solution innovation.

Zhujidongli's humanoid robot Oli

This business judgment mainly stems from Zhang Wei's observation of the current situation of robot implementation. In 2025, implementation is a sword hanging over most robot manufacturers. Everyone is trying in scenarios such as factories, healthcare, and retail. However, Zhang Wei believes that "letting robots replace humans is something that can easily create illusions".

In his opinion, the fixed investment in hardware and software for robot implementation only accounts for 10% - 20% of the entire value chain. The subsequent deployment and maintenance are the biggest costs.

It's not hard to find that the current situation of robot implementation in the industry is that to replace a worker with a humanoid robot, it often requires "several top - notch algorithmic doctors". Moreover, the more projects a company takes on, the more it loses. Zhang Wei said bluntly, "The last 10% of robot implementation is enough to kill the first 90%."

In early August, "Intelligent Emergence" had a three - hour conversation with Zhang Wei. Rarely seen in public, he continued his usual sharp - tongued style, liked to convey anti - consensus views, and was good at highly abstracting his own opinions.

He shared his observations on robot implementation, his judgment on the debate about robot forms, and his views on the currently popular robot "brain". The following is the transcript of the conversation between "Intelligent Emergence" and Zhang Wei (slightly edited).

OPEX, the biggest illusion of robots

"Intelligent Emergence": What do you think are the bottlenecks preventing robots from being truly implemented so far?

Zhang Wei: I think embodied robots have already been implemented. Companies like Xingyuanzhe and Mekabot are AI - driven robot companies. They've been working in factories for quite a long time. Although it's tough, they've actually been implemented. They have a "brain", just not a perfect one. They have perception, positioning, and decision - making abilities, not just pure programming.

"Intelligent Emergence": In your opinion, what is the standard for implementation?

Zhang Wei: We have our own judgment on the commercial implementation of robots. It's easy to make robots work, but don't think that just because they're working, they're implemented. The key is how much effort it takes to make them work. What's even more difficult is whether robots can truly systematically replace humans to improve efficiency and reduce costs.

So, why is it difficult for robots to be implemented? The reason is that people see robots seemingly replacing a worker and then calculate the pay - back period for robots replacing workers. But in fact, this is just the tip of the iceberg.

I think in the entire value chain, letting robots replace humans is something that can easily create illusions. It only accounts for 10% - 20% of the entire value chain. It's not difficult to use the robot's body and software to replace humans, but maintaining and deploying the robots is the biggest cost.

When it comes to robot implementation, people often focus on CAPEX (Capital Expenditure), that is, the fixed investment, and forget that in fact, OPEX (Operating Expenditure) is the biggest and most difficult part.

"Intelligent Emergence": What does the specific OPEX of robots include?

Zhang Wei: You need to estimate and predict what technical variables are needed for maintenance and operation, and at the same time, you need to have a good understanding of the business.

However, in reality, those who understand the business don't understand the technology, and those who understand the technology don't understand the business. So, the estimation deviation is very large, which makes people easily underestimate this matter.

Even if 90% of the robot implementation has been achieved, the remaining 10% may take one or two years or even more time. 90% seems quite high, but the last 10% can be fatal. For example, to replace a worker, you may find that you still need to involve several top - notch algorithmic doctors.

I think the current embodied "brain" is a very valuable technical variable. But if the business model is not well - thought - out and the implementation scenario is not clear, saying that you want to use it in a factory means you don't even have the qualification to lose money. It's not the start of making money, but the start of losing money.

This is the lesson from the previous generation of robots. Spending so much money on maintenance will drain the company. The more projects a company takes on, the more it loses. This is the common situation.

It's the same with Robotaxi. It's not a technical problem. Its business essence is an operational problem. The entire business closed - loop requires many links, A, B, C, D, E. No matter how good A is, if B, C, D, and E are not mature, it won't work.

"Intelligent Emergence": We can see that some robot companies' commercial actions are not to find a high - requirement scenario for implementation as you mentioned, but to make small gains along the way. Is this also a consensus in the industry now?

Zhang Wei: I think there are people who think this way, but not all. You need to have a deep understanding of the entire business chain of the scenario you choose, not just the technology. You need to understand the maturity and scarcity of technology in each link.

I think it's quite difficult for robot manufacturers to cover multiple fields and implement in various scenarios. Because the last mile requires a lot of time and effort, and the accumulated threshold behind it is often even higher than the technical threshold. It tests the industry's know - how, the company's position in the industry, and the advantages of the model. Unless you think your core value lies in a deep understanding of the red - ocean industries.

Most of the scenarios in the robot field that can be profitable have become highly competitive. In factories, everything that can be automated has been automated, and the remaining tasks don't seem to be easy.

Zhujidongli's humanoid robot carrying boxes

Making the upper body of a robot is not creation

"Intelligent Emergence": Based on this business understanding, how do you promote the implementation of robots? How do you deal with the fierce competition in the red - ocean market?

Zhang Wei: Currently, it's hard to see immediate applications for humanoid robots. In the long run, they don't have a significant efficiency improvement. However, I think it's a very important form and the optimal solution for CAPEX.

Because for any specific task, as long as the task is simple enough and the volume is large, a general - purpose robotic arm can handle it. Since the birth of robots, the transformation from automation to robots has been a change from specialized to general - purpose.

From the perspective of robot forms, we think there are four types of robots:

The simplest is the robotic arm with only arms;

The second type is the wheeled dual - arm robot that you can see in 90% of the exhibitions. I call it a person in a wheelchair;

The third type is the real humanoid robot;

The fourth type is the robot with only legs and no upper body.

The wheeled dual - arm robot solves the problem of moving on flat ground. The "legs" of the robot solve the problem of adapting to different terrains to move from point A to point B.

"Intelligent Emergence": You've chosen the latter two forms. Why are the legs so important?

Zhang Wei: When we talk about the body, we also need to talk about the "cerebellum" and motion control. The "cerebellum" and the body are coupled. There's nothing easier to make than a robot. Robots are much easier to make than cars. The reason why people can't make them is whether they can control them well.

In the 1980s, industrial robots emerged, and traditional industrial motion control algorithms were developed in the industry, giving birth to the four major robot families (ABB, Yaskawa, KUKA, and Fanuc).

This algorithm model, based on dynamics, kinematics, and calculation, is already very mature and can be basically applied to wheeled dual - arm robots. Therefore, the wheeled dual - arm robot doesn't need AI - enabled "cerebellum".

We think that the forms of wheeled dual - arm robots and robotic arms are mature and highly competitive. The entire technical solution is very mature. It could be done before, and it can be done now. It's just that there's a current trend. This is not creation but a supply - chain logic.

Why do we choose to make legs and humanoid robots? Because legs and humanoid robots are new. A humanoid robot with legs must have an AI - enabled "cerebellum" because it needs balance. AI - enabling can greatly shorten the entire development cycle.

Zhujidongli's humanoid robot Oli

"Intelligent Emergence": What's the difference between having an AI - enabled "cerebellum" and not having one?

Zhang Wei: Whether the current "cerebellum" is AI - enabled or not, it is based on models. The difference lies in the qualitative change in the way people use models.

In the early days of the four major robot families, people used models by understanding and deriving them through symbols - people designed the controller through their understanding of things and models and the deduction of symbols.

By the time of Boston Dynamics, the way of using models changed from symbols to calculation. I need to calculate a solution in some optimized way. The essence of AI - enabling, I think, is to generate a lot of data with models in the simulator and then train the controller through the data. This is the most fundamental change.

Currently, the robot "brain" is being tested in various scenarios in a piecemeal way. The humanoid form is the one that can best exert the value of the "brain".

"Intelligent Emergence": Who are your customers and what are their characteristics?

Zhang Wei: We've divided them into three parts, IDS - Innovator, which includes scientific research institutions and technology companies and is mainly responsible for creating new technologies; Developer, these people usually don't invent new models but develop new functions based on existing technologies; The third part is System Integrator, who integrates various technologies and functions to develop solutions for specific applications.

We don't serve end - users. We serve the innovation process. IDS corresponds to technological innovation, development innovation, and solution innovation respectively. We are the "NVIDIA of embodied intelligence". Models are not my way of making money.

"Intelligent Emergence": When did you come up with this IDS customer system?

Zhang Wei: It was finalized this year.

"Intelligent Emergence": I feel that your model is different from others. Your positioning is more like a tool - type company. But everyone is focusing on scenario implementation this year. It seems that you're more cautious about implementing in various scenarios?

Zhang Wei: No, we also do it, but in a way of collaborating with our IDS customers. We don't have the know - how to enter specific scenarios. We can say that we've taken a step back compared with most robot companies.

In the implementation of embodied intelligent robots, we focus on the underlying platform, providing the body, the "cerebellum", and the model development toolchain, corresponding to the iPhone, iOS, and Xcode. We help our customers develop "apps" to use the robots. Different "apps" complete different tasks, and we aim to cultivate a developer ecosystem to empower various industries.

Don't directly copy the autonomous driving model when making robots

"Intelligent Emergence": There is no consensus in the robot industry on the importance of data. Some people think that data should be collected after the model is determined, while others think that good robot models can't be developed without data. What's your judgment on data?

Zhang Wei: My judgment on data is that if there is enough data, there's no need to argue about end - to - end. Robots can do any task. But the problem is that this "if" doesn't hold.

"Intelligent Emergence": Does the data here refer to real - machine data?

Zhang Wei: It refers to valuable data.

"Intelligent Emergence": It seems that you don't emphasize real - machine data but pay more attention to simulation data?

Zhang Wei: Reinforcement learning on real machines is an important recent technological progress. Autonomous driving is a relatively simple form of embodied intelligence. In the