Conversation with Mr. Zhihui: The window of embodied intelligence is fleeting, and the challenge is aimed at the idol Musk

Source: Qubit

Author: Hengyu

Original title: "Conversation with Mr. Zhihui: The window for embodied intelligent entrepreneurship is fleeting, and the challenge is aimed at the idol Musk"

Musk is my idol, and I am honored to have the opportunity to compete with him on this track.

Zhihui Jun who said this is very clear about what he wants to do when he leaves Huawei to start a business:

Benchmark Musk, make general-purpose robots, and the competing product Optimus Prime. It is not only facing the stars, but also rushing to commercial implementation. The final price of the product is no more than 200,000 yuan.

To put it in the most popular terms, he is making embodied intelligence that combines large models and robots.

The idea began during his student days.

As early as 2015, before he became the owner of Station B, Zhihuijun, who was a student, participated in organizing a group of friends who met through competitions and prepared to "become an individual":

At that time, ordinary people's first-hand experience of AI was in applications such as face recognition or beautification. The concept of large models had not yet been born. The proposition of "making robots have the ability to think and act like humans" seemed a bit too far away from being realized. out of reach.

Therefore, after graduation, Zhihui Jun did not continue to work only on robots. Instead, he chose to enter the AI track and became an AI algorithm engineer at a major manufacturer. He worked in the field of software algorithms for several years.

Until ChatGPT broke the ground and large-scale model intelligence emerged, OpenAI proved that vigorous stacking can really produce qualitative changes.

Looking back at his robot dream, Zhihui Jun saw the feasibility of this path.

So, I left the chrysanthemum factory and devoted myself to starting a business.

So within half a year, he led the team to come up with a working prototype.

At the press conference in August, the self-made robot that was proudly invited to the stage by Zhihui Jun had a steady pace, and the walking effect was comparable to that of Musk’s Optimus Prime’s debut last year.

It is with such a work that his start-up company Zhiyuan Robot has become one of the few targets to receive heavy investment from VCs and major manufacturers at a time when investors are calm but the market is hot.

In the meeting with Qubit after the press conference, he expressed his satisfaction with the press conference, the team and the overall progress, and also expressed his confidence in going along this road to the future.

He also told us in front of his workstation that his desktop wallpaper has been changed for a long time:

What would he think of the high valuation of his company when he devoted himself to starting a humanoid robot business? What is Zhiyuan's next plan? Where will the commercialization landing scene be? After starting a business, what should I do with the B-station account of the top 100 up owners?

Qubit asks, Zhihui Jun answers, everything is in this conversation.

If it’s later, maybe you won’t have to start a business

Mr. Zhihui was the first group of people to perceive this wave of environmental changes and put them into practice.

From the outside world, when he announced his business at the end of last year, robots and large models were hot topics in the field of technology, especially when ChatGPT just debuted, which shocked everyone with violent aesthetics.

At that time, the track was far less crowded than it is today. It is not surprising that technology practitioners are moved by the wind.

But Mr. Zhihui revealed his state of mind at that time for the first time:

**If you choose a little later, maybe there is no need to come out to start a business and do this. **

In order to explain this sentence, he pulled out a timeline for the competing Tesla Optimus Prime:

The prototype was displayed in September last year. In May this year, it demonstrated the ability to recognize the surrounding environment, store memory and accurately control the handling of objects. In July, it said that 10 units had been produced. It is expected to conduct a walking test in November, and it will be practical in Tesla's own factory next year. sex test.

The action is very fast and the actual effect is amazing.

Musk, a man who never plays by the rules, has been advancing the pace of mass production at a rocket-like speed since he announced his entry into the robot track. According to him, the number of robots will exceed humans in the future.

Seeing all this, Zhihui-jun was overwhelmed.

At the same time, he also observed the potential of combining robots and large models.

Microsoft's ChatGPT for Robotics, Google's Palm-E, RT-1, RT-2, VoxPoser, RoboCat and many other efforts are all trying to transfer the capabilities and knowledge of large image language models to the field of robotics.

Among them, the most sensational Google RT (Robot Transformer) series has demonstrated excellent generalization capabilities in the field of crawling in paper explanations and demo videos.

It is not easy to do this, the core threshold lies in the data**.

Just to train it, Google spent 17 months and collected 130,000 pieces of real robot data from 13 robots - this is probably a small trick because the RT model is open source and the data is temporarily closed source.

Zhihui Jun admitted that although he has been preparing his own action task data set from the beginning, the data that can be used to train his own products is currently "only a few thousand."

A comparison, the big gap is obvious.

The intuitive gap in numbers is enough to explain what Mr. Zhihui said to qubits, "The current node, the place where Expedition A1 needs to be improved the most is the AI generalization ability brought by data", which is enough to explain why this "will be a relatively long-term Layout".

It can also partially explain the doubts of netizens in the live broadcast room that the display time is short and the display ability is not as cool as imagined when the robot is released.

"I think that in order to achieve the ideal practical effect, a large amount of real scene data is still needed. Our time is too short, and we have not accumulated enough in this regard." Zhihui Jun said.

Competitors won't give you a chance to breathe. Because of this, Mr. Zhihui said that one of the next focus of the team's work is to build its own data center**.

It is planned to land in Lingang in the next few months. The main purpose is to build a scene and simulation platform, fill in motion data, and improve generalization capabilities.

How is the data generated? Zhihui Jun has three conclusions:

• **Supervised learning data. **Rely on people to do demonstrations, control robots to do various operations such as sorting, and collect real data in the process.
• ** Simulation data. **You need to build your own simulation platform, with a better rendering engine, physics engine, and better human-computer interaction process device.
• **AIGC generates data. **It is mainly the supplement and expansion of real data and an important means of low-cost data enhancement.

Building a data center is one of the key points of work, and the other key point is to iteratively reconstruct the hardware structure and enhance the motion performance of the robot body. **

According to Mr. Zhihui, the team will use the speed and efficiency of agile software development to iterate hardware.

This is a very subversive and challenging thing.

Here is a little gossip.

In April this year, Wisdom Army submitted Nezha, a self-made bipedal robot from Station B, and said at the end of the video, "If nothing else happens, Nezha will become an Easter egg at the press conference."

Of course, according to the classic plot, if there is no accident, there will be an accident :D, Nezha did not attend the press conference.

That’s it↓

Qubit helped you find out that the reason is that several motors used by Nezha were rejected by the supplier, and the order placed in March was not received until July, resulting in insufficient development time.

Mr. Zhihui said: "I will continue to complete this project when I have time later. The Ace Pigeon must fill in any pitfalls."

Become a unicorn in half a year, and there is another hidden line of commercialization behind it

After reading this, you can probably notice that the robot body of the Zhiyuan humanoid robot project still needs several iterations; and the AI capabilities are limited by the current lack of training data, and it will also take some time to accumulate.

Generally speaking, the product seems to be still some distance away from being launched.

However, for such an entrepreneurial project, the half-year valuation went straight to 1 billion US dollars.

Is this reasonable? ? ?

Hearing this question, Mr. Zhihui did not directly answer "whether it is worth it or not", but only replied that in fact, the financing idea was not finalized at the beginning. During the period, he referred to the suggestions of many industry leaders and seniors.

The team’s initial idea was very simple, which was to make a demo first and then increase the valuation naturally.

"But starting a business is obviously not a simple matter. In a context where the general economic environment is not prosperous, in order to integrate resources and attract talents, capital endorsement is required." Zhihui Jun said,** "This is not a bad thing. .”**

Working efficiently and effectively, quickly adjusting ideas and strategies, and advancing things with a results-oriented approach, this is Mr. Zhihui’s style of doing things.

His style ultimately determines Zhiyuan, at least the working style of Zhiyuan's R&D team.

Every early member of the technical team was personally interviewed by him, and most of the 30-odd people were self-reported, and Mr. Zhihui, who slept five or six hours a day, was deeply moved:

Everyone thought I was a master of time management before, but now I can say with shame that there are a lot of people in our company who are as lazy as me...

During the entire communication process, Qubit noticed that he emphasized two keywords, "cost reduction" and "application scenarios".

These two are the common pain points of the entire track at present. How the team solves the pain points will definitely be the key to them getting heavy bets from top investors such as Hillhouse, CDH, Matrix Partners, Gaorong, Lanchi, and BV Baidu Ventures.

Let’s hear Mr. Zhihui’s opinion——

Let’s talk about lowering costs first.

Now Zhiyuan’s slogan is to control the price of humanoid robots within 200,000 yuan.

This is about the same as the price of US$70,000 that Musk announced, while the price level of domestic similar humanoid robots is around 500,000 RMB. The cost of Boston Dynamics Atlas, which everyone loves to see, is US$2 million.

Mr. Zhihui said bluntly: "It's not that we hope to achieve 200,000 yuan, but if the price cannot be achieved at 200,000 yuan, there is no way to realize commercial landing."

As for why it is 200,000, he said that taking the new energy automobile manufacturing industry as an example, if 200,000 robots replace some manual positions, the return on investment period can be 1 to 2 years.

Mr. Zhihui also briefly described Zhiyuan’s method to control costs for mass production**.

**One is to take the self-research route as much as possible to reduce costs and increase efficiency.

Components such as joint motors and dexterous hands account for more than half of the hardware cost, and there is still a feature mismatch in the existing supply on the market. Independent research and development of core components can reduce the cost by more than half.

The second is to adopt some ideas similar to those used by Tesla in building cars, using software and algorithms to supplement the accuracy requirements of the hardware and reduce hardware costs.

Such as abandoning the harmonic reducer and choosing a planetary reducer, the visual closed-loop solution used on the dexterous hand, and so on.

Let's talk about landing application scenarios.

Zhihui Jun said that it is expected to be commercialized in the second half of next year. It will be applied in the industrial manufacturing field first, and service application scenarios such as homes will be moved to the back. At this stage, one thing is highlighted: "the scenarios are relatively simple, and the tasks are relatively complex."

"Did this route be explored while walking, or was the goal set at the beginning?" “We basically finalized this implementation plan at the earliest stage when the team was less than 10 people.”

At the same time, it was stated that

Many people will compare our or Tesla's robots with Boston Dynamics, which is actually inappropriate. To achieve commercialization, the most correct logic should be: on the premise of meeting the functional and performance requirements of the application scenario, implement the solution at the lowest possible cost. ** So in the scene where it can walk and move things, there is no need to make it capable of backflips.

Now, the signs of final implementation of this route set half a year ago are becoming increasingly clear.

The latest industrial and commercial trends show that BYD has taken a stake in a subsidiary of Zhiyuan. In addition, Qubit has previously learned from Zhiyuan that the company has been in close negotiations with leading domestic smart car manufacturers and 3C manufacturers.

Therefore, before the official announcement, it was speculated that the first part-time job location of Expedition A1 in the field of industrial manufacturing, not surprisingly, is the BYD Automobile Factory.

In addition to cost and implementation scenarios, as the team deepens its understanding of embodied intelligence, Zhiyuan also holds some other differentiated cards.

For example, the qubit dug out a hidden line of the company's commercialization** from Jun Zhihui's mouth——

If universal humanoid robots are the mid- to long-term plan and ultimate vision for commercialization, then in the process of moving towards this end point, the team will also have some "laying eggs along the way" product forms.

What is the specific form? Mr. Zhihui was as strict as ever and kept it secret, but he still revealed something.

He has heard many questions, asking why the robot should be made into a human form. Compared with other special forms (mechanical arms, wheeled), is it thankless?

Regarding this issue, Mr. Zhihui has two thoughts.

On the one hand, this will be a long-term investment process, so please don't overestimate the short-term value, and don't underestimate the long-term value**.

The human form is Zhiyuan's first step towards the ultimate form, which is why he named this robot "Yuanzheng".

On the other hand, choosing to do this thing (human form) is not done because it is easy, but rather because it is difficult.

The universal humanoid robot involves the most comprehensive robot technology stack. In the process of its realization, various cutting-edge technologies (self-development and optimization of various technologies such as visual servoing, MPC, SLAM, LLM/VLM, middleware, etc.) are laid eggs along the way. , can give birth to many innovative and specialized robot products**, "everyone will see these results one after another in the future."

"Netizens, please rest assured that the Bilibili account will not become a company-specific marketing account"

Outstanding technology, beautiful resume, the aura of a big factory, millions of fans, and coincides with the outbreak cycle of new technologies: large models, embodied intelligence, AIGC... After starting a business, he served as the CTO of the team and led the company to quickly In half a year, the company has nearly 100 employees, and the market valuation exceeds 1 billion US dollars.

A series of stories with a halo came over, and the onlookers couldn't help but re-examine Mr. Zhihui at this time.

How would he define himself now? Qubit put this question to Zhihui Jun himself.

Mr. Zhihui didn't show any hesitation, he just said that his positioning for himself had not changed much.

** An engineer first, an entrepreneur second. **

I may be an atypical entrepreneur. The motivation for doing these things is based on personal interests. I have also been lucky enough to achieve some small achievements: I shined in a big factory in the early days, gained some halo, and gained a lot of success online. popularity, and then suddenly ran out to start a business. Everyone around me was shocked at first. I have always considered myself an optimist.

There is another sentence that he said without hesitation——

"Since there is no chance of regret in life, then insist on believing that every step I have taken so far is the most correct choice I have made."

“Every step is the right choice” may also include many people lamenting about leaving Huawei and breaking away from the “genius” tag.

In front of the qubits, he did not hide his gratitude for the honor he had worked in Huawei, and also mentioned,

The old club is doing some great things, but the exploration of more future fields like robotics may be more suitable to be done in a small innovative team. I hope I can inherit the fine tradition of 'the sky is full of stars'."

Then, as he often did, he emphasized again that he was neither a genius nor a teenager.

It can be felt that Mr. Zhihui hopes that the outside world will shift his attention from a specific tag to what he wants to do. **

Interestingly, he also advised everyone not to start a business too early. "For students, it is recommended that they work for a few years and accept the beatings from society before they have a clear understanding of how society and companies operate, haha."

At the press conference at that time, he also expressed his opinion: One of the most effective ways to test the value of a technology company is to see whether it can be commercialized.

Otherwise, no matter how good the technology is, it is easy to fall into self-excitement.

After talking about this, Mr. Zhihui expressed his thoughts. Now that he has started a serious business, the project cannot be based on personal whims and whims.

After recruiting people and taking money, you need to think more about the strategic development direction of the company, and "you have to be responsible for so many brothers and sisters in the company."

But obviously, he has his own place to enjoy himself: Station B. **

"Personal account? I've been too busy these days (so I haven't updated it)." Zhihui Jun explained, saying that he had absolutely no intention of digging a hole and running away. "I will update it later when I have some free time."

As for the subsequent contributions to station B, will still be in the original style, the original taste may occasionally be mixed with some entrepreneurial daily life.

But he assured that it will not become a purely company's marketing window.

(He hinted that after all, Zhiyuan has a separate official account, everyone is welcome to follow it~)

"It's also learning from Musk. He has done a good job between company management and personal account operations."

One More Thing

At this point in our conversation, how could we not ask Ace Pigeon, when will the next video update of Station B be?

"It will be certain this year, and it will be certain before the end of the year."

What is the content related to?

"It's still a robot, a certain hole that was dug before, this is the next video."

Well, with my authorization, we put the words here for him.

Cuckoo.