Shulou(Shulou.com)11/24 Report--

Musk is my idol and it's an honor to have the opportunity to compete with him at this track.

Zhihui Jun, who said this, knew very well what he wanted to do when he left Huawei to start a business:

For the standard Musk, do general robot, competitor Optimus Prime, not only for the star sea but also for commercial landing, the final price of the product target does not exceed 200,000 yuan.

In the most popular parlance, he's combining big models and robots with embodied intelligence.

The idea began with his explorations as a student.

As early as 2015, when he had not yet become the master of Station B, Zhihui Jun, a student, participated in organizing a group of friends who knew each other in the competition and prepared to "be a person":

At that time, ordinary people's personal experience of AI was still such applications as Face Recognition or Beauty, and the concept of big model had not yet been born. The proposition of "To make robots have the ability to think and act like human beings" seemed a little too far from realization.

Therefore, after graduation, Zhihui Jun did not continue to toss only on the robot, but chose to enter the AI track, became the AI algorithm engineer of the big factory, and worked in the field of software algorithm for several years.

Until ChatGPT broke the sky and large model intelligence emerged, OpenAI proved that strong stacking can really produce qualitative changes.

Looking back at his robot dream, Zhihui Jun saw the feasibility of this road again. So, leave chrysanthemum factory, devote oneself to start a business.

So in half a year, he took the team out of a prototype.

At the August conference, Zhihui Jun proudly invited his own robot to the stage with steady steps, and the walking effect could be compared with last year's debut of Musk Optimus.

With such a work, his start-up Zhiyuan Robot has become one of the few targets to get VC and big factory bets at a time when investors are calm but the track is hot.

In a meeting with Qubit after the conference, he showed satisfaction with the conference, the team and the overall progress, and confidence in going down this road.

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

How would he view his company's high valuation if he devoted himself to starting a humanoid robot business? What is the next step of Zhiyuan's plan? Where will the commercial landing scene be? After starting a business, what should I do with the B station account of this top 100 up master?

Qubits ask, Zhihui Jun answers, everything is in this conversation.

Later, perhaps he would not have to start a business. Zhihui Jun was the first 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 issues in the field of science and technology, especially ChatGPT, which shocked everyone's violent aesthetics.

At that time, the track was far less crowded than it is today. It's no surprise that technologists are moving. But Zhihui Jun revealed his mentality at that time for the first time:

If he had chosen to be a little later, perhaps he would not have needed to start a business to do this.

To explain this sentence, he pulled out a timeline for his competitor Tesla Optimus Prime:

The prototype was shown last September, the ability to recognize ambient storage memory and accurately control object placement was demonstrated in May, 10 units were produced in July, walking tests are expected in November, and practical tests will be conducted at Tesla's own factory next year.

The action is very fast, and the actual effect is also wow.

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

When Zhihui Jun saw all this, his heart surged. At the same time, he sees the potential for combining robots and large models.

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

Among them, the most sensational Google RT (Robot Transformer) series, in the paper exposition and demo video, all show very good generalization ability in the field of capture.

It is not easy to do this, the core threshold lies in data. Just training it, Google spent 17 months collecting 130,000 real robot data from 13 bots-probably a small trick for RT model open source and data temporarily closed source.

Zhihui Jun admitted, Although also from the beginning in preparing their own action task data set, But at present can be used to train their own products data,"Only thousands."

The contrast was obvious.

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

It can also partially explain the question that when the robot is released, the netizens in the direct broadcast room have a short display time and the display ability is not as cool as imagination.

"I think it takes a lot of real scene data to get really good results. We haven't accumulated enough time in this area. "Zhihui Jun said so.

Competitors don't give you much time to breathe. Because of this, Zhihui Jun said that one of the team's next priorities is to build its own data center. It is planned to land in Hong Kong in the next few months, with the main purpose of building scenes and simulation platforms, filling in motion data and improving generalization ability.

How is the data generated? Zhihui Jun's summary has three parts:

Supervised learning data. Rely on people to do demonstration, control robots to do sorting and other operations, 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 a better human-computer interaction process device.

AIGC generates data. It is mainly the supplement and expansion of real data, and it is an important means of low-cost data enhancement.

Building a data center is one of the key points of work, and another is to iteratively reconstruct the hardware structure to enhance the motion performance of the robot ontology. According to Zhihui Jun, teams iterate hardware with the speed and efficiency of agile software development.

It's a very disruptive and challenging thing to do.

Here's a little gossip.

In April this year, Zhihui Jun submitted a self-made two-legged robot Nezha at Station B, and said at the end of the video,"If nothing unexpected happens, Nezha will become the colorful egg of the conference."

Of course, according to the classic plot, if there were no accidents, there would be accidents: D, Nezha did not attend the press conference.

That's it

Quantum bit help you find out, the reason is Nezha above several motors used by suppliers pigeon, March orders received in July, resulting in insufficient development time.

Zhihui Jun said: "Later free or will continue to complete this project, pigeon king pit required." "

Half a year into a unicorn, there is another commercial hidden line behind reading here can probably be detected, Zhiyuan humanoid robot project robot body, still need to iterate several versions; and AI ability is limited by the current lack of training data, also need to accumulate for a period of time.

Overall, the product seems to be a long way from landing.

However, it is such a start-up project that the half-year valuation goes straight to $1 billion. Is that reasonable??

Hearing this question, Zhihui Jun did not directly answer to the end "value or not," just replied, in fact, financing ideas are not finalized at the beginning, during which a lot of industry leaders and senior suggestions were referred to.

The team's initial vision was simple: to make a Demo first and then let nature take its course to raise the valuation.

"However, starting a business is obviously not such a simple matter. Under the background that the general economic environment is not prosperous, to integrate resources and attract talents, there is no endorsement without capital. The author states,"This is not a bad thing. "

Efficient and practical work, rapid adjustment of ideas and strategies, result-oriented promotion of things, this is Zhihui Jun's style of doing things. His style ultimately determined the work style of Zhiyuan, at least the Zhiyuan R & D team.

Every early member of the technical team was personally introduced into the company by him. More than 30 people were basically self-rolling. Zhihui Jun, who slept five or six hours a day, felt deeply:

Before everyone thought I was a master of time management, now I can be ashamed to say that our company is so liver like me, and there are a bunch of…

Throughout the communication process, qubits noticed that he emphasized two key words,"cost reduction" and "application scenario."

These two are common pain points throughout the current track. How the team solves the pain point must be the key for them to get the heavy bets of Hillhouse, CDH, Jingwei, Gao Rong, Lanchi, BV Baidu Venture Capital, etc.

Let's listen to Zhihui Jun's opinion_

Let's start with keeping costs down. The slogan Zhiyuan shouted to the outside world was that the humanoid robot landed at a price of less than 200,000 yuan.

This is about the same price as Musk's $70,000, while the price level of similar humanoid robots in China is around 500,000 RMB, and the popular Boston Dynamics Atlas costs $2 million.

Zhihui Jun bluntly said: "It is not that we hope to achieve 200,000 yuan, but if we cannot achieve the price of 200,000 yuan, there is no way to realize commercial landing." "

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

Zhihui Jun also roughly described the method of Zhiyuan to control costs for mass production.

One is to adopt self-research route as much as possible to reduce cost and increase efficiency.

For example, joint motors and dexterous hands account for more than half of the hardware cost, and there are still mismatched characteristics in the existing supply in 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 Tesla's car-building, supplement the precision requirements of hardware with software and algorithm, and reduce hardware cost.

Such as giving up harmonic reducer and choosing planetary reducer, vision closed-loop scheme used on dexterous hand, etc.

Let's talk about landing application scenarios. Zhihui Jun said, It is expected that the second half of next year commercialization landing, First applied in the industrial manufacturing field, Family and other service-oriented application scenarios are ranked back, At this stage highlight a "Scene is relatively simple, Task is relatively complex."

"Is this route groped out as you go, or is it a goal set in the first place? "

"At the earliest stage when the team was less than 10 people, we basically finalized this landing plan. "and expresses:

Many people will compare our or Tesla's robots to Boston Dynamics, which is actually inappropriate.

The correct logic should be to implement the solution at the lowest possible cost on the premise of meeting the functional and performance requirements of the application scenario.

So in a situation where it can walk and carry things, there's no need to have it capable of backflips.

Now, this route set six months ago, the final landing signs are clearer.

The latest industrial and commercial developments show that BYD has taken a stake in Zhiyuan's subsidiaries, plus the previous qubit learned from Zhiyuan, the company has been in close negotiations with domestic head smart car manufacturers and 3C manufacturers.

From this, before the official announcement, it was speculated that the first place for Expedition A1 to work in the industrial manufacturing field was BYD Automobile Factory without accident.

In addition to the cost and landing scenario, as the team deepened their embodied intelligence, Zhiyuan also held some other differentiated cards.

For example, qubits are dug from Zhihui Jun's mouth to a hidden line of commercialization of the company_

If the universal humanoid robot is the medium-and long-term plan and final vision of commercial landing, then the team will also have some product forms of "laying eggs along the way" towards this end point.

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

He had heard many questions about why robots were made in human form, and whether it was a waste of effort compared to other specialized forms (robotic arms, wheels).

In response to this problem, Zhihui Jun has two aspects to think about.

On the one hand, this will be a long-term investment process, we do not overestimate the short-term value, do not underestimate the long-term value. Humanoid is the first step toward the ultimate form of the AI, which is why the robot is called Expedition.

On the other hand, doing this (human form) is chosen not because it is easy, but because it is difficult.

Universal humanoid robot involves the most comprehensive robot technology stack. In its realization process, various cutting-edge technologies (self-research and optimization of various technologies such as visual servo, MPC, SLAM, LLM / VLM, middleware, etc.) lay eggs along the way, which can lead to many innovative special robot products."Everyone will see these achievements for six times in the future."

"Netizens can rest assured that the account number of Station B will not become a special marketing number for the company." Excellent technology, beautiful resume, aura of large factory, millions of fans, and coinciding with the outbreak cycle of new technology: large model, personal intelligence, AIGC... After starting a business, he served as the team CTO and led the company to grow rapidly. In half a year, the number of the company was nearly 100, and the market valuation exceeded US $1 billion.

A series of stories with halo came crashing down, and the onlookers could not help but re-examine Zhihui Jun at this time. How would he define himself now? The qubit threw the problem to Zhihui himself.

Zhi Huijun did not show any hesitation. He only said that the position he gave himself did not change much.

First an engineer, then an entrepreneur.

I may belong to an atypical entrepreneur. The motivation to do these things is based on personal interests. I am also lucky to have made some small achievements: early in the big factory, I gained some halo, gained a lot of popularity online, and then suddenly ran out to start a business. At first, everyone around me was shocked.

I have always considered myself an optimist.

There was one more thing he said without hesitation-

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

"Every step is the right choice," which may also include many people leaving Huawei and leaving the "genius"tag.

In front of qubits, he did not hide his gratitude for having won honors for his work at Huawei. He also mentioned:

There are some great things being done by the old team, but more future-oriented exploration like robotics may be better done in small innovation teams.

I hope I can inherit the fine tradition of 'scattering all over the sky'.

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

I can feel that Zhihui Jun hopes that the outside world will shift the attention to him from a specific tag to what he wants to do itself.

Interestingly, he also advised everyone not to start a business too early. "for students, i suggest that they should first work for a few years and accept the beating of society before they can have a clear understanding of the way society and companies operate, haha. "

At the press conference, he also expressed his view that one of the most effective ways to test the value of a technology company is to see if it can be commercialized.

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

Speaking of this, Zhihui Jun revealed his heart. Since he started his business, the project could not be carried out according to his own whimsy. After organizing people and taking money, we need to think more about the strategic development direction of the company,"responsible for so many brothers and sisters of the company."

But apparently, he has his own place: Station B.

"Personal number? I've been too busy lately (so I haven't updated). "Zhihui Jun explained that he absolutely did not mean to dig a hole and run away." After a little leisure, it will still be updated. "

As for the post B station submission, it will also be the original style, the original taste, may occasionally be mixed with some entrepreneurial daily. But he promised that it would not become a purely corporate marketing window.

(He hinted that Zhiyuan has a separate official account after all. Everyone is welcome to pay attention ~)

"It is also to learn from Musk. He has done a good job between company management and personal account operation. "

One More Thing Talk about this, how can you not ask the king pigeon, when will the next B station video update be?

"This year definitely, definitely before the end of the year. "

What does the content relate to?

"Robot again, one of the holes dug before, this is the next video. "

Well, with my authorization, we've put his message here.

Gu.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.