Summary

Zhao Zhelun divides quadruped-robot evolution into three generations. Boston Dynamics represents rule-based control, with Spot’s extremely high cost; Unitree cut prices through Chinese supply chains and design, after which the industry shifted to reinforcement learning, though robots still “walk blind” and rely on remote controls. Now embodied hardware, nearly a decade of autonomous-driving experience, and large models are converging at “the crossroads of three technologies.” Vita Dynamic has used that convergence to launch what he calls the first quadruped robot that does not require a remote control; it will officially go on sale by year-end, nine months after the company was founded and after 3 rounds of financing totaling approximately RMB300M. A new round is under way.
Zhao Zhelun’s response to Wang Xingxing’s claim that VLA is a “foolproof architecture” is restrained and pragmatic. VLA data sources are indeed inefficient and have generalization problems, “but there is no better solution at the moment, so we should choose the best solution currently available”; moreover, autonomous mobility is closer to the mature R&D paradigm of autonomous driving than autonomous manipulation is, which is the core reason the company is starting with quadrupeds and not manipulation.
His sharpest industry judgment is that Unitree’s hardware is “absolutely top-tier, even first-tier, in China and globally.” But the other unavoidable path for robotics is autonomy across perception, understanding, planning, and decision-making, and “the people who possess this cognition are not currently at Unitree—they are in today’s autonomous-driving and AI teams.” That also identifies the talent challenge for the next phase of embodied intelligence.
Going To C from day one was a deliberate contrarian choice. CEO Yu Yinan’s lesson from the previous AI cycle is that the customer in the To B business chain is the channel and commercial function: “a lot of the technology can simply be bought or carried over”; only To C can push the technical ceiling and retain the best AI teams. Choosing a quadruped over a humanoid likewise runs against financing incentives: “you have to give up a better or easier financing opportunity to do something more grounded, and that becomes a choice made by only a few.”
The underlying logic behind leaving Li Auto is an investment-grade framework. Physical AI has two axes—scene complexity and tolerance for error: driving is not complex (people learn it in 2 weeks to 1 month), but its tolerance requirement is “the highest of any product in the world I can imagine,” leaving L4 with a gap that is difficult to close. Autonomous driving therefore is not an independent operating unit: “the kind of intelligent driving needed to sell a car is the kind of intelligent driving it should become.” He wants physical AI to commercialize on its own.
He openly acknowledges the bubble: “There is definitively a bubble today.” The strategy is not to put the company in a position where it “cannot withstand the bubble bursting”—one product, 60-70 people, and normal operating expenses. He readily accepted the host’s metaphor: soap bubbles burst and disappear, but beer foam eventually settles into good beer.
His 10-year wager is that WRC will become the equivalent of the 1851 Great Exhibition in London within 5-10 years. Overseas visitors have already shifted from the “industry owner” mindset of the auto-show era to “coming purely to learn”; China has a chance to give birth to a robot culture in the way car culture emerged in Europe and the US. The end-state resembles WALL-E: robots of many forms, not necessarily humanoid, increasing “the happiness index of the physical world.”

Deep dive

1. Nine months in: RMB300M raised, 60-70 people, launch by year-end

The rapid-fire introduction establishes the basics: Zhao Zhelun is 31, an early Tesla China employee, founder of Garage No. 42, and Li Auto’s head of intelligent-driving products for 5 years. Last December, he co-founded Vita Dynamic with Yu Yinan, positioning it as “China’s first consumer-grade embodied-intelligence company.” By May this year, it had completed 3 rounds of financing totaling approximately RMB300M; a new round is under way, with valuation undisclosed.
The team was built from zero to 60-70 people across hardware, systems, software, AI, products, and marketing. The first product will officially launch by year-end, and the company currently has zero revenue. The annual OKR is not a simple volume target: “The most important thing is shipments… it should be NPS and shipments.” For the first-generation product, the bigger test is whether users buy it and “feel that this thing is genuinely pretty good.”

2. WRC was “a chance to get on the table,” and overseas visitors changed their posture

As early as March, he told the CEO explicitly: “This WRC opportunity is our chance to get on the table.” The 30-40-person R&D team arrived 2 days early, and more than 20 media representatives joined a workshop in the still-under-construction exhibition hall the day before the show opened—“that was the only window.” The venue was then packed; instead of touring during the final 2 days as planned, the team was so busy that they only managed one lap around the show by noon on the last day.
The more revealing change was in the audience. For at least the past 10 years, he had watched Europeans, Americans, and Japanese and Koreans visit Chinese auto shows to study China’s EV supply chain; “they still felt like the owners of the industry,” looking for what they could take home and use. At this WRC, there were few Americans, but visitors from Japan and Korea, the Middle East, and Southeast Asia were “coming purely to learn”—to see how far China’s robotics industry had progressed and what they could take home to sell.

3. The 10-year wager: WRC will be this era’s Great Exhibition in London

His forecast comes with a clear time frame: “In 5-10 years at most… China’s robotics conference will become something people from around the world come to see, and it will represent the highest level of robotics in the world.” His benchmark is the Great Exhibition in London around 1851, whose exhibits included steam engines and textile machinery—the source of the era’s energy.
His corresponding definition of robotics is a new “way of moving through space.” Whether it is a dexterous hand, robotic arm, or quadruped, “the essence is to move a mechanical object in a highly orderly way from point A to point B in space… to complete a certain task with certainty.” That capability can apply to any scenario and generate broad value.
A signal of US demand arrived the morning after the show. David, a friend familiar with the US market, called early and asked: “Can you bring me 2 units to the US?” American houses are larger and more older people live alone, so a product that moves autonomously without a remote control has a value proposition in the US that is “really not quite the same” as in China.

4. Three generations of robot dogs: from rule-based control to reinforcement learning, but still “walking blind”

The first stage was the era of robot control theory led by Boston Dynamics: programming and rule-based algorithms, with the coordination of 12 motors handled entirely through internal rules. Motor and hydraulic actuation were extremely expensive. Spot was “very strong in robot control theory,” but costly.
The second stage is represented by Unitree, which cut costs through Chinese supply chains and clever design. The Go To version sold to schools and researchers still costs nearly RMB100K, but is far cheaper than Spot. In 2022-23, the industry shifted from rules to reinforcement-learning gait training, moving from “one set of rules for flat ground and another for stairs” to adaptation across all terrain. The technical framework is almost identical to today’s exoskeletons.
But this generation is still called “walking blind”: gait changes rely on physical sensing rather than vision, and navigation still depends on a remote control. That is the dividing line for his third generation.

5. Three technologies converge into the third generation: autonomous mobility without a remote control

His company’s thesis is that 3 technologies are relatively mature today: quadruped hardware, which can “go up mountains and into water, and outperform normal human athletes”; nearly 10 years of autonomous-driving experience, where autonomous mobility is simply “moving from a faster transportation environment into a slower, more complex living space,” with lower speeds but more complex, less rule-describable, contact-rich scenarios; and natural interaction enabled by large models. “Our quadruped is the result of 3 technologies reaching a crossroads.”
The product form follows from that convergence: without a remote control, users can issue natural-language commands such as “come here,” “go there,” and “follow me.” Once both hands are free, “you can do a lot of things”—carry objects, film while following the user. Asked whether this is an industry consensus, he left room for uncertainty: “Most people should be able to agree, but it seems no one has broken it down and discussed it seriously in this way.”

6. Responding to Wang Xingxing’s “VLA is for idiots” argument: with no better option, choose the best one available

On the first day of WRC, Wang Xingxing made the provocative claim that the VLA architecture was “relatively foolproof” and that embodied data was insufficient. Zhao Zhelun accepted the premise: “A lot of the data in the VLA architecture comes from inefficient sources. In cars, there is already a very good data loop, but on robots the source material is relatively inefficient.” But “there is no better solution at the moment, so we should choose the best solution currently available.”
The second layer of his analysis is more important: upper-limb autonomous manipulation and lower-limb autonomous mobility should be viewed separately. On mobility, data collection, generalization, and using world models to extrapolate are closer to the autonomous-driving R&D paradigm; “VOA, or VON… can run relatively better on mobility.” The R&D paradigm for manipulation “is indeed not particularly efficient.” That is the core reason the company is entering through quadrupeds and not manipulation.

7. Why start with a dog instead of a humanoid: hardware maturity determines whether it can go into your home

He directly addressed the impression that the company chose an easy problem and disappointed its fans. The logic is straightforward: a startup going from zero to one “has to deliver a product,” which means choosing a mature hardware platform and a viable AI path. Quadrupeds are “definitely one of the most mature machine forms” in hardware; robotic arms come next, while “bipeds and dexterous hands are actually not particularly mature today.”
His test for maturity is worth remembering: “Making a demo and producing a video for you is not a problem, but if you want me to put this robot in a Hacker House and let you use it however you want, it will run into problems.”
In capability building, he considers the company full-stack: perception, interaction, execution units, AI’s large and small brains, controller chips, and energy systems. “The only difference from a humanoid is the execution unit”—whether there are 10, 20, or 30 motors. The sparrow may be small, but it has all the vital organs; once the system works, the organization will be ready for upper-limb manipulation and humanoid forms.
He also acknowledges that this path is unfavorable for fundraising: “Choosing a quadruped is more difficult than choosing a humanoid in terms of valuation and financing… you have to give up a better or easier financing opportunity to do something more grounded, and that becomes a choice made by only a few.”

8. Going To C from day one: Yu Yinan’s lesson about the chain-owner logic of the last AI cycle

Every robotics company they studied before starting up was To B, so they went the other way. The first part of their DNA is the need to deliver a product for users: “Companies with very strong technical capabilities generally serve users directly with their products.” Only that creates a chance to push the technical ceiling.
The second part came from Yu Yinan, one of China’s first-generation AI researchers who lived through the previous “four AI tigers” boom and shakeout, when the tide receded and Horizon Robotics surfaced. His conclusion was that the chain owner in To B is the B-side customer, so the company’s core capabilities become channels and sales: “A lot of the technology can simply be bought or carried over.” Only To C can make “the best technical teams in the industry willing to stay here.”
There is also an irreversible strategic judgment: once a company goes To B from zero to one, “it is difficult to go back and do To C.” The organizational model and R&D methods become fixed, so the road has to be chosen at the beginning.

9. Hands-on user research: robot dogs previously had very low NPS

Around February this year, the team conducted research in what Li Auto internally called “getting both hands dirty.” They manually searched Xiaohongshu, Weibo, Channels, and Xianyu for every user who had bought a quadruped robot, obtained contact information, and conducted in-depth interviews lasting 1-2 hours.
The conclusion was clear: there is a market for quadrupeds, but “first, people buy them to try something new, and second, they are easy to leave sitting around… although some people buy them to try them out, their NPS is very low.” The 4 major pain points were the need for a remote control, low battery life, loud noise at home, and frequent breakdowns. In March, the team made a special visit to a Roborock factory to study quality management at a mature robotics company; repair and return rates for quadrupeds had historically been quite high.

10. Not a toy: pricing constraints create a chain reaction that cannot support an AI team

The host observed that most robot dogs on Tmall and Xiaohongshu are sold as children’s toys, with monthly sales of 1,000-3,000 units. Zhao’s answer was a chain of business logic: once positioned as a toy—unless it becomes a designer collectible, which requires a different, artist-led model—“its pricing is necessarily constrained… The chain reaction is that it cannot use very good technology and cannot support the AI teams behind it. Commercially, the business simply does not work.”
They want something “fun and useful”: a robot dog that people can genuinely take outside today. Toy motors lack sufficient torque and the legs are too short to go outdoors; Unitree’s Go To can go out, but “you have to use a remote control to push it along… it is the same as playing with a remote-control car.”
The next layer of value is a multi-function backplate. The robot dog is a general-purpose mobile platform whose “ability to move and arrive basically reaches everywhere a person can walk.” The backplate supplies power and communications and can mount equipment—imaging devices, robotic arms, storage, “even equipment like a projector or a refrigerator can be placed on me.”

11. The trap of full-stack development: algorithms proven on Unitree were unusable on the company’s own hardware

Apart from components such as cameras, LiDAR, and battery cells, the system architecture, mechanical structure, and exterior design of the hardware are all developed in-house. That meant “making up all the coursework, making up the original Robotics course.” Before April, the company ran 2 tracks in parallel: the AI team used Unitree quadrupeds for secondary development and algorithm testing, while the hardware team built its own platform. Around April 28, the first-generation hardware A-sample was completed.
When software and hardware were combined in May, they hit a wall: “The things our algorithm people had made very maturely on Unitree’s hardware were completely unusable on our hardware… The gap here was huge.” He nevertheless gave the rival generous credit: Unitree’s hardware is “absolutely top-tier in China and globally, even first-tier.”
His honest progress marker is: “As of today, I cannot say it is completely solved; we have solved 80-90%.” But “we will definitely solve it completely before launch this year.” The payoff is organizational capability: bringing people from 2 completely different R&D paradigms—AI and robotics—onto one project, so once it works, “when I build more complex forms in the future, I can absolutely do it.”

12. Unitree’s 2 paths, and the judgment about where the talent is

He divides robotics capabilities into 2 paths. The first is extending hardware control: building platforms with more joints and degrees of freedom, greater complexity, and lower cost. “Unitree is pursuing the first path… its hardware keeps getting better and better, and that is also where Wang Xingxing and Unitree are very strong.”
The second is autonomy: moving from perceiving the environment to understanding information to planning, decision-making, and control—the AI path no one in the industry can avoid. This is the sharpest judgment of the discussion: “This path is our challenge, and it is also a major challenge for Unitree, because the people who possess this cognition are not currently at Unitree.” Asked where they are, he answered: “They are in today’s autonomous-driving and AI teams.”
The reverse is equally true for AI companies: their challenge is robotics. “AI teams need people who can accommodate robotics, and robotics people need to accommodate AI. Only then can a complete organization emerge.” That is the shared challenge for the next generation of embodied-intelligence companies.

13. The framework behind leaving Li Auto: L4’s tolerance gap, and why intelligent driving is not an operating unit

When end-to-end intelligent driving became the industry’s focus last year, he saw the possibility of building robots. Cars are planar mobility, while robotic arms, quadrupeds, bipeds, and dexterous hands “are essentially very similar”; their R&D paradigms are transferable. Behind this “Tesla has played a major role in driving the whole thing.”
But he also saw “a very large gap” in bringing L4 to market, and proposed a 2-axis framework: physical AI has a scene-complexity axis and a tolerance-for-error axis. Driving “is definitely not difficult in terms of complexity; people can definitely learn to drive in 2 weeks to 1 month.” AI is good at complex problems, but tolerance for error is exactly where AI is weak. Autonomous driving has “the highest tolerance requirement of any product in the world I can imagine,” because the cost of failure is extreme.
The commercial consequence of that gap is that L4 “is not itself an operating unit.” The industry’s terminology and marketing confusion stem from this: “The kind of intelligent driving needed to sell a car is the kind of intelligent driving it should become.” He did not want to spend another 5 years building something that could not become an independent business. Since 2024, he has “particularly wanted to do robotics,” even believing that once robots reach sufficient maturity, “they may in turn go and do L4; all of these things are connected.”

14. Li Xiang’s serious wide-eyed look: “If you have really thought it through, I still support you”

The resignation scene remains unchanged. There were only 2 people in the office. Li Xiang “stared at him with his big eyes” and asked, “Zhelun, have you thought this through?” He looked at Zhao with “a gaze I could not evade… At that moment, I could not lie to myself.” After Zhao confirmed his decision, Li Xiang said something that moved him deeply and that he would remember for life: “If you have really thought it through, I still support you.”
Li Xiang’s later input covered products, brand, and users. The Li Auto ONE-era methodology was that “the product is the brand”: without a real brand, “the greatest point of contact users have with you is your product; they learn who you are through the product,” so there is no need for anything flashy. Zhao added an observation: Li Auto has ample cash reserves and wants to do AI, yet has still made no appearance in robotics—suggesting Li Xiang also believes today’s hot humanoids “are not that mature yet.”
The partner list is a signal of mass-production intent: JD.com, Volcano Engine, DeepRoute.ai, and Horizon Robotics, as well as, according to him, likely Hesai and EVE Energy. “If you were just making a demo, you really would not need these partners to come together and do all this.”

15. Choosing a co-founder means handing over your life’s time; entrepreneurship is writing on a blank page

The standard for choosing a partner has nothing to do with capability: “Choosing a co-founder is equivalent to giving him your time.” The people he truly trusts are “brothers who have fought alongside you in the past.” He and Yu Yinan formed a bond through their work on Journey 3 and Journey 5 platforms at Li Auto and Horizon Robotics; he describes Yu as technically demanding, strong in engineering, open-minded, and not rigid. He also admits to a vulnerable side: if he had been entirely on his own, “I might not have been able to do robotics at all,” and might have gone back to making content. Yu found him and “pushed me forward another step.”
His contrast between working with Li Xiang and starting his own company is vivid: before, “Xiang-ge had already built most of the framework, and you were doing multiple-choice or fill-in-the-blank questions. Today it is a blank page.” The question is whether 5 years of growth has actually produced his own judgment about the market, users, and technology trends.
He retained Li Auto’s methodology—what problem the product solves, who the user is, and what “we want and do not want” at WRC—but discarded meetings that lasted 6 or even 10 hours: “We might talk for 3 hours and still not have said anything.” Communication is much more efficient at a small company. He also spent 6 months filling in his hardware gaps across motors, batteries, controllers, manufacturing, and quality management.

16. The end-state by scenario: outdoor quadrupeds, indoor wheels plus arms, and grippers over dexterous hands

Within the first 1-2 months of starting the company, he had worked out 2 product lines on a 10-year horizon. Outdoors, quadrupeds are “closer to the ultimate form”—more stable than bipeds and better across terrain than wheeled robots—with photography, exercise companionship, tea delivery, and every other outdoor service layered on top. Indoors, feet are a burden: walking is noisy and easy to topple, so wheels plus robotic arms are the right answer for household chores.
Simplification is strategy: “If wheels will do, don’t use legs; if one robotic arm will do, don’t use 2; if a gripper will do, don’t use a dexterous hand.” The first goal is to simplify the technology and product; the second is to lower the purchase threshold and cost. Today, a robot priced at RMB30K, RMB40K, or RMB50K is considered cheap, but he does not believe users can accept those prices. The company is cutting costs through supply-chain integration and in-house development; retail pricing has not been announced.
His assessment of Tesla’s teleoperation-for-data approach is that it is “suitable for Tesla but not for us.” Tesla has its own factories for unlimited trial and error, “100 or 1,000 times the R&D resources of a Chinese startup,” and needs a general-purpose narrative that matches its valuation and brand. The truly valuable teleoperation scenarios, such as high-risk jobs in paint shops, “are actually not particularly numerous,” while a small company cannot afford to invest purely in data collection.

17. A definite bubble, a robot culture born in China, and WALL-E in 10 years

He does not dodge the bubble question: “There is definitively a bubble today.” For people inside the industry, the bubble may even be beneficial by accelerating development, provided “you do not put yourself in a position where you cannot withstand the bubble bursting.” Vita Dynamic’s answer is restraint itself: 1 product, 60-70 people, and normal operating expenses. “Even if the bubble bursts, we will still be a relatively healthy company.” He called the host’s metaphor “particularly good”: soap bubbles burst and disappear, while beer foam eventually settles—there may be less good beer than people imagine, but there is still some, and it may well be valuable.
As a practitioner, he was not dazzled by WRC, but he saw the industry “eliminating the floor.” Last year, videos of robots falling over or rolling around on the floor were common; this year, they were rare. “Although it did not break through the ceiling as expected,” the baseline is rising. What moved him most was the creativity of children who bought tickets and signed and doodled on brand pillars: “Robot culture may have a chance to be born in China, a bit like car culture in Europe and the US.” Why do cars race in F1? “China has no answer to that question; the answer is in Europe and the US.” This time, the answer for robots may be in China. People are “a little too tense” about robot sports; culturally, “we should relax a bit.”
His 10-year vision has 2 analogies. The first is the golden age of Japanese and American electronics in the 1980s and 1990s—Walkman, Game Boy, digital cameras, Tamagotchi—ultimately integrated by Steve Jobs into the iPhone. “That process was beautiful, and it will happen again in robotics.” The second is WALL-E: robots of many forms, not necessarily humanoid, each with their own vitality, serving people in different spaces. Vita in Vita Dynamic means life. That also informs the product philosophy: if you can avoid speaking, do not speak. Drawing on LOVOT founder 林耀’s Warm Technology, he says that when verbal expression falls short of expectations, trust can be lost; tone, inflection, and body language can communicate better. By pulling people away from electronic information, “robots may have a chance to bring the happiness of the physical world back.”

The AI Whistleblower Initiative: Supporting AGI Insiders When It Matters Most, w/ founder Karl Koch Oren Cass & Noah Smith Debate the True Impact of Tariffs