4 Hours with SEER Robotics’ Zhao Yue: First Listed Robot Brain Firm

2026/07/06 · Source link · AI Summary archive

Summary

SEER Robotics is not really betting on training an all-knowing, all-capable embodied brain today; it is first taking control of the data gateway to real machines. Its control system supports more than 2,000 machine types and over 400 kinds of sensors, with tens of thousands of units installed and the base “basically doubling every year.” That gives it diverse, naturally aligned, continuously generated field data at low cost. Zhao Yue’s foundational view is that “data is more or less equivalent to the model”: model techniques may lead by only 3 to 6 months, while installed base, scenarios, and time cannot be accelerated so easily.
Zhao Yue’s core reservation about the current humanoid-robotics path is not a bias against the form factor, but that the data and ROI loops have not closed. Humanoid data faces three hurdles at once: it is hard to obtain, incomplete across modalities, and difficult to screen for quality. In industry, the most economical answer for moving one ton of goods is neither 10 humanoids carrying it nor a humanoid driving a forklift, but “making the forklift itself embodied.” His market split is clear: To B offers the larger incremental opportunity, while To C offers the larger disruptive one.
SEER Robotics is training vertical experts such as forklift and cleaning specialists before trying to distill all its data into one general-purpose model. Zhao believes the future robot will more likely use a mixture-of-experts architecture, calling the specialist required for each task; he even argues that “it’s not that the GPT moment for robotics hasn’t arrived—the Transformer moment may not have arrived either.” The plan is to let the “forklift master” and “cleaning auntie” create value and generate data in bounded environments, then combine the raw materials when a new paradigm emerges.
SEER Robotics is moving from the world’s No. 1 mobile-robot controller to a platform, with its moat built on years of taking on the dirty, difficult, exhausting work others did not want. It supports more than 400 classes of sensors and a broad range of field requirements, building the union of customer needs rather than the intersection: “If you change one, I’ll adapt one.” At sufficient scale, that naturally becomes a de facto standard. The disruptive risk is equally clear: if an edge model can understand the protocol of any component and generate its own control capabilities simply by plugging it in, today’s accumulated toolchain could be erased in one stroke by a higher-dimensional technology.
High private-market valuations and SEER Robotics’ public-market path represent two entirely different risk-reward structures. The high-flying companies are buying a tiny probability of an enormous future payoff; companies with revenue and deployment are immediately judged on market size, PS, PE, and ROI. Zhao warns that much of the fundraising in China is “actually borrowing,” and a RMB5B financing round can carry RMB5B of potential buyback obligations. Even with $1B in hand, he would spend the next 3 to 5 years expanding supported machine types from 2,000 to 20,000, building overseas channels, and increasing installations—not buying data at roughly RMB500 an hour to train a general-purpose brain.
Zhao traces SEER Robotics’ organizational capabilities back to RoboCup: failure becomes systems debt, while championships come from exhausting the details, the opponent, and uncertainty. An own goal in the 2012 final, caused by forgetting to substitute the goalkeeper, gave him the lasting belief that “anything you think is a problem will come looking for you sooner or later.” The 2013 penalty-shootout win over CMU and the 2014 2:0 regulation-time victory came from 8-millisecond decisions, frame-by-frame log analysis, reverse-engineering programmers’ magic numbers, and even deliberately creating the “largest gap” for CMU to target. That capability later became the company’s approach to product development, field logs, and customer corner cases.
SEER Robotics’ long-term thesis is that every equipment company will eventually become a robotics company, so the B2B market may never have one giant that supplies every robot body. No company monopolized the world’s production lines during the automation era; the intelligent era may likewise not need one standard robot. Controllers, data formats, and heterogeneous coordination may become the more durable infrastructure layer. Zhao sees “one to 100” as the stake for the next “zero to one,” and defines first place as a moving target that can be approached but never attained. If someone makes “intelligent machines without barriers” work better, SEER Robotics should “die fighting.”

Deep dive

1. SEER Robotics Debuts as the Global No. 1 but Still Sees Itself as a Slime

卫诗婕 calls the newly listed SEER Robotics a “true hidden champion”: it sells the brains—the control systems—of mobile robots worldwide and holds the No. 1 global market share. Zhao Yue still describes the team’s starting point as “a little nobody.”
The company’s mascot is a slime: a pun on SLAM, and a game monster that can either be killed with one slash or kill the player with one slash. It can change form, matching SEER Robotics’ ambition to provide systems for every kind of robot.
Zhao admits to an inner rebelliousness that wants to “always see ourselves as a little nobody and always try to overturn the table,” but rejects calling the company a big boss: “Once you become the big boss, someone else may eliminate you.”

2. Refusing Corner Cases Became the Starting Point for SEER Robotics’ Rise from the Margins

When Zhao was doing R&D, he “especially hated rejecting customer requests.” If a request existed, there had to be an unseen reason behind it; the right move was to dig deeper with the customer, not dismiss it as unreasonable on the first pass.
His belief is close to absolute: “I don’t believe there is such a thing as an unreasonable customer.” The corner cases that manufacturers overlooked or considered bizarre became the entry point for SEER Robotics to work 100 times harder.
That inclusiveness became part of the company’s mission: “Accelerate an open, diverse intelligent civilization and make intelligent machines accessible to everyone.” Zhao even says the company has “no values,” because an organization trying to create a diverse world cannot impose too many constraints on itself first.

3. The Long Middle Between Wrong and Right Is Supposed to Be Chaotic

The “wrong” and “right” on SEER Robotics’ walls can switch places when viewed from different angles. Zhao believes things change dynamically and perspectives differ, so it is better to take a few more steps than rush to label them; many things become “right” only after you keep doing them.
His definition of entrepreneurship is similarly plain: keep solving problems, then see who solves them faster and who can still climb out after falling into a pit. “Experience teaches; easy lessons can be learned. Teaching people never works.” Every pitfall described by motivational slogans still has to be stepped into personally.

4. A Hunan Town, Teacher Parents, and an Exam Ace Produced a Student Who Resisted Constraints

Zhao grew up in Shaoyang, Hunan, in a small town. Both parents were teachers but rarely controlled him. As a child, he spent more time on the street following other children around and catching small animals; he says he was always the follower, not the leader.
He was not a conventional top student so much as an “exam ace”: someone who did little studying but still scored above average. Entering Zhejiang University’s Chu Kochen Honors College eight-year medical program also reflected a fair amount of simply going with the flow.
His parents rarely made decisions for him, which later became a management principle carried into the company: do not leave people entirely unsupervised, but let them make their own choices first and bear the consequences.

5. The Clash Between Medicine and Electronic Information Eventually Drove Zhao Out of the Eight-Year Program

Zhejiang’s eight-year medical program began with a nonmedical undergraduate degree and “maybe 30 or 40” prerequisite medical credits, followed by doctoral training in the fifth year. The goal was not to produce ordinary clinicians, but physician-scientists who could take clinical problems back to the lab and then feed the results into practice.
Zhao chose electronic information engineering because he wanted to see Zhejiang University’s supposedly elite “double-E” major. Once he began studying it, he found engineering to be a system of derivable, reviewable logic, while medical experiments often required repetition, patience, inspiration, and luck.
Studying medicine during the day and building robots across campus at night caused the two modes of thinking to grow increasingly “split.” After roughly 3 months in the fifth year, he concluded he could not immerse himself in medicine, withdrew, and prepared for an engineering graduate entrance exam without telling his parents beforehand: “If I told them, the three of us could only be anxious together.”

6. The Lab Replaced the Dorm, and Robotics Turned from an Extracurricular Credit into Life Itself

By the end of undergrad Zhao rarely returned to his dorm. Graduate school was more extreme: after leaving his clothes in the dorm, he almost never showed up again. The lab’s soccer field was covered with carpet; he tuned robots during the day and lay down to sleep there at night.
Someone was present in that lab almost continuously for 4 or 5 years. The air conditioning, lights, and equipment rarely stopped. Zhao does not call it “involution,” because to the people involved it felt like spending time in an Internet café—just fun.
He initially joined the university robotics competition only to earn 2 second-class activity credits. In a 3-person team, someone had to do the work; he kept going, won first place, and was then admitted to the more advanced RoboCup lab.

7. A 0:2 or 0:3 Loss Made the “Wheel Picker” Determined to Fight Back

Around 2009, Zhejiang’s team was beaten 0:2 or 0:3 by Dalian University of Technology in the domestic final. Zhao was only a picker: when a robot’s wheel came off, he ran onto the field, retrieved or swapped the machine, and hurriedly screwed the wheel back on.
Because he was somewhat overweight, each sprint shook the floor and the overhead camera. The opposing team even shouted, “Little fatty, don’t run so fast.” The poor result turned the humiliation into a concrete thought: “Next time, I have to beat them back.”
After a year of training, Zhejiang won 5:0 in Lanzhou. Zhao describes growth like a game: a mini-boss knocks you down, so you go back to level up and put on better gear before returning to settle the score. Continuous positive feedback reinforced both his enthusiasm and his persistence.

8. Early Simulation Meant Building Your Own Wheels, While GPUs Were Already Driving Decisions in 2009

There was no mature Isaac-style platform at the time. The team built its own simulator on physics engines such as ODE and used CUDA to calculate the state of play, dividing the position space into millimeter-scale cells to speed path and strategy searches.
Such simulation treated the robot as a whole, which was enough to study trajectories, formations, and intelligent decisions. Modeling every joint, complex motion, or reinforcement learning remained far beyond what those tools could do compared with today’s detailed simulators.
Zhao says sim-to-real depends on the level of the problem. For overall trajectories and game decisions, the tools of the time were adequate; for joint dynamics and fine manipulation, they were nowhere near today’s capabilities.

9. RoboCup Compressed 6 or 7 Robots into an 8-Millisecond Centralized Decision Cycle

An overhead camera shared the entire field image with both teams. The system had to identify the robots, ball, velocities, and formations, predict the opponent’s behavior, and generate simultaneous control commands for 6 or 7 robots.
The image refreshed roughly every 8 milliseconds, and the complete decision had to finish within those 8 milliseconds. Robots could reach 3 meters per second, so a millimeter-level position difference could determine a pass or defensive move.
Soccer appears to have few inputs, but its state space is enormous, like Go, with an additional layer of strategic interaction between two sides. Zhao found competition more interesting than writing papers because the uncertainty came from a real opponent rather than a closed problem.

10. Opponent Logs, Randomized Tactics, and Magic Numbers Formed an Early Version of Adversarial Learning

The team collected match images and logs from opponents, studied their preferred zones, defensive weaknesses, and set-piece choices, then learned tactics through rules or neural networks. The hardest opponent was not the most complex one, but the “unpredictable opponent.”
SEER Robotics’ predecessor therefore introduced randomness deliberately: run strategy one this time, then randomly switch to strategy two next time, preventing the opponent from inferring the fourth-game playbook from 3 matches.
Zhao also reverse-engineered programmers’ coding habits. Thresholds were often 5 or 10, rarely primes or other irregular numbers. If he inferred that an opponent used 5 as a boundary, he would place a robot at 6: “Once it’s over 5, he leaves—and 6 is the closest point to his half.”

11. RoboCup Used a 2050 Endgame to Break an Expensive Problem into Leagues Anyone Could Enter

RoboCup began in 1998, the year of the World Cup. Its professors set the ultimate goal of assembling a robot soccer team by 2050 that could defeat the human World Cup champion. Zhao believes the field appears to be approaching that goal in recent years.
To reduce the cost of a single school building a full human-sized team, the competition was split into bipedal, wheeled, simulated, and other leagues, further divided into kid-size, middle-size, and adult-size categories with different resource requirements and research difficulty.
Zhejiang competed in the small-size wheeled league, focusing not on bipedal locomotion but on multi-robot coordination, passing, and autonomous decisions. Each category would make its own breakthroughs before ultimately contributing to the 2050 objective.

12. An Unknown Team Knocked Out the Defending Champion in 2012, Then Left the Final with an Own Goal

Zhejiang reached only the quarterfinals in 2011. The year Zhao dropped out and retook the graduate entrance exam, he had no exam pressure and spent nearly the entire year in the lab “picking at details,” embedding new tricks from planning algorithms to team coordination.
In the first game of the 2012 world championship in Mexico City, Zhejiang faced the defending champion: a combined team from Thailand’s 4 best universities, with 20 to 30 people in attendance. Zhejiang had only 4 or 5 people, yet won 2:0 or 3:0, drawing congratulations even from the organizing committee chairman.
The teams met again in the final, but the Thai side had spent 5 sleepless days studying Zhejiang’s tactics. In the first half, Zhejiang’s goalkeeper chipped the ball, hit a defender’s backside, and deflected it into its own goal; the team eventually lost 1:2. Zhao compared the moment to the regret the protagonist team had to carry in Slam Dunk.

13. Zhejiang Beat CMU on Penalties in 2013, Then Took the Coaching Duel to Its Limit in 2014

When CMU returned in 2013, it did not initially view Zhejiang as its main rival. Both teams went undefeated into the final, where regulation play could not separate them; Zhejiang won the penalty shootout and claimed the first world championship ever won by a Chinese team in the physical-robot category.
The teams met again in 2014, when Zhejiang beat CMU 2:0 in regulation. When one side called a timeout to adjust tactics, the other immediately called one after seeing the change. It was no longer a simple robot-versus-robot contest, but a live battle between 2 coaches.
Zhao considers it one of the best small-size finals ever played. Algorithms, formations, opponent analysis, and in-game adjustments were all pushed to the limit; the result no longer hinged on any single technical point.

14. “Anything You Think Is a Problem Will Eventually Come Looking for You” Became an Entrepreneurial Imprint

Before the 2012 final, Zhao had already noticed that the goalkeeper’s chip was too low and joked about avoiding an own goal off the defender’s backside. A junior member seriously identified the best substitute, then forgot to make the actual substitution under pressure; the feared problem materialized in the most theatrical way possible.
Zhao formed a lasting rule from the episode: “If you think something is a problem, it may not happen now, but it will definitely show up later.” A problem can be solved later, but the debt cannot be treated as nonexistent.
Emotional stability does not come from willpower alone; it depends on preparation. The more contingencies and remedies you have, the less you panic. In 2013, the team built both breadth and depth to an “insanely detailed” level. Under that degree of preparation, Zhao says, “it would have been hard not to win.”

15. The Smarter CMU Was Made Predictable by a Deliberately Created Gap

CMU’s robots constantly repositioned and always passed to the player in the largest open space. Other teams chased them closely and were often beaten by half a step; Zhejiang instead deliberately created a seemingly enormous gap.
Because CMU would inevitably select that point, its intelligent, dynamic strategy became predictable again. Zhejiang’s rear robot accelerated in advance and moved to enclose the ball and receiver together.
Zhao keeps the example because it captures a competitive principle: do not merely follow a strong opponent’s moves. Understand the mechanism behind its decisions, then make it walk into your optimal solution.

16. The Championship Was Only RoboCup’s Surface Goal; the Real Output Was the People It Planted

Zhao later discussed RoboCup with one of its founding professors, who said that defeating the human champion was only the explicit goal. More important was cultivating people who genuinely loved robotics and sending them into the industry like seeds.
His examples include the captain of the 2004 champion team, who later founded Kiva; Amazon acquired the company for more than $700M. Several Chinese participants also went on to start robotics companies.
In 2017, Zhao returned as a corporate mentor and led students from Shanghai Jiao Tong University to another championship. He also wanted to follow CMU’s example by publishing papers and some code: “Student competitions should be more open.”
After the first championship in 2013, Zhejiang’s recruitment pool jumped from more than 20 people to more than 300. Yet 50 or 60 would still drop out round by round, leaving fewer than 10—not the people with the most impressive résumés, but those willing to sleep in the lab for years.

17. Recruiting Starts with “What Do You Love?” Because Passion Cannot Run Against the Company’s Direction

Zhao asks almost every interviewee: “What do you love most? What makes you passionate?” He does not require a perfect overlap with the company, but the angle between the 2 directions should at least be acute.
When one controls student said he really wanted to write novels, Zhao advised him to go write immediately rather than work for the sake of working. Taking a job at the cost of abandoning what you truly want to do for life is too expensive.
Passion is difficult to fake. When people discuss something they genuinely love, their eyes light up and they can answer detailed follow-ups without pause. Zhao’s own answer is to push the boundary of robotics, solve problems, and put the solutions into the field so the feedback loop closes.

18. The Robot Pyramid Says Defects Should Be Eliminated as Close to the Base as Possible

Zhao breaks a robot into a pyramid rising from mechanical hardware and electronic circuits through low-level software and the operating system, communications, algorithms, and models.
The rule is: “If you can solve something at a lower layer, never extend it upward.” A mechanical defect should be fixed mechanically; a circuit problem should not be pushed onto low-level software; something low-level software can handle should not become an algorithmic patch.
The higher up a lower-layer problem is compensated for, the greater the cost and the more the system’s extensibility suffers. The principle comes from competition engineering and forms Zhao’s first-principles view of layered architecture and the end-to-end debate.

19. Zhao Believes More in MoE-Style Expert Systems and Says Robotics Has Not Reached Its Transformer Moment

For complex, variable robots constrained by power consumption, Zhao expects the future to look more like MoE: call the small expert appropriate to the scene rather than constantly running one all-knowing, all-capable model.
An end-to-end full-scale model is “beautiful,” but Zhao remains explicit and cautious: “It’s not that the GPT moment for robotics hasn’t arrived—the Transformer moment may not have arrived either.”
He therefore will not make a firm bet today on a universal model. SEER Robotics will first validate, close the loop, and accumulate data in small, clearly bounded environments, then test a larger unified architecture when the data and paradigm are mature enough.

20. Top-Down Brain Companies Are Not Wrong; They Simply Accept a Steeper Probability Distribution

A new startup that rebuilt from the bottom would not have enough time, and it would struggle to explain why it should retrace an incumbent’s lead. Starting from the model at the top of the pyramid is often the only realistic story it can tell and finance.
Dollar-based investors are more willing to buy a grand future. Of 1,000 companies pursuing that kind of top-down ambition, perhaps only 1 becomes something like Anthropic—but that 1 will probably come from the group. Of 1,000 pragmatic entrepreneurs, perhaps half will be decent, but extreme returns may never emerge.
卫诗婕 points out that the path is difficult to validate or falsify in the short term. Zhao acknowledges that “aiming high” can make action purer, but success may be enormous and failure may be “total ruin.” Neither style is abstractly right or wrong.

21. SEER Robotics’ Modest Fundraising Reflects Its Preferences, Cash Flow, and View of Where Rewards Come From

Zhao admits he is “not good at fundraising.” Zhejiang’s practical culture also leads the company to understate itself: after achieving 80 or 90 points, it may communicate only 70 or 80, which is a weakness in capital markets.
He cares more about the reward from customers paying than the reward from investors funding. In the early years, cash flow was adequate and the company was not forced to make fundraising its top priority.
On the club of companies valued at RMB10B or more, he says, “We’re not envious; after listing, we’ll have plenty of money too.” In China, fundraising often comes with buyback clauses, so raising RMB5B can also mean taking on RMB5B of potential debt.
Money merely means the company is “holding it in custody.” It must eventually answer what value it created with the funds; without that value, a full bank account is just debt waiting to be called.

22. Hot Money Can Compress the Trial-and-Error Cycle, but Industry Consensus Is Not Truth

When resources are limited, a path may take 3 years to test; with enough funding, that can be compressed to 6 months. Zhao believes this helps not only an individual company but also allows the industry to eliminate wrong answers faster.
Chinese talent moves quickly and secrets have short half-lives. After one company validates a path, the entire industry may follow within 3 months. VLA, world models, real-machine data, simulation data, and collecting data without a proprietary body have all gone through rapid divergence followed by rapid consensus.
Even without investing as aggressively, SEER Robotics can benefit from other people’s trial and error. 卫诗婕 jokes that this is the mentality of a “bad kid”: let others raise more money and eliminate more answers, then SEER Robotics absorbs the consensus.
Zhao sets one standard for genuine consensus: it must create verifiable value for customers or society in deployment. Otherwise it is only synchronized market storytelling.

23. New Technical Paradigms and New Product Forms Must Be Validated Separately

SEER Robotics views embodied intelligence along 2 axes: old technology to new technology, and old product forms to new product forms. Bipeds, wheeled humanoids, and quadrupeds are product forms; VLA and world models are technical paradigms.
The 2 dimensions do not need to be tied together in a single all-or-nothing test. A new paradigm can first enter a mature existing scenario, while a new form can initially run on mature control methods.
This split lets SEER Robotics follow the latest models without putting all its chips on the upper-right corner of “new technology × new form,” avoiding 2 layers of unresolved uncertainty at once.

24. Humanoid Data Faces Three Hurdles: Acquisition, Completeness, and Screening

First, humanoid robots lack large amounts of existing human-operation data. Second, even if vision and action are captured, touch, force, and other critical modalities may be missing, leaving the data incomplete.
The third hurdle is more subtle: the high degree of freedom makes it hard to define a good action. A cup can be picked up from different directions with different gestures; none is necessarily wrong, so screening criteria cannot converge quickly.
Autonomous driving would not treat red-light violations as good data because the rules are clear. Humanoid manipulation lacks such a clean evaluation boundary. Zhao therefore does not want to validate a new paradigm on humanoids first.

25. Forklifts Are SEER Robotics’ First Real-World Testbed for Generative Control

Forklift environments have clear boundaries, abundant human-driving data, and no need for the fine touch and full-body degrees of freedom required by a humanoid. Task completion, efficiency, and safety all have assessable standards.
SEER Robotics is already building vertical models for forklift scenarios. They are no longer merely rule-based, but use the current generation of data-driven, generative methods, and the company says it is seeing some results.
Data can continue to flow at low cost only after a scenario creates value. If the robot does not create value for the customer, the company has to buy data or provide equipment for free to collect it, making the process expensive and unsustainable.
Zhao’s ordering is: “Models are of course important, and the technical paradigm is important, but the scenario may be relatively more important.”

26. To B’s Focus on Unit Economics Makes Specialized Tools Better Than General Humanoids in the Short Term

Zhao’s market view is that To B offers the larger incremental opportunity while To C offers the larger disruptive one. Without attractive unit economics, B2B robotics cannot scale.
There are 3 ways to move one ton of palletized goods: 10 humanoids carrying it, 1 humanoid driving a forklift, or making the forklift itself embodied. The third is cheapest and most consistent with the history of humans inventing tools to raise productivity.
“Tools were designed for people, so humanoids are the most general-purpose” is a valid argument. So is the alternative: bring the cost of embodying each specific tool down far enough.
He does not oppose humanoids loading and unloading, and even a “thousand-armed Guanyin” with 5 or 10 arms is possible. The standard is not whether it looks human, but the ROI in the specific scenario.

27. If Every Equipment Company Can Build Robots, There May Be No “Robot Company” in B2B

Zhao’s “provocation” is: “In the future every equipment company will be a robotics company, so there will be no robotics companies.” Manufacturing giants such as Foxconn will gradually develop robotics capabilities.
No single company produced the overwhelming majority of industrial production lines during the automation era. Factories assembled the line best suited to their own products, efficiency, and yield, while standardizing only infrastructure layers such as PLCs.
Robots are production tools for the intelligent era, and the production lines they serve are inherently nonstandard. If each industry can define its optimal body, there is no reason to assume one standard humanoid will cover all B2B use cases.
Rebuilding industry as a humanoid manual-work line could even be a step backward: “People don’t want to work that way, and robots certainly won’t.”

28. Existing Scenarios Create an Innovator’s Dilemma but Also Supply the Rarest Model Feedback Loop

卫诗婕 recounts 王启彬’s change of view. He once thought incumbent companies with customers and scenarios were more likely to win, then worried that existing revenue would bind them to old paths.
Zhao accepts the risk but reverses the question: the most important thing for a model is still data, so SEER Robotics first answers where the data comes from and only then how to build the model.
Beyond “only companies that train their own models are qualified to define good data,” he goes one step further: only when a model is actually used in a scenario can you know which data is effective.
RoboCup’s value was its unified evaluation system. Without field results, models, sensor layouts, and data formats can all claim superiority, but there is no way to know which is actually better.

29. SEER Robotics Defines Good Data as Real, Diverse, Aligned, and Sustainable

Zhao is a “committed real-machine-data advocate” and believes real-machine data must have a scaling law, even though the industry has not yet found the corresponding paradigm.
Good data begins with real scenarios and diversity. Next, data from different sensors, actuators, and machine types must be naturally unified and aligned rather than purchased and then subjected to enormous cleaning costs.
The third condition is continuous, low-cost generation. If every hour of data is expensive and can be obtained only through fundraising, it lacks a natural commercial foundation and cannot support long-term model scaling.
These standards are not static specifications. What counts as “good” must ultimately be iterated through model training, field use, and customer value.

30. More Than 2,000 Machine Types and 400 Sensors Make the Controller a Natural Data Bus

SEER Robotics’ controllers support more than 2,000 machine types and over 400 sensors, capturing data from lidar, cameras, and actuator trajectories across a broad range of B2B forms and scenarios.
Every sensor connects to the controller and is processed through unified extrinsic calibration and automated labeling tools. Whether it is Company A’s lidar, Company B’s camera, or Company C’s servo, the resulting data is captured in SEER Robotics’ format.
The scenario makes the data nearly free. When a robot malfunctions, customers often send the logs to SEER Robotics themselves; those corner cases resemble autonomous-driving takeover data and are high quality.
Zhao says this infrastructure is the source of the company’s confidence. Not rushing to mix everything into a large model does not mean it is unprepared for the future.

31. The “Forklift Master” and “Cleaning Auntie” Work Separately Before Being Blended Later

SEER Robotics currently keeps data segregated by scenario. Forklift data trains the “forklift master,” while cleaning data trains the “cleaning auntie,” avoiding a blind mixture of humanoid, cleaning, and logistics data.
The first benefit is that each expert can deploy and create value today while continuously receiving new data. The second is that all experts still share the same underlying format.
If a genuinely effective unified paradigm emerges, the raw materials from hundreds or thousands of scenarios will already be ready. SEER Robotics can then “throw everything in and see what comes out.”
This is Zhao’s version of the MoE path: build each expert first, then discuss routing and fusion, rather than pretending to possess an omniscient model on day one.

32. The Controller’s First Strategic Role Is to Turn Installed Base into Future Model Optionality

SEER Robotics wants more companies to create different robots on its controllers. The controller lowers the development barrier, ensures data-format consistency, and uses the installed base to keep expanding the data source.
The current installed base is in the tens of thousands and is “basically doubling every year.” Zhao believes that as long as SEER Robotics maintains the largest installed base and the data remains highly aligned and inexpensive, it will not be caught off guard in the future.
Compared with a single-body company, the difference is not only richer data. A single B2B form cannot be universal; SEER Robotics must rely on many partners to place its controllers across specialized bodies.
The strategy serves current revenue while buying a long-term option on a model paradigm that has not yet appeared.

33. Adapting to Every Standard Instead of Declaring One Made SEER Robotics the De Facto Standard

卫诗婕 points out that as more controllers are sold, users increasingly work within SEER Robotics’ interfaces and data system. The company does not claim to set a standard, but it is creating one from the bottom up.
Zhao believes top-down standard-setting campaigns usually fail because partners receive no direct benefit from complying. A standard emerges only after a product creates value and reaches enough users.
SEER Robotics does not ask lidar, camera, or servo makers to rewrite their protocols. It accepts single-line, multi-line, explosion-proof, different-brand, and even different protocols within the same product line: “If you change one, I’ll adapt one.”
He invokes the Tao Te Ching: “The movement of the Way is reversal.” The more aggressively a company imposes standards around itself, the less likely it is to become one; the more it lowers itself and embraces differences, the more likely the industry is to adopt it.

34. SEER Robotics Is Not Selling Data Yet; Even with $1B, It Would First Reach 20,000 Machine Types

SEER Robotics does not plan to sell data at this stage because Zhao believes it is not yet “good enough” and the internal loop has not closed. Monetizing too early would let different customers define the data and create massive internal friction.
Over the long term, he believes good data should be shared. If the core business is already profitable, data is simply a high-quality by-product, and sharing it is more natural than locking it away.
If SEER Robotics suddenly received $1B, Zhao still would not immediately train a general-purpose brain. He would expand supported machine types from 2,000 to 20,000, accelerate overseas channels, and increase installations and diversity over the next 3 to 5 years.
Market data that once cost about $100 per hour later fell to roughly RMB500. Zhao still considers that expensive, and more importantly, “it isn’t the data my own scenarios truly need.”

35. Model Know-How May Lead by 3 to 6 Months, but Data Must Be Bought with Money, Scale, or Time

Zhao acknowledges that brain companies will develop their own techniques for data production lines, automated labeling, and industrialized training, while model algorithms themselves also have barriers.
But once a method is proven effective, Chinese competitors can spread it quickly. He describes that barrier roughly as “a 3-to-6-month barrier.”
Data cannot be replicated immediately with a paper. A company must either spend enough money to buy it, install enough machines to generate it, or spend enough time accumulating it; at least one of those costs must be paid.
SEER Robotics therefore has not made “raising enough money” a strategic objective. Fundraising matters only when it can be converted into more installations, scenarios, and high-quality data.

36. From Anxiety About Being Called “Old School” on VLA to Re-centering the Strategy on Data

Two years ago, an investor told Zhao that he would not even read a BP without VLA and asked about SEER Robotics’ VLA progress. He even predicted that the old methods would be replaced within 6 to 12 months.
SEER Robotics had only experimented with VLA and had not made a real commitment. Zhao admits that the investor’s “highly dramatic” description of the future triggered more anxiety than reading papers alone, as if the industry had suddenly switched to a new game he did not yet understand.
The team spent roughly 6 months studying and observing rather than immediately overturning the old system. The spotlight then shifted from VLA to world models, and Zhao realized that change was not moving as fast as the market imagined.
Once the anxiety shifted from methods to data, he looked again at SEER Robotics’ installed base and scenarios and realized that “on this particular question, we actually aren’t nervous.” The objective became clear again.

37. The New Brain Team Combines Cloud Computing, Autonomous Driving, Academia, and Internal Transfers

SEER Robotics broke down the required capabilities: internet and cloud companies are good at large-scale data processing and cost reduction; autonomous-driving talent understands training methods and engineering optimization; academia is closer to frontier paradigms.
The team therefore combines external advisers, external “people who understand,” and internal transfers rather than rebuilding from scratch.
The R&D organization had about 100 people at the time. The company opened exploration to anyone interested and lightly applied the brakes to the old rule-driven path, gradually shifting time toward data-based methods.
Zhao initially worried that veteran employees would resist learning. He found the opposite: people who truly love robotics were more worried that the company would not pursue new technology. “You can’t stop them.”

38. New Methods Do Not Replace the Old Organization; They Lay Eggs Along the Way and Upgrade the 90-Plus-Point System

SEER Robotics moved its most promising people into the new team, then promoted new leaders from the old business. The transformation also became a round of talent renewal.
Methods developed through new-model exploration are migrated back into the old control system. Optimization that once moved slowly from 90 to 91 points may find a different route through the new paradigm.
卫诗婕 cites Microsoft Research Asia’s research-product-feedback loop as a reference. Zhao says the key is not only technology but whether the organization allows research-oriented and deployment-oriented people to be evaluated differently.
A multi-robot scheduling team might even buy a robotic arm to study reinforcement learning and find ideas on another platform. A project need not map directly to someone’s job or prove short-term performance first.

39. SEER Robotics Maintains Order Through Energy, Interest, and Closed-Loop Problem Solving—not a Unified Value System

Zhao says that deliberately organizing cultural activities and CEO speeches can shape employees into one fixed personality. SEER Robotics would rather let people “not know what the boss is thinking” and preserve an open atmosphere.
The organization includes big-company hires, people who grew within the company from their student days, and employees who moved from warehouse management into technical leadership. “There are many problem employees and many strange people,” which Zhao sees as energy rather than a defect.
Almost none of the core employees left during the industry’s hiring wars. Zhao gives no single explanation, but sees freedom, low constraints, and intrinsic motivation as important factors.
Sleeping in the lab or working from home is not the measure of overwork. The key is whether people are happy and whether responsibility or unresolved problems make them keep working voluntarily.

40. The First Company Tried to Build a Standard Robot but Could Not Even Standardize Emergency Stops

Around 2015, ABB, Fanuc, Yaskawa, and Kuka represented the successful paradigm of selling complete robot bodies. Zhao’s team naturally concluded that the robotic arm was the “hand” and that they should build the mobile “feet,” offering chassis series by size and payload.
Customers could not even agree on the correct basic emergency-stop behavior: cut power immediately, mechanically lock the brakes, merely disable the system so it could be pushed, or slow down first and then engage the brakes. Each answer made sense under a different site-safety logic.
Components could not be standardized either. A state-owned enterprise might require fully domestic supply, while another customer would accept only overseas brands certified by its approved supplier system. Every requirement was reasonable, but corner cases made a standard body impossible to define.
Looking back, Zhao says the problem was not that customers were too diverse. The team was wrong to assume the robot body could be a standardized product.

41. Foxconn Asking for All the Drawings Prompted SEER Robotics to Sell Only the Control System

When Foxconn launched its “million robots replacing people” initiative, it believed it could manufacture every part of the mobile robot except the controller. It wanted to buy the system but demanded the drawings for the complete machine.
A conventional robot-body company would treat those drawings as core know-how. Zhao’s team did not have strong mechanical capabilities or any corresponding pride in keeping the work, and ultimately decided to hand over everything.
Once the body was opened up, previously incompatible requirements could be handled at the software layer. Emergency stops could support 4 or more configurations, all radars could be adapted, and interfaces could be left open for customer choice.
Zhao says the moment was “very happy”: the team no longer had to reject one reasonable request after another or build a new robot for every scenario.

42. The Controller Platform Was Not a Day-One Blueprint but a Strategic Summary Written After the Fact

When the team shifted toward controllers, it had not first written a “platform strategy.” It simply felt that the path was more correct and more consistent with its DNA. Only as customers and configurable modules accumulated did the platform characteristics emerge.
Zhao believes roughly 80% of company strategy is “a summary written after the fact,” with the remaining 20% being deductions built on existing facts.
The second venture therefore completed the loop: if the system can help anyone build a robot, SEER Robotics does not need to prove that others cannot do it and then manufacture the robot itself.
The company moved from defining products to supporting others: “Help people build the robot they have in mind.” Stepping back opened a much larger boundary.

43. The First Company’s Biggest Asset Was Knowing the Cash-Flow Limit and When to End

The first company spent more than a year in a state where paying salaries left it with almost no money and little certainty about the following month’s payroll. Zhao says perhaps only he and the head of finance knew; even the other co-founders did not.
He worried that public panic would create a vicious cycle: skipping trade shows would reduce leads, which would hurt revenue further. He therefore tried to keep the team from changing its behavior because of cash flow and treated cash flow as simply another problem to solve.
Fundraising, loans, and personal borrowing were tools, not emotional burdens. Zhao says he still slept well: “Do your best and accept every outcome.” Emotion did not enter the problem model.
The 5 founders later became 3. Unable to align among themselves or reach agreement with new investors, they chose to shut down the old company, explain the decision to all sides, and restart rather than continue consuming time in limbo.

44. SEER Robotics’ Controller Is Both Edge Hardware and the Vehicle for Moving from a Rule-Based Brain to a Data-Based Brain

Zhao calls the controller the robot’s brain. Three to 5 years ago, a rule-driven planning-and-control system was the best brain available and does not lose that status merely because models exist today.
The old brain described the environment and planned trajectories through thousands of rules, while also embedding small models for functions such as recognition. The new brain makes more actions and decisions data-driven, but both must run on physical hardware.
The controller connects locally to sensors and actuators; all real-time data cannot simply be sent to the cloud. It is an edge brain that can also work with cloud models.
SEER Robotics initially focused on industrial mobile robots such as AGVs and AMRs, then expanded into mobile manipulators, wheeled humanoids, intelligent forklifts, and delivery robots.

45. SEER Robotics Used “Encircling the Cities from the Countryside”—Dirty Work Instead of a Single Technical Ambush

The previous No. 1, American company Kollmorgen, had acquired a Swedish company. When SEER Robotics entered, it did not find one grand trend the incumbent had completely missed; it found countless small unmet needs.
The route began with customers, brands, functions, and machine types that the giants considered unimportant. Where others supported 1 configuration, SEER Robotics supported 4; where others made customers navigate 3 configuration pages, SEER Robotics reduced it to 1 and automated the process.
Supporting hundreds of radars sounds unglamorous, and engineers may not immediately see the value from any single small customer. That is precisely the kind of dirty work established giants viewed as beneath them.
As the dirty work was steadily consolidated into one product system, peripheral capabilities gradually encircled the main battlefield. Zhao sees victory as “the accumulation of countless small points,” not one essential demand nobody else had imagined.

46. SEER Robotics Rejects Vertical Integration and Lets People Who Understand the Scenario Define the Robot

Leading companies such as Geekplus and Hikvision tend toward full-stack development, vertically integrating from upstream components through downstream delivery for integrated optimization and thicker margins.
SEER Robotics made the opposite choice: “If we don’t need to build it ourselves, we shouldn’t.” It does not manufacture components or robot bodies, and tries to leave delivery to partners, limiting itself to the system, toolchain, and enablement layers.
The reason is not that it cannot do everything, but that “there will always be a company that understands the user better than we do.” The customer or vertical service provider knows the product, workflow, and site constraints best.
A livestock customer used a SEER Robotics controller to build an automatic cattle-and-sheep feeding robot, a form the team had never imagined. What made Zhao happiest was not the order but that “he built something we could not have imagined.”

47. The Xinyun Platform Turns Controller Customers into Creators and Overseas Demand into Orders

A factory may simultaneously need robots for loading and unloading, cleaning, cross-floor transport, sorting, and more, yet no single company can supply them all. That supply-demand gap gave rise to SEER Robotics’ Xinyun platform.
Customers that have built products on SEER Robotics controllers can receive R&D enablement, sales leads, and orders through the platform. 卫诗婕 summarizes the model as “not just serving creators, but building a creator economy.”
Overseas customers see Chinese robots as cost-effective but do not know whom to buy from, and cannot learn multiple APIs, manuals, and maintenance systems at once. The platform packages and connects different Chinese robots through unified software.
Domestic partners gain a faster route overseas, while foreign customers need to learn only one tool. Controller sales, robot categories, and platform transactions reinforce one another in a flywheel.

48. A Leading Company Can Build Its Own Controller but Would Struggle to Rebuild Everyone’s Controller

Zhao considers it entirely rational for body companies such as Unitree to develop their own controllers, just as automakers do not want to “lose their soul.” Controller hardware alone is not the highest barrier if the company does not go deep into chips.
The real distinction is between a controller that works well for yourself and one that works well for everyone. Roughly 80% of SEER Robotics’ customers were not originally considered robotics companies by capital markets.
A later entrant seeking to become a platform must not only recreate the functions but also explain why customers should volunteer to be guinea pigs again. Without a higher-dimensional technical advantage, migration cost and time become barriers.
More than 400 sensors, different surfaces, configurations, mechanisms, and network failures create countless combinations. Only real exposure teaches you how to optimize them; a later entrant may “not even know how to copy it.”

49. SEER Robotics Builds the Union of Needs, Not the Common Denominator; It Is More Like Word Than a Standard Hardware Product

The product principle was fixed on day one: after collecting 100 scenarios, do not build only the intersection everyone shares. Implement the outer requirements unique to scenarios A, B, and C as well.
A single customer may use only 20% of the software, and each customer’s 20% is different. The complete toolchain is therefore like Word: most people understand only part of it, but believe they can almost always find a function for their text-editing need.
Standardized hardware is easy to copy and price down in China. A support tool containing 50,000 requirements, thick manuals, and field workflows is much harder for a product manager to redefine from a prototype.
That is why the controller looks like a box but is not a single product. Its moat lies in software, tools, and organizational workflows accumulated over time.

50. SEER Robotics Fears a High-Dimensional Model That Automatically Understands Everything More Than a Rival Controller

Zhao imagines a future in which a sensor is plugged into a controller and an edge “little lobster” system understands the protocol, builds the algorithm, and learns how to use it on its own. The value of manually adapting 400 brands would then collapse instantly.
If loading a model into a control unit lets a robot understand requirements, capability boundaries, and surrounding components, development engineers may no longer be needed and today’s platform ecosystem could be bypassed entirely.
He sees this as the largest risk and the direction SEER Robotics must approach deliberately: “If it reaches that point, perhaps what we do will have no value at all.”
The path ultimately approaches AGI and may take 5 or 10 years. SEER Robotics cannot protect itself by denying it; it can only expand the current toolchain while pushing itself toward models and agents.

51. Multimodality and Agents Will First Rebuild Deployment Rather Than Immediately Eliminate the Entire Control Layer

SEER Robotics has historically been stronger in traditional sensing such as lidar and magnetic navigation. Zhao sees substantial room for vision, environmental semantics, and sensor fusion, with sound and other modalities also eventually entering robotics.
Field workers should not have to configure robots by dragging items on a screen. If robots can understand natural language, explain faults, and troubleshoot automatically, the deployment barrier will fall sharply.
He estimates that agents could replace roughly 90% of field software-delivery work, perhaps within 6 to 12 months, because deployment is fundamentally about translating customer language into technical language.
In the future, an end customer could give the requirement directly to an agent, which would call the tools and solve it. Only genuinely new problems would flow back to SEER Robotics’ R&D organization.

52. Tens of Thousands of Functions Are Being Broken into Small Modules Agents Can Safely Call

SEER Robotics is gradually decoupling its existing functions into microservice-like modules, reducing coupling, stabilizing interfaces, and opening them for agents to combine.
The goal is not to make agents infallible, but to ensure each module is independently safe and that calls are observable. “What can be observed can be controlled”; errors cannot pass directly through the system boundary.
Existing interfaces were mainly designed for people. The company must expose lower-level capabilities as well, or it will constrain the space of combinations an agent can produce.
Zhao describes the goal as an experimental field with a firewall: agents can flexibly dispatch robotic capabilities, but every action is logged and constrained.

53. The Hardest Part of a Platform Is Not Matching Transactions but Quality, After-Sales Service, and Trustworthy Recommendations

Once there are thousands of partners and highly customized products, the largest challenges are quality management and long-term after-sales service. Component suppliers may discontinue products or go bankrupt, while complete machines lack a unified maintenance standard.
SEER Robotics’ first step is not to list everything. It selects solutions that have already been validated. If there are 10 similar components, it tries to select the best value, usually offering customers 2 or 3 verified options.
Customers can still name a brand outside the list, even simply because of a relationship with the boss. SEER Robotics will adapt and test it and explain the capability boundaries, while leaving the choice to the customer.
The team once publicly listed where specific brands were weak, then decided it should not play industry god. It now provides a whitelist rather than a blacklist and presents test reports objectively.

54. The Speed of China’s Component Supply Chain Means Self-Developed Parts Need a “Must-Build” Rationale

A lidar unit cost RMB10,000 to RMB30,000 a decade ago. Today, products costing a few hundred yuan can approach that performance, while domestic supply has rapidly expanded from single-line and multi-line to explosion-proof and safety-grade products.
Zhao expects Chinese components to take more global share. One component may have 20 competing suppliers in China and only 1 or 2 overseas, making it difficult to match the pace of performance and price iteration.
A humanoid company should of course develop a component itself if doing so directly accelerates deployment or produces an irreplaceable technical solution. If it is only a short-term cost-saving move or an attempt to capture a problem others have not noticed, the long-term risk is high.
Some current supply-chain price increases are merely temporary supply-demand imbalances. Zhao expects prices could fall back within 6 to 12 months; if components also enter larger markets such as autos and industry, scale will drive them down faster.

55. SEER Robotics Chooses Markets It Likes and Understands, and Uses One-to-100 to Fund the Next Zero-to-One

Zhao first looks at what the team likes and does well, then at the market. If market size mechanically determines the business, the company will have “no human character and no personality.”
SEER Robotics is best matched with customers that understand their scenarios and can define requirements but do not want to make the initial investment in a full robotics R&D system: “Treat me as your R&D department, and treat yourself as the product manager.”
He does not oppose scale, but rejects doing one-to-100 merely for money. The real purpose is to build data, installations, customers, and capabilities that create momentum for the next technological self-disruption.
After the controller’s zero-to-one came scale, which created the platform’s next zero-to-one. As the platform scales, it accumulates data for the next zero-to-one in embodied models. Each old table is being converted into chips for the next one.

56. COVID-Era Localization Created a Double “Now I Get It,” but Did Not Change SEER Robotics’ Long-Term Choice

During COVID, overseas components and chips could not enter China, so customers had to try domestic controllers. What began as a trial became a shift in procurement habits once they found the performance acceptable.
SEER Robotics was also forced to replace some overseas components and found that domestic alternatives cost less while performing adequately. Customer acceptance rose as the company’s own costs fell, sharply improving value for money.
Main compute chips remain somewhat weaker, but passive and power components are already highly competitive. Zhao says SEER Robotics does not manufacture components precisely because domestic supply-chain competitors are “too strong.”
After listing, the company has more money and is increasing R&D investment. Zhao still worries that excessive focus on short-term financial metrics will distort behavior; financial growth must serve the next round of innovation rather than become the endpoint.

57. The Humanoid Boom Brings Installations and Funding, but Valuable Data Still Has to Be Exchanged for Value

Mobile-robot, robotic-arm, model, and component companies are all moving into humanoids because the market broadly believes the category could be enormous. Some genuinely believe it, some are following the capital narrative, and some need to enter the field to understand it.
SEER Robotics currently has tens of thousands of installed units, doubling annually. Zhao notes that humanoid robots are also shipping in the tens of thousands, but it is unclear whether the resulting data can be used effectively, especially in industrial settings.
Sensitive customers do not want to send back all their data. SEER Robotics is therefore considering exchange mechanisms: lease robots to customers while retaining ownership and monitoring the assets online; customers receive a shorter payback period.
Another exchange would use cloud models to reduce deployment cost, with customers allowing connectivity and data return in exchange for more intelligence and easier deployment. “If you want data, what are you giving me in exchange?” The model must first create equivalent value.

58. Autonomous Driving Proved the Importance of a Real-Machine Loop; Embodied Intelligence Adds Manipulation

Zhao calls autonomous driving a lower-level form of embodiment: it mainly solves navigation while adapting to the environment. Robots must also manipulate objects and change the environment, adding execution, force control, and touch.
Autonomous-driving entrepreneurs moving into embodied intelligence is natural, and they should at least avoid some pitfalls. People with road experience usually insist on real data and human involvement early rather than relying entirely on simulation from day one.
Simulation can be used to train actions, and the tools are relatively mature. But vision involving lighting, textures, and real environments still has sim-to-real gaps. World models can change textures and augment samples, but the original samples should preferably come from the real world.
To advocates of synthesis who say not to use a single frame of real-machine data, Zhao gives no theoretical verdict: “A black cat or a white cat—it’s a good cat if it catches mice.” The test remains whether the specific scenario can deploy.

59. Customer Sites Reveal the Industry’s Truth: Reliability and Generalization Are Still Hard to Combine

Zhao looks at the companies that have raised the most money and command the highest valuations by comparing what they say with what customers say they have actually delivered. As a service-layer company, SEER Robotics can more easily obtain unvarnished field feedback.
His conclusion is that most performance falls short of expectations. Claims of thousand-unit orders and sustained mass production contain “a lot of exaggeration”; the industry is still taking its first steps. Tesla’s plans have repeatedly been pushed back, and complexity cannot be solved by money alone.
A 90% success rate is nearly meaningless in industry. Even 99% means 100 robots could generate a failure every day, a burden that people and processes cannot sustain.
What can currently be deployed is a choice between 2 options: use reinforcement learning to make one task highly reliable but not generalizable, or tolerate failure in low-risk scenarios such as folding clothes in exchange for generalization. Zhao says he “has not yet seen” a case that combines reliability, safety, precision, and zero-shot capability.

60. Capital Bubbles Create Deceleration, Breakups, and Repeat Founding—but Also Eliminate Answers for the Industry

卫诗婕 invokes the ofo experience as a warning that a rapid inflow of hot money often causes organizational bloat and distorted behavior. Embodied-intelligence companies have recently seen frequent executive departures, team breakups, and new company formations.
Zhao has also observed that, especially in the first half of this year, new companies appeared faster than anyone could imagine, with large numbers of team members leaving to start businesses.
The cause is not only entrepreneurs chasing money. Investors also actively assemble new teams: if an existing project is too expensive, they persuade core members to leave and invest in a new company at a lower valuation, creating a two-sided rush.
He still sees bubbles as potentially positive. Investors’ money is better spent exploring the future than sitting in a bank; failure at least eliminates one wrong answer, and robotics still has many answers to eliminate.

61. SEER Robotics Chooses Two Diagonals: New Technology for Old Forms, and Infrastructure for New Forms

In the 2x2 matrix of old versus new technology and old versus new product forms, SEER Robotics’ first path is to apply new models to mature forms such as forklifts and cleaning robots, using existing customers, scenarios, and data to drive deployment.
The second is to provide control, joints, mobile chassis, and toolchains for new product forms, allowing model-focused companies to avoid rebuilding real-time control and hardware adaptation.
SEER Robotics is not abandoning the most dangerous and attractive upper-right corner of “new technology × new form.” It is approaching it through the 2 existing capability paths. Zhao admits that if he were starting from scratch without historical accumulation, he might bet directly on that quadrant.
He defines SEER Robotics’ current role as a technology enabler and “deployment leader,” not a company claiming to possess the final paradigm. Once it sees a deployable possibility, it should be able to follow within 3 months.

62. Heterogeneous-Robot Coordination Is the Missing Systems Layer Beyond Single-Machine Intelligence

Customers care about 2 things: how much ROI one robot can create, and whether multiple robots can form a coordinated system. Individual strength does not guarantee collective intelligence.
Factories typically deploy logistics, cleaning, and automated equipment before adding humanoids. A humanoid therefore faces competition for right of way, map formats, APIs, data, and inconsistent model interfaces from day one.
SEER Robotics’ downstream customers often buy from one another. If A does not make a machine type, B supplies it, and both serve the same factory. Cross-brand, heterogeneous coordination is therefore a real requirement.
Car brands do not necessarily need to coordinate directly, but robots must work together toward one production objective. Zhao believes embodied-intelligence companies broadly underestimate this layer.

63. Big-Tech Entry Depends on the Market’s Waterline; SEER Robotics Can Fight Only Through Customer Value

Giants such as Nvidia and Huawei could theoretically move down into control and coordination. Zhao says a startup cannot rule that out; when a giant enters is primarily a calculation of investment and return.
If market returns remain below the giant’s waterline, it will not act. Once the market is large enough and returns clearly exceed investment, entry becomes nearly unavoidable.
If someone makes “intelligent machines without barriers” work better than SEER Robotics, Zhao says the company should “die a worthy death,” with “the best outcome being to die as quickly as possible.”
That does not mean giving up voluntarily: “It definitely does not mean dying without a struggle, but dying in the struggle.” The final test remains who can satisfy demand better and lower the customer’s barriers.

64. “Not Being No. 1 Is a Disgrace” Means Staying Incomplete Forever, Not Occupying the Summit

Zhao’s social-media signature reads, “Not being No. 1 is a disgrace,” yet he also says he does not like being No. 1. It took him a long time to explain the apparent contradiction.
First place, in his mind, is a target that can be approached but never achieved. Once the controller reaches No. 1, the target shifts to data volume, deployments, and the next technology, preventing the company from stopping at a completed state.
If an organization believes it is already No. 1, it will start rejecting the 101st unusual requirement, assuming that the first 100 customers have proven the rules correct. That is the point at which fullness begins to turn into decline.
The Tao Te Ching says emptiness is safer than fullness. The value of pursuing first place is not the trophy but continuously lowering one’s position and keeping the remaining gaps visible.

65. Long-Term Strategy Bets on Certainty; Short-Term Fads Require Only Speed of Following

Zhao refuses to bet on one specific milestone 6 months out. If a question can be definitively resolved in 6 months, it may not have a high barrier and may be an arithmetic problem rather than strategy.
He defines strategy as a “long-term, irreversible investment.” The longer-term direction is actually more certain, while short-term technology cycles are harder to predict precisely.
SEER Robotics firmly believes 2 things: real-machine data will exhibit a scaling law, and future edge controllers will run large models, potentially without the cloud. Zhao does not bet whether that happens in 3 months, 1 year, or 10 years.
The organization needs the ability to follow quickly once a paradigm demonstrates deployment value. That ability comes from the RoboCup era, when the team studied stronger rivals frame by frame before eventually overturning CMU.

66. Top-Down Companies Are Valued on Dreams, While Deployed Companies Face Immediate PS, PE, and ROI Scrutiny

Zhao follows overseas paths such as Figure and PI and mentions early VLA experiments based on π0.5. He also tracks every domestic company that raises huge sums for a simple reason: “We’ve never seen a lot of money, so we’re curious what they plan to do with it.”
Brain companies valued at RMB10B or more, such as Xinghaitu, represent a different pricing logic. The more deployed a company is, the more calculable its market size, PS, PE, and ROI, and the stricter the investor scrutiny; the more dream-driven it is, the more it is valued on long-term possibility.
SEER Robotics cannot become a completely non-deployment-oriented dream company overnight, and does not envy that pricing model. It will, however, study how well-funded companies allocate resources and look 6 or 12 months further ahead.
Zhao’s warning is that if a company raises money without accumulating data, scenarios, installations, or model-engineering capability, it may ultimately be “left with nothing but money.”

67. The Robotics Moat Is Not One Clever Idea but Turning Dirty, Difficult Work into Organizational Capability

Zhao does not believe a single clever method can command lasting fees in China’s To B market. Once disclosed, the method will quickly be followed by competitors and the supply chain.
More durable barriers come from sustained work: data pipelines, R&D workflows, field logs, adaptation systems, deployment tools, and the organizational capability to consolidate that dirty work into reusable products.
The hardest part for field engineers is not physical labor but being pressed by customers while unable to solve the issue and forced to wait for R&D. SEER Robotics therefore uses low-code, modularity, and broader permissions to increase frontline autonomy.
Zhao reviews field-issue conclusions every day “like scrolling short videos.” Every deployment is treated as a RoboCup: log the failure, analyze it, and fold the result back into the product and SOPs.

68. SEER Robotics Measures Talent Not by Credential Density but by an Organization’s “Vitality Index”

Around 20 people at the company have served as CTOs elsewhere, yet Zhao still does not believe anyone can master all of robotics across mechanics, electronics, controls, software, algorithms, and models.
The ideal employee does not fill every weakness. They understand multiple fields and go deep in one direction; AI and team collaboration can cover weaknesses, while strengths should be pushed further.
He does not measure talent density by 985 schools, C9 universities, or QS rankings. He looks at whether employees generate new ideas every day, dare to experiment, and whether someone can close the loop on the experiment.
If vitality depends on 1 or 2 brilliant minds while everyone else executes uniformly, even an organization of excellent people may have “everyone step into the same river and drown together.”

69. The More AI Produces Correct Answers, the More an Organization Must Preserve the 1% of Individuality

After GPT, SEER Robotics arrived at a counterintuitive conclusion: AI will provide answers that are unlikely to be wrong for 99% of questions, but if everyone accepts similar answers, choices will also converge.
Zhao recalls—without complete confidence—a story involving a car and an engine. A conventional model would probably advise the person to communicate, prepare a plan, and persuade the board, rather than support leaving the existing framework and starting again.
Industry-changing moves often come from the 1% who refuse convention. A model can support an aggressive decision when explicitly prompted, but the person first has to step outside the default frame.
That is why “the purer a person is, the more likely they are to achieve something in a professional field.” It is not a guarantee of worldly success, but individuality may help someone pass through a short-term phase that is unpopular, difficult, and correct.

70. The Stability of the Three Founders Comes from Admitting Weaknesses Rather Than Maintaining Perfection

All 3 founders came from the RoboCup team and slept in the lab together for years. Zhao became captain and the No. 1 person not through formal appointment, but because “whoever carries more of the load becomes captain.”
One co-founder acts like a “Red Guard,” directly pointing out problems inside the company. The organization assumes that “having no problems is the biggest problem,” and even the R&D teams Zhao manages are criticized.
Another co-founder is a utility player who has worked across 7 or 8 roles. When a new business is unclear, he first mobilizes resources and uses his relationships to figure out roughly how it works, then hands it to the specialists.
Zhao believes “a lot of trust is trust based on weakness.” Admitting that he is not good at external storytelling makes real mutual trust easier than hiding the flaw and maintaining a perfect image.

71. Zhao Is Not Eager to Tell Stories Externally but Sees Entrepreneurship as a Life Free from Constraints

He admits that a CEO should handle external communication, but prefers the front line of product and R&D, organizational evolution, and back-office support. Another co-founder runs the “Teacher Ye Plays with Robots” video account, with roughly 60,000 to 70,000 followers, naturally filling that capability gap.
Zhao describes his role as “front-line work plus back-office support”: move wherever needed while helping others release their capabilities.
Entrepreneurship satisfies the part of him that dislikes constraints on thought and behavior. But the absence of constraints can itself become the greatest constraint because he must be responsible to employees, customers, and investors.
His definition of happiness is not grand: “At least at this stage, I can still happily do what I want to do.” That is enough.

72. The Tao Te Ching Helped Zhao See That SEER Robotics Has Always Practiced “Non-Forcing”

Some people say Zhao has “very little karmic burden,” meaning that few things become lasting entanglements and anxiety struggles to hold him for long. He began reading the Tao Te Ching and realized that many of his own and the company’s practices already shared its underlying tone.
He interprets wu wei as “not acting recklessly”: act in accordance with the rules rather than doing nothing. Building controllers and supporting others from the rear is also an active choice to occupy a lower position.
“When the work is done and the task accomplished, the people all say, ‘We did it naturally,’” comes closest to his ideal management state: after the job is complete, employees feel it happened naturally through everyone’s efforts rather than through a campaign launched by the boss.
Embracing differences, not forcing standards, and remaining empty rather than full all echo SEER Robotics’ organizational and product logic of openness and diversity.

73. Anime Provided an Escape from the Logical World and Reinforced the Idea That First Place Must Be Emptied

When Zhao withdrew from medical school and prepared for the graduate exam, he tuned robots during the day and studied at night. A co-founder recommended anime as a way to relax. The habit stuck: occasionally he would watch an entire long series at 2x or 3x speed in a day, or use cheats to finish a game in one day.
Fullmetal Alchemist, Code Geass, Gundam, and Attack on Titan were “masterpiece anime” that suspended everyday logic through acceleration, scale, and utopian world-building.
Code Geass left a particularly strong imprint. The protagonist first makes himself the common enemy of the world, then has his friend kill him at the summit, sacrificing himself to reunite the world.
卫诗婕 uses the story to reverse-engineer Zhao’s view of first place: once you reach the top, you must quickly find the next state and empty yourself, or the person on the summit will gradually become the dragon.

74. Romanticism Sets the Mission; Pragmatism Executes Every Step

Zhao admires Kazuo Inamori and Mao Zedong because both held strongly idealistic visions: the former believed that extreme effort would bring “divine inspiration,” while the latter said, “All reactionaries are paper tigers.”
In management and strategy, however, both respected objective laws, concrete circumstances, and detailed execution. Zhao admires that combination: unwavering belief in a grand goal that may not be rationally justified, paired with executable and logical actions.
SEER Robotics’ idealistic side is its conviction that the robotics world will inevitably be diverse and that every company will eventually become a robotics company. Its materialist side is doing the controllers, installations, adaptations, data, and field problems one by one.
The company’s “wrong” and “right” wall closes the loop. The middle is often chaotic and the correct path impossible to see, but by continuing to move and changing the viewing angle, a starting point that looked wrong may become part of the right path.

Why Now is the Time for the App Layer | Why Startups Should be TokenMaxxing | Mike Mignano, USV 陈天桥's Chief Scientist on AI Self-Evolution in as Little as 6 Months