77. The Vision and Ambition of AgiBot and Migo, and the Race in Embodied Intelligence | A Conversation with 姚卯青

2026/06/16 · Source link · AI Summary archive

Summary

姚卯青把具身智能的进度定在“连GPT-1都没到”——当前全行业真机数据的量级，比真正基模所需少四到五个数量级。 大部分惊艳Demo靠后训练即可做出，与基模无关；他判断真正的涌现需要1亿小时数据，“等到我们有1亿小时数据的时候，我认为这个前夜就来临了”。按路线图，10亿小时级别数据预计在2027—2028年达到；主持人判断明年达到1亿小时有希望。
AgiBot的营收曲线是“每次都是加两个零”：2023年只卖1台样机、约RMB300K，2024年RMB60M，2025年已超过RMB1B。 这分别对应量产元年、商用元年和2026年部署元年，并支撑Deng的“三五八战略”（3/5/8年对应RMB1B/RMB10B/RMB100B营收）。精灵系列被确认是现金奶牛、较快实现ROI的产品，但新品先发护城河只有“1年左右”。
Migo的剥离源于与红杉的一顿饭：对标Scale AI，但具身数据需要进入物理场景、管人、投设备，门槛和复杂度更高。 2026年规划1,000万小时产能，等效约1万人按正常工作强度工作1年；若达成，姚称其“绝对领先”。真机数据价格约RMB500—1,000/小时，100万小时约RMB1B。
数据军备竞赛的本质“不是谁拥有多少量，而是谁在多快的速度内拥有多少量”。 年初AgiBot World的100万条、3,000多小时数据已是最大规模真机数据集；到2025年底，Generalist称拥有20万小时数据，但姚判断应主要是无本体数据。Nvidia GR00T N1的技术报告显示，其使用的真机数据中80%来自AgiBot World。他转述Generalist与Google DeepMind技术人员的共识：若数据是唯一决胜变量，中国公司现在就可以宣布自己一定是赢家。
反共识判断：车不是“最简单的机器人”，把车归为具身智能“有点牵强”。 车永远安全第一，高速场景下100—200毫秒级反应和大模型的推理速度、概率性错误都构成限制；机器人低速、可人工兜底，反而有机会部署VLA、世界模型等前沿模型。但要落地成功，“没有人比做过自动驾驶的人更有机会”：量产硬件、数据闭环、供应链和质量体系都是智驾从业者的优势。
商业化认知在1年内发生变化：机器人抢场景要“沉下心、撸起袖子，甚至跪得下来”，因为“工业不相信故事，也不相信眼泪，相信的就是ROI”。 AgiBot称已在龙旗南昌工厂3C主产线稳定运行1个月后进行全天直播，完成2,300多次任务、100%成功、0失误；姚称Tesla和Figure尚未做过类似活动。
融资与竞争哲学是：Deng说“我们想要钱随时都可以融，但没必要”，AgiBot给投资人的股权可能还不到一半，强调克制稀释、把股权留给团队。 姚引用余凯的护城河观：不是绝顶聪明的算法，而是工程问题日积月累形成完整体系；“聪明可能只占10%，剩下90%是完整体系的能力”，“大家可以去做聪明人，我们来做好老实人”。
技术路线判断是：VLA和世界模型“终局来看可能都不是最终形态”，未来更可能是上下游分层配合。 语言高度压缩但有损，世界模型更能捕捉液体流动、玻璃破碎等物理规律；两三年后，机器人真实场景回流数据可能成为主要增量来源，依据是GPT-3时代的历史类比以及对scaling law普适性的判断。

Deep dive

1. Embodied AI Is Not Even at GPT-1: The Missing Fuel Is 100M Hours

The sharpest calibration in the episode came when 姚卯青 punctured investor expectations. 诗婕 relayed the message: “Investors really care about this. They hope you’re already at 2—please, please, preferably 3.” 姚’s answer: “Actually, we haven’t even reached 1. I don’t think we’ve reached 1.” He believes embodied AI is not yet at GPT-1, and that the truly decisive moment may not arrive until GPT-3.
The bottleneck is not algorithms: “I don’t think algorithms will be the absolute bottleneck or constraint, but data is definitely the biggest problem.” Embodied AI is “nowhere near emergence”; everything is still a visual-to-motor mapping exercise, without the feeling that “intelligence seems to have awakened.”
His timeline: “The eve before the eve. Once we have 100M hours of data, I think that eve will arrive.” The line was cut into the opening and serves as the episode’s central thesis. He sees 100M hours as the threshold at which the eve of emergence may begin; the roadmap puts 1B-hour-scale data in 2027–2028.

2. 姚卯青’s Self-Portrait: Refined on the Surface, Wolfish Underneath

Asked about his personality, 姚 first deconstructed the “refined” label. As an operator running a startup, refinement is one side of the job; he also needs a strong dose of wolfishness.
He defines himself as clear-eyed about reality: when planning and assessing technology and commercial direction, he “doesn’t blindly follow hot trends” and instead looks at the underlying business logic. His career path—Waymo to NIO to AgiBot—was not planned. “Greatness cannot be planned”; he simply followed the technology wave and kept making himself a participant in it.

3. Tsinghua Electronics: An Unofficial Alma Mater for Embodied Startups

The Tsinghua imprint is its motto, “Actions speak louder than words”: practical, understated, and focused on doing the work. 姚 believes the Department of Electronics is probably one of the hardest majors to enter; his gaokao score was good, but at Tsinghua he was in the second tier. The first tier consisted of students admitted through recommendation, many of them Olympiad gold medalists.
The coursework was heavy and broad—analog and digital circuits, integrated circuits, programming, communications, signal processing, machine vision, optics, and microwave engineering. It “builds a systems capability” and develops the ability to absorb pressure, epitomized by the “devilishly hard junior year.”
A Shanghai Electronics Department committee annual meeting became a roll call of embodied-AI founders: 陈一伦 of Zhihang, 王潜 of Ziliang, and 丁文博 of Tsinghua Shenzhen International Graduate School. Others absent included 王鹤 of Galaxy General, 高继扬 of Xinghai Tu, “Snowflake Beer Bro,” 江哲源 of Songyan Dynamics, and 申宇君. The alumni network is close, but “this is an intensely competitive industry, and companies will fully compete with one another.” Meet on the field.

4. USC and Silicon Valley: Academia Has to Connect to Industry

USC’s biggest impact was its practical orientation. Many professors came from industry rather than spending their entire careers teaching. He cited views including 李飞飞’s: university resources are too limited to compete with industry, while the AI era is a contest for compute and resources.
He also saw the pull of the US academic ecosystem and top scholars. USC’s best-known program is its film school; the university is financially strong and relatively well positioned in infrastructure and faculty.
Moving to Silicon Valley after graduation was the natural employment choice. His self-deprecating description is that, in today’s language, his cohort already counts as “middle-aged tech workers.” When he graduated in 2014, deep learning was just emerging, and his PhD work was still closer to semiconductor physics.

5. Google Ads: Transformer Went Into Production the Year It Was Born

He first encountered applied AI at Google’s advertising team. Display Ads used then-new visual algorithms including ResNet and GoogLeNet for recommendation and matching.
The key detail: Transformer was created at Google in 2017, and the paper was titled Attention Is All You Need. That same year, the team used a Transformer-based framework—still described at the time as machine translation—to generate advertising headlines. Large-scale campaigns could not rely on people writing copy by hand every day.
His motivation for moving on was straightforward: using technology only for advertising “didn’t seem hardcore enough.” He looked internally for teams applying frontier AI while doing fundamental technical work. In 2017–2018, autonomous driving was the hottest AI segment, with TaaS—Transportation as a Service—offering enormous room for imagination.

6. Three Years at Waymo: From Demo to Rider-Only

Getting into Waymo was much harder than getting into Google. By then Waymo had spun out of Google X into an Alphabet sister company on equal footing with Google. Candidates went through roughly 7 or 8 interview rounds, starting with 3 rounds of the hardest on-site LeetCode problems. Even internal Google transfers faced strict screening.
Tsinghua gave him an unofficial edge. Several leaders in the Perception team he wanted to join were Tsinghua alumnae, including 2 who were 4 years ahead of him and had PhDs from Stanford. “Alumni share an extra layer of trust and an extra layer of familiarity.”
Over more than 3 years, he watched Waymo move from technology demo to rider-only commercialization: no safety driver in front, only passengers inside. At the time, service was limited to Chandler, a district of Phoenix, Arizona—roughly like operating only in Jiading or Qingpu in Shanghai.

7. Returning to China and Moving to L2: The Operational Burden of L4

Returning home was a decision shaped by both family and the times. He and his partner are both only children from Jiangsu, Zhejiang, and Shanghai. In late 2021 and early 2022, China’s new-energy industry was at its peak, and autonomous driving was a hard technical differentiator between new-energy vehicles, ICE vehicles, and traditional joint-venture brands.
He uses HD maps to illustrate why L4 moves slowly. Every traffic light requires precise annotation of the number of bulbs, colors, orientation, shape, turn arrows, and 3D position. US infrastructure changes relatively slowly; in China, roads and lights can change at any time, requiring teams to rescan maps in real time. Heavy-asset fleets and mapping operations make geographic expansion slow.
L2, by contrast, can expand to customers once the vehicle is sold. That faster route to commercialization was the core logic behind his move from pure L4 Robotaxi work into passenger-vehicle L2/L3 development.

8. Why NIO: One Platform Across the Lineup

What attracted him was the systematic product definition. NIO emphasized using one NT 2.0 platform, with lidar, 4K cameras, and 4 Orin chips standard across the lineup. “It’s like buying an iPhone”: different sizes do not mean the operating system and apps are adapted or stripped down. For technical teams, supporting multiple platforms makes it difficult to optimize any one of them.
He still remembers 李斌’s decision on the roof-mounted lidar “forehead”: “You’ve designed so many cars in the past—has anyone really remembered your designs? If we make it like this today, maybe it will be remembered for generations, maybe it will be remembered as a disaster, but people will definitely remember you.” Roof-mounted lidar is now becoming standard across the industry; some have even glued fake lidars onto lower-end cars. “Be the first. Be the leader.”

9. What NIO Shipped: BEV, General Obstacle Detection, and AEB

NIO was among the first companies globally—and, in 姚’s telling, the first—to put lidar into mass-produced vehicles, using lidar from RoboSense. This continued his Waymo-era lidar-led approach: L4’s safety redundancy requirements could not be met by pure vision.
The production stack included lidar perception, camera-lidar fusion, BEV algorithms that became popular after 2021–2022, and general obstacle detection based on lidar and Occupancy Network. Neural-network-based end-to-end AEB was also one of the earliest functions they brought to mass-market vehicles across the industry in 2023, materially improving the AEB experience.
Mass production changed his understanding of the data flywheel’s scale. Waymo’s fleet of several hundred vehicles was the largest fleet of its time, but NT2.0 and NT3.0 vehicles now number more than 1M. The corner cases those vehicles encounter every day were “unimaginable for the original Robotaxi fleet.” Engineering can push triggers to the entire installed base, automatically slice specific scenes, and send them back.

10. The Essence of Mass Production: Delivery Discipline and Branch Management

Robotaxi companies are more R&D-oriented—“in plain English, there is no clear deadline.” Mass production is delivery-oriented: requirements are collected, reviewed and locked, the solution is locked, the system is developed, and then it goes through simulation, proving-ground, public-road, and shadow-mode testing. The cadence must also match vehicle software releases, with at most a 1-week extension for a safety bug or a hotfix at launch.
Parallel development across multiple versions requires a complex branch-management system. Code has a mainline and branches; when several versions run in parallel, teams must manage common features, bug fixes, and merges.
Google’s trunk model—one version and one codebase across a company of more than 100K people—is not easily replicated. It depends on powerful infrastructure and test gates, supported by tens of thousands of infrastructure engineers. Any team moving quickly in the real world will face multiple branches and versions.
The same lesson carries into robotics. A robot may have a longer lifecycle than a car, and OTA can make it “get smarter the longer it is used.” Lifecycle management across the full product, with the data flywheel feeding upgrades, remains “quite rare.”

11. The Autonomous-Driving Camp’s Edge: Procurement Is Not Online Shopping

Robots are less like consumer 3C electronics and more like cars and other power-intensive systems. They combine sensors, actuation, energy, intelligence, and control; the underlying hardware and software engineering systems overlap with automotive, as do supply chains and manufacturing. Many robot components come from the new-energy vehicle supply chain.
The US peer mentioned by the host was described with uncertainty—“probably Sunday.” 姚’s lesson was that one team initially thought it could assemble robots for mass production by buying cameras and wires on Taobao or Amazon. It later discovered that “procurement” is a discipline: setting specifications through an SOR, auditing factories, evaluating quality and technical capability, comparing multiple suppliers, and maintaining A, B, and, when necessary, C suppliers for continuity.
The garage robotics startups that the US media described as the first wave of failures lacked a mass-production mindset. “We were prepared to fight a hard battle from day one.” Over the past few months, the hardest work has been building the supply chain.

12. Against Consensus: Cars Are Not the Simplest Robots

姚’s exact view: “People in autonomous driving and automakers will say a car is also a form of embodied intelligence. I think that’s a bit forced.” The dividing line is the operating environment. Cars put safety first—“one mistake in ten thousand is an irreversible tragedy”—while edge-model inference is limited and deep-learning algorithms are probabilistic. BEV, world models, and other autonomous-driving foundation models therefore cannot represent the entirety of a vehicle’s running system; rules, small models, and high-speed real-time systems remain underneath.
Robots have more room for error. A failed factory task causes an economic loss and can be covered by human intervention. “A car can’t have an accident and then turn back time so someone can hit the brake.”
Speed is different too. At more than 100 km/h, human reaction limits may be 100–200 milliseconds. Robots usually operate at low speed in pick-and-place settings, where a momentary error is less likely to cause irreversible harm. That creates room to deploy VLA and world models, but they cannot simply become the complete autonomous-driving system.
On the trend toward integrating the architectures of cars and robots, as discussed in a Xiaomi interview, he remains cautious: “There isn’t that deep a fusion.” Mature technologies such as localization, navigation, and PNC can be reused, but the specific algorithms above the electrical/electronic architecture, operating system, middleware, and data links remain very different. Robots today are primarily end-to-end; autonomous driving cannot be purely end-to-end.

13. “Foundation Model” Is a False Concept: The Data Gap Is 4–5 Orders of Magnitude

He first dismantles the industry’s marketing language. Foundation models carry high perceived value and excite investors, so many embodied-AI companies label themselves “embodied foundation models.” But saying “my foundation model is good, therefore my demo is good” is a bait-and-switch. Most demos can be produced with post-training; even without a foundation model, a team can collect data for one demo from scratch and post-train it to look excellent.
When 诗婕 pressed him on the scale, he clarified that the gap was not one-quarter but “4 to 5 zeros”—10,000x to 100,000x. The analogy is that discussing foundation models when Transformer had just appeared in 2017 or BERT had only just crossed 100M parameters in 2018 would have made little sense. The concept becomes more meaningful at trillion-parameter and hundred-trillion-scale training corpora, when emergence begins to appear.
The data problem is not just volume. The industry still cannot fully specify what information an adult—or even a 4-year-old child—needs to operate autonomously in the real world. He rejects the idea that vision is enough: even losing hearing for a day would make many tasks difficult, before adding touch, temperature, and force feedback. Information modalities, volume, and scenario diversity are the core fuel for moving from motor mapping to something that looks like intelligence.

14. Academia vs. Engineering: One Seeks the End State, the Other Ships the Paper

Academia’s mission is to work on what industry has not yet imagined. Once a clear benchmark exists, academic value should not lie in using tricks to win a fraction of a point, but in original work and architectural breakthroughs from 0 to 1.
He believes current architectures cannot support robots working like humans. People may need more than 2 systems or fast-and-slow systems; robots may need multiple layers: motor-facing communication and control at up to 1,000Hz; a 10–20Hz VLA system for path and trajectory planning; a 1–2Hz world model or vision-action model for stage-level goals; and an upper layer resembling “OpenClaw,” handling long-horizon reasoning and planning in dynamic environments.
Do not underestimate the home. “A nanny may spend more effort thinking than moving.” After pressing an elevator button, deciding which elevator will arrive first is already a complex reasoning task.
Engineering turns ideas into products: performance targets, resource allocation, functional safety, and the endless practical problems and bugs of real-world deployment. Publishing a paper, building a few demos, and uploading them to GitHub is not the end of the job.

15. The Rarest Talent: People Who Have Fixed Bugs and Fought With Teammates

He cited OpenAI infrastructure engineer 温家怡: “It is relatively easy to train an engineer to do academic research, but turning a pure researcher into an engineer requires more barriers and more hard-earned experience.”
The scarce talent is not the person with the most top-conference papers or benchmark wins, but the person who has personally deployed systems, fixed many bugs, and argued with the team. Engineers at AgiBot who publish papers or top benchmarks have also gone through the full process of fixing problems with hardware, software, and other teams.
He described an assumption that may have existed in some early DeepMind and PI teams: take the best algorithm, connect it to a perfect robot, and intelligence will emerge. “That doesn’t exist.” Infrastructure, hardware, software, and algorithms have to be developed together. That is also why he sees an opening for people from autonomous driving.

16. Joining AgiBot: A Halted End-to-End Startup and a Sequoia Introduction

In 2023–2024 he tried to start an end-to-end autonomous-driving company, with backing from his former employer and Sequoia among other investors. On review, he was clear-eyed: autonomous driving had already developed for 7 or 8 years since around 2016. Investors had a reasonably clear view of which business models and technologies could work, along with the painful lessons. There was less FOMO and excitement.
At the same time, the first embodied-AI companies were just being formed, and multiple investors encouraged him to take a look. Sequoia was helping AgiBot recruit talent and contacted him directly.
What persuaded him was the chemistry between people and the willingness to commit resources. Deng was “very clear that he wanted to go all in on AI,” positioning AgiBot as an AI company rather than simply a humanoid-robot company. Humanoids were the vehicle for bringing AI into the physical world. After a short conversation, Deng decisively committed the initial resources—an unusually large investment for the embodied-AI industry in early 2024.

17. Deng’s Style: Low Profile, Disciplined Fundraising, Equity for the Team

Deng’s style runs against industry convention. Many AI founders like to stand in front of media and investors to express their views; Deng is more low-key and spends his time on internal planning and management.
AgiBot did not disclose new financing amounts for a long time because “financing is not the ultimate purpose of entrepreneurship.” It is only a milestone or tool needed to achieve the objective. 姚 says the equity given to investors may be less than half, reflecting a restrained fundraising pace and a desire to leave more ownership for the team and employees.
Asked whether this was a Huawei-style approach, he answered, “It probably is,” summarizing it as broad-based sharing of value creation.
Deng’s signature line at an ecosystem-partner conference was: “We can raise money whenever we want, but there’s no need.” 姚 added that AgiBot sometimes politely declines investors’ desire to invest, avoiding unnecessary dilution at an unnecessary stage.

18. The Other Side of the Capital Boom: Industry-Wide Inflation

As an early mover growing quickly, AgiBot has also faced the downside of the boom. Talent, supply chains, product prices, customers, and market resources have all become more expensive as large numbers of players entered the sector. Even employees choosing AgiBot may receive high offers elsewhere, forcing AgiBot to respond.
His analogy is Silicon Valley real estate and living costs, inflated by the US stock-market bubble and the concentration of highly paid workers. The hotter robotics becomes, the more intense the competition for people and materials—in effect, industry inflation.
Early deployment of company-wide equity incentives has helped counter the market offensive and retain employees.
He was measured on the expansion of the “RMB10B valuation club.” Valuations have risen unusually fast, and not every company above RMB10B is operating at the same level. But he believes all of them have strong teams that can develop within their respective niches.

19. How to Evaluate an Embodied-AI Company

First, look at hard investment: “The team, compute, algorithms, and data—none of these can be skipped.” If a company has raised a lot of money but leaves it in wealth-management products, “something is missing.”
Second, assess management maturity: Is the organization clear? How do R&D, product, and delivery work together? Are targets, performance management, and performance incentives institutionalized?
Third, examine the supply chain, manufacturing, and quality systems. Embodied AI starts with complex, high-value hardware that must be safe, stable, and consistent. Otherwise, a company can only produce a few laboratory prototypes.
He recalled that when he previously participated in building data-collection and training sites, many customers said they had bought hardware from several vendors but only AgiBot’s equipment actually worked. Some devices never completed commissioning after being powered on. The source presents this as feedback from some customers, not a universal pattern.
His closing line: “Embodied intelligence must first be embodied. Without embodiment, intelligence exists only in papers.”

20. Embodied AI Will Not Repeat Autonomous Driving’s Decade

He has revised his view. He once expected embodied AI to take at least 5 or 6 years, and possibly 8 or 9, like new energy and autonomous driving. Now he thinks it may compress those industries’ long cycles into 2 or 3 years. The drivers are foundation-model acceleration, veterans from other industries avoiding old mistakes, a larger industrial opportunity, and the inflow of capital and talent under national strategy.
Autonomous driving’s lesson is to close the loop in a focused niche first, such as logistics-delivery vehicles or L2/L3 driver assistance. There is no universal L4. Robotics likewise cannot claim that difficult long-term targets such as the home are around the corner without a defined ODD, or it will quickly be disproved.
Factories are a step-by-step starting point: “You can say entering a factory is not the hardest thing, but first you have to prove you can do A, then move from point A1, A2, and A3 gradually to B.” No technology arrives from outer space overnight to carry the industry to another stage.

21. AgiBot’s Three Businesses: Labor, Silicon-Based Stars, and Frontier Exploration

The business split is clear. 姚’s embodied business focuses on wheeled robots, including Genie G1 and G2, plus the upcoming G2 Air lightweight model and G2 Max heavy-duty model. The core mission is “to get work done”—manipulation and task intelligence. 王闯’s general-purpose business makes full-size bipedal humanoids; its newly released A3 is positioned as a “silicon-based star” for explanations, reception, entertainment, performances, and commercial endorsements. 志辉 leads the Lingxi line and X Lab; X2 and the Webster backflip fall under athletic intelligence.
Why put robots on stage? 姚’s answer: “Emotional value is also a form of productivity, and it is a large market.”
Three types of intelligence—interaction, athletic, and task intelligence—are developing separately today. The end goal is human-level intelligence, with the 3 eventually converging. Innovation is not confined to one department; every business unit and product line has teams working on next-generation technology.

22. Will the Stars Clash? BU Structure, Self-Funding, and Weighted Coefficients

On whether strong personalities can coexist, 姚 punctured the drama. People enjoy gossip and imagined plots, but if responsibilities and collaboration mechanisms are clear and everyone is too busy to keep up, there will not be the persistent interest conflicts outsiders imagine.
The 3 leaders hold many meetings each week and support one another. Guangzhou Metro’s security-screening guidance, patrol, and information robots, for example, use interaction technology and products from other departments. The management rule is to get the beginning right: management holds regular strategy sessions, employees have OKRs, and goals and paths are clarified first.
Deng has implemented a BU structure. Each BU has product, technical, and commercialization targets; budgets are allocated according to the scale and difficulty of the targets, not revenue alone. Wheeled robots entering harder industrial settings receive a weighted coefficient for commercialization results.
Fine-grained management should be implemented while things are going well, when the cost and downside are lower. This is 姚 relaying a management view, not attributing it to a specific person.

23. The AgiBot Incubation Wave: Market-Based Coopetition, Not Infighting

On rumors that internal disagreements caused the spinouts, 姚 said “not really.” AgiBot’s resources will not be the bottleneck; any clear and valuable objective can receive the resources it needs.
Subsidiaries are spun out first to serve the wider industry better, and only secondarily to become leaders in a single category. Migo will do normal business with AgiBot. If another customer raises a need first, capacity will be allocated on a market basis—first come, first served.
Will competitors use AgiBot-affiliated services? His observation is that many embodied-AI companies still recognize the value of AgiBot’s data. They join competitions, use open-source data, and proactively reach out when data services launch. One reason is decent industry relationships; another is that duplicating heavy-asset data collection is painful.
The industry is nowhere near a zero-sum game; coopetition is common. 姚 expects that within 1 or 2 years, or 2 or 3 years, leading teams will all need tens of millions or hundreds of millions of hours of data. If a provider can reliably supply high-quality data, customers will take it all.

24. From Closed to Open: The Industry Needs More Smart People

诗婕 raised the contradiction: last year 姚 said AgiBot would pursue an Apple-style closed hardware-software stack, while today it is open-sourcing datasets and building an ecosystem. His explanation: “Capital moves fast, but technology and products are not moving fast enough.” More smart people need to join, with usable robots, data, and benchmarks.
His stance on technical competition was almost a declaration: “If a company is afraid of technical competition, it might as well surrender first.” He believes he will be a core participant in the next major technical shift.
He rejected the idea that AgiBot changed course because other companies also entered the RMB10B valuation club and closed systems no longer work. Financing has no necessary relationship with true corporate success; many heavily funded companies still fail to match their technology and product iteration.
AgiBot is also investing several hundred million RMB in an industrial ecosystem fund, developing university, application-development, and delivery partners. Competitions provide baselines, platforms, and data; many participants ultimately beat the baseline, with some clearly outperforming AgiBot. Those are exactly the people AgiBot wants to know.

25. “Microwave Mode”: Mass-Producing a Robot in 2 Months

After joining, he quickly entered “microwave mode” and began to run hot. In the first half of 2024, few companies could truly deploy VLA on a robot and demonstrate it live to investors. AgiBot did it in roughly 1 or 2 months.
The first Genie went from idea in late August to design in September, supplier and contract-manufacturer selection, line construction, trial production, and volume production in “just over 2 months.” The normal industry cycle is at least 9 months. He acknowledges that luck played a role.
2 levers drove the acceleration. First, AgiBot selected a mature contract-manufacturing supply chain, allowing flexible staffing and borrowing 3C’s quality, process, workforce-management, and information-security systems. Second, a small-team founder mode put 20 or 30 hardware, software, and systems staff on site at the factory until the first robots rolled off the line.
This approach works only from 0 to 1. Genie T2 took nearly 9 months from definition to production ramp. Once volume reached more than 10K units, a complete project-management and quality system became necessary.

26. Choosing Linghou: Suppliers Believe ROI, Not Stories

The reality of supplier selection in 2024 was blunt. Investors might believe the long-term robotics story, but suppliers first need orders and survival. They will not invest ahead of time because of a promise of 100K units a year in the future. Actuator, motor, reducer, electronics, camera, radar, and battery vendors all look at ROI.
By the second half of 2025 and into 2026, the situation had changed. Automotive and new-energy supply-chain companies began reaching out proactively.
The final choice was Suzhou Linghou Robotics. 姚 said its valuation was recently reported to have reached RMB5B. The key tests were its capability system and whether it genuinely believed AgiBot would improve and was willing to collaborate aggressively. Linghou had experience serving leading 3C customers, with validated capabilities in quality, safety, information security, and dynamic capacity.
Robots have a stronger boom-and-bust production pattern than cars. At certain moments they can overwhelm a contract manufacturer; periods such as the October 1 and May 1 holidays require overtime, more workers, and additional lines. To B demand is less stable and predictable than passenger vehicles: once the right scenario and customer are found, demand can appear suddenly.

27. Current Supply Is Tight: Zero Safety Stock

The supply chain remains constrained: “The factory finishes production and the units are taken away immediately. There is no safety stock.” The ideal would be monthly sales of 10K units with 2K–3K held in inventory to absorb demand volatility, but that is not yet possible.
Upstream suppliers are under severe pressure. Quality management, training, enablement, and capacity are all stretched, and some suppliers have abandoned projects halfway through because they were exhausted.

28. Why Genie Took Off: An Industrial Benchmark and Only a 1-Year Moat

Investors rate the Genie line highly because it was the industry’s first industrial-grade embodied-AI operating platform, with high-end sensors, actuators, and compute. The downside followed quickly: competitors began building similar products.
The moat is quantified with unusual honesty: “Around 1 year.” AgiBot has nearly 2,000 full-time and co-creation employees, yet completing the full IPD process still takes 9 months. Other teams following the same rigorous standard would struggle to finish in less than 1 year.
The fundamental demand driver is labor. Hiring in China is getting harder as the demographic dividend fades, and younger workers do not want to spend their careers in factories. Research found that many industrial jobs put people into strict takt times and synchronized rhythms, asking them to operate like robots. Demographic change will intensify the labor challenge.

29. Cash Cow and Pricing War: Enterprise Pricing and Tiered Channels

The host called Genie a cash cow and the product most likely to reach ROI fastest. 姚 replied, “You can put it that way.” The pricing logic is clear because Genie primarily serves enterprise customers: compare it with the wage cost of an equivalent worker in the factory, and if the robot covers that cost over its full lifecycle, the company can set a corresponding price.
Genie has largely avoided the price war, although other niches have been affected locally. It would be unrealistic to expect complete immunity.
AgiBot uses a tiered system covering a terminal-customer list price, an MSRP, and separate pricing for elite and VIP channels. Without a pricing architecture, the market can quickly fall into disorder.

30. The 3-5-8 Strategy: How the Ambition Adds Two Zeros Each Time

The chain of numbers is straightforward: 1 prototype sold in 2023 for perhaps RMB300K; RMB60M in revenue in 2024; more than RMB1B in 2025. The first year of mass production corresponds to several hundred units and tens of millions of RMB; the first commercial year expands into the RMB1B range; 2026 is defined as the first year of deployment. 姚 summarizes it as “adding 2 zeros every time.”
He admits that a startup targeting RMB1B in revenue and 100K units shipped in 3 years is “bold,” but says the target comes from reviewing consecutive growth. The key variable is opening a real application market, entering scenarios, and closing the value loop for customers.
The path has to stay focused. With limited resources, the company must maximize resource ROI and break through scenarios that match current technology, are most willing to co-create, and have the greatest potential for scaled replication.
Is AgiBot the most ambitious company? 姚 says it “can be called that”: “If you’re going to do it, do it best. Be a pioneer, not a follower.” He sees robots as potentially the next mass industrial terminal after smartphones and cars—and as an entry point for physical AI and productivity.

31. In the Age of Endless Demos: How to Tell What Is Real

姚’s warning to himself is that robots are hardware and can produce demos every day. Practitioners must determine whether a demo solves a fundamental technical problem or merely creates a temporary visual effect.
His caution about narratives is direct: “You can’t believe as many stories as an investor.” Narrators often tell audiences what they want to hear—for example, claiming to have discovered a scaling law or achieved emergence, then pairing the claim with a demo that creates a misleading impression.
The practical test is to start from the end. If the technology is genuinely good enough to solve the problem, someone will use it and may even pay for it. Checking whether customers have placed orders and how far performance is from practical deployment reveals much of a demo’s real value.
But moving from demo to mass production takes at least 9–12 months. Final validation may not come for a year.

32. The Biggest Change in 1 Year: You Have to Get Down on Your Knees

The biggest update in his thinking over the past year is that a robot competing for real-world scenarios must “settle down, roll up its sleeves, and even get down on its knees.” Industry does not believe stories or tears; it believes ROI. The analogy comes from how automotive suppliers respond to OEM customers, cut costs, and operate under strict payment terms.
The hardest organizational leap is moving from product-centric to customer-centric. Product-centric means taking money to build a demo and then raising more money. A real commercial loop requires solving the customer’s problem and delivering value.
Customers do not particularly care whether a robot is humanoid, monkey-shaped, or dog-shaped. They care whether service is stable, economical, reliable, and fast. Some robots used to “walk into a factory, film a video, and walk back out.” AgiBot has to make them “stay in the factory and not come back out,” or customers will return them.

33. The Longqi Nanchang 24-Hour Livestream: Three Layers of Credibility

姚 said AgiBot and Xinhua News Agency livestreamed humanoid robots working an entire day last Tuesday on the main production line at Longqi’s Nanchang factory in Jiangxi. The first credibility test was the “main line”: this was not an experimental area, there could be no bottleneck upstream or downstream, and 3C production lines run at a high takt.
The second was “all day”: this was not a few minutes of livestreaming. The robots had already operated stably for 1 month before AgiBot dared to broadcast a full day.
The third was global broadcast. Inviting Xinhua to livestream the event reflected confidence in stability. 姚 said the robots completed more than 2,300 tasks with 100% success and 0 errors.
Compared with last year’s 8-hour demonstration of relatively decoupled material-box loading and unloading, this event put the robots on the highest-takt 3C line, where they had to coordinate with workers and perform precision operations. 姚 said Tesla and Figure had not yet held a comparable event, without claiming they were incapable of doing so.

34. The Scenario Race: Head-to-Head on Boxes, Co-Creation Everywhere Else

Material-box handling is a broad industrial use case that most teams can address. Competition therefore comes down to product, price, and service: the winner is whoever is good, cheap, and well supported.
姚 said AgiBot was the only team at some customer sites to meet the technical requirements, hit the task takt, and achieve the required success rate. When the material boxes changed, it adapted and went back online within 2 hours while maintaining 100% success, ultimately passing formal acceptance.
Entering a real production line cannot be done by guesswork. Teams need to spend several days on site observing workflows and labor pain points, integrate with the factory MES, and complete communications, joint debugging, and validation. Robots still do not have plug-and-play generality.
Early co-creation customers are often companies that can see future value during an industrial transition and are willing to adopt technology ahead of the curve. 姚 believes overseas technology giants are also actively exploring the introduction of embodied AI into supply chains and manufacturing.

35. AgiBot’s Differentiator: There Are Plenty of Smart People

姚 agreed that the industry’s narratives are converging and that there is no shortage of people who can do what you can do. AgiBot’s answer is: “There’s no shortage of smart people today. Everyone can go be the smart people; we’ll be the honest, hardworking people.”
The quantitative version is: “Smartness may account for only 10%; the remaining 90% is the capability of a complete system.” AgiBot wants to build that 90% first and stay ready while it connects the technology and penetrates real-world scenarios.
He cited 余凯 of Horizon Robotics: the true moat is not an exceptional algorithm that “no one else has,” but the accumulation of small engineering fixes every day until every brick forms a complete wall.
AgiBot’s complete system includes corporate governance, mass-production capability, talent development, and technical planning.

36. VLA and World Models: Lineage, Limits, and Neither as the End State

World models “come back like a period,” appearing whenever autonomous driving, gaming, or embodied AI becomes hot. At core, they digitize and neuralize physical laws and state-transition rules.
VLA is the extension of language models and multimodal language models into embodiment. 姚 pointed to Google’s PaLM-E in the first half of 2023: PaLM plus Embodiment, applying a language model to robot-action post-training.
The latest wave of interest in world models came from work such as Wayve’s GAIA-1 and Tesla’s efforts around 2023, as well as advances in video generation. World models can capture spatiotemporally continuous physical details such as liquid flow, shattering glass, and soft-object deformation.
VLA’s limitation is the representational gap between language and action. Language is highly abstract and information-dense: a novel may take a few KB, while a film adaptation may require tens of GB. Language is highly compressed but lossy.
His end-state view is explicitly provisional. VLA developed primarily from language, while world models developed more from 2D video and third-person views. Robots operate in a 3D physical world from a first-person perspective, so neither is a complete final form. The future may involve a layered architecture in which the 2 work upstream and downstream.
VLA will still be necessary. Robots need to understand open-ended instructions, map them into plans, and then into locomotion, manipulation, and interaction. That requires the logical-reasoning space embedded in language.

37. GO-2, Top Rankings, and an Automated Data Loop

AgiBot’s value is “pursue excellence; if we do it, we do it at the industry-leading level.” 姚 said the recently updated GO-2 model introduced an action chain of thought, asynchronous execution, and other mechanisms, reaching first-place-level performance on LIBERO-Plus, Genie Sim 3.0, and other leaderboards.
The trigger system from autonomous-driving shadow mode has been deployed on robots. If precision loading and unloading at Longqi’s factory fails to place an item for an extended period, rules and monitoring models automatically trigger data return. The team then determines whether the issue is software or hardware, or whether changes in line position, lighting, or other factors exposed insufficient robustness and generalization. New training data is added and a new software version is pushed out.
The engine of iteration is not leaderboard chasing or demo production, but validating, refining, and iterating the entire system around real scenarios. AgiBot’s embodied business unit has nearly 500 people covering hardware, quality testing, edge and cloud software, algorithms, and data.

38. The 2–3-Year Inflection: Returned Robot Data Becomes the Mainstream

Data sources are stage-dependent. The industry is currently in a launch phase, absorbing simulation, real-robot, human-wearable, and human-centric data wherever possible. Once real-world deployment becomes the main source of incremental data, robot-returned data will become the dominant source. 姚 expects that shift in 2 or 3 years.
The reasoning is historical analogy. If today corresponds to the early Transformer and BERT period, reaching GPT-3 took roughly 3 years. 姚 also believes scaling law should apply across other AI paradigms.
诗婕 pointed out the difference: language models only convinced the outside world that the direction worked at GPT-3, while embodied AI has not yet reached GPT-1 and the industry is already piling in.

39. Migo’s Origin: A Dinner, Scale AI, and the Higher Bar for Physical-AI Data

The starting point came from Sequoia investors, who raised Scale AI over dinner. 姚 said Scale AI’s last financing may have valued it at more than $20B before Meta effectively absorbed it and Alex Wang joined Meta.
Scale AI mainly handles data labeling in the digital world: internet video, images, and autonomous-driving data are relatively easy to obtain. Labelers have computers and can begin after 1 day of training; the core advantage is largely labor operations.
Embodied data requires entering physical environments and first solving access to supermarkets, hotels, and factories. It also requires hiring and managing people, deploying robots, cameras, grippers, and wearable equipment, and paying for large-scale video storage, processing, and compute. 姚 therefore believes the complexity and barriers of embodied data should support a platform worth more.
The confidence comes from existing capacity. AgiBot has spent 1 year building data operations, and AgiBot World has significant industry influence. Nvidia’s GR00T N1 Tech Report shows that 80% of its real-robot data came from AgiBot World, and only part of the dataset had been used at release. Papers normally disclose data sources and shares, allowing the usage mix to be inferred.

40. US vs. China Data Markets: The US Does Not Crowd In, China Has Not Produced Scale AI

姚’s view of the market structure is that many US industries have a clear leader and runner-up, while a third player may not exist. In China, a hot industry can attract hundreds of companies. China has a large data-labeling business but has not produced a Scale AI-style leader because competition quickly becomes a price contest: Shanghai versus the west, adults versus students and vocational-school workers, with costs continually pushed down.
When the technical barrier is low, everyone thinks, “If you can do it, why can’t I?” The scene, labor, equipment, and compute requirements of embodied data may change that structure.
Scale AI’s loop is instructive: accumulate data through the data business, train its own models, and then use those models for pre-labeling and automated labeling to improve service efficiency. 姚 sees the model as relevant.
On the concern that companies may be “building a plank road in the open while secretly marching through Chencang,” he acknowledges that proprietary data can create an advantage. But he believes the technology itself holds few secrets; the real contest is the data pipeline, mix, quality, and full-stack system. The industry may ultimately need a public, one-stop data-services platform rather than every company duplicating heavy investment.

41. The Data Arms Race Is About Speed, Not Stockpiles

The race is already underway. When AgiBot World launched earlier this year, 1M data points and more than 3,000 hours were the largest real-robot dataset. By the end of 2025, Generalist claimed 200K hours; 姚 believes those were primarily data without a robot embodiment. The host also mentioned Sunday’s data strategy, though the name was presented with uncertainty.
姚 expects this year to become the year of the data arms race, with companies openly targeting 1M, several million, and tens of millions of hours. Sitting out and watching could mean falling behind.
The core formula is: “It’s not about who has how much data, but who has how much data at what speed.” If it takes 2 or 3 years to build the data scale others had 2 or 3 years earlier, the data may be proprietary but still 2 orders of magnitude behind.
Customer language confirms the urgency: “as many as possible, as soon as possible.” Prices may decline as scale rises, but for large technology companies, 1M hours and roughly RMB1B in purchases is still not an especially large number.
姚 relayed a consensus between Generalist and Google DeepMind technical staff: if data is the only variable that determines the winner, Chinese companies can declare victory today. China has industrial environments, relatively low labor costs, and a mature supply chain.

42. What Makes Good Data: Real, Diverse, and Related to Failure

Migo has 2 advantages. AgiBot develops multiple algorithms in-house and understands what data they require. It also has a data-operations system that keeps data teams dynamically aligned with R&D and algorithm teams, allowing customized collection and validation of data effectiveness.
Quality has several dimensions: stable links across the chain, including time synchronization and sensor-to-actuator calibration; diverse distributions, with the same pick-and-place task varying by position, lighting, background, and other factors; and dynamic adjustment by model stage, with targeted collection of setups where the model still fails.
The clarification around “dirty data” matters. Dirty does not mean low quality; it can include diversity and failure samples. Seeing only successful cases creates survivor bias. Robots also need to know how to recover after failure.
姚 said that in imitation learning, if the system sees only successful experiences, execution deviations gradually amplify, with failure probability rising as the square of execution time. Reinforcement learning is more like a child learning to walk: it explores through crawling, falling, success, and failure to learn what is stable and what is not.
The floor remains data quality. A blacked-out camera or a trajectory off by several centimeters is not useful dirty data; it is still low-quality data. “Garbage in, garbage out.”

43. The Simulation Debate: No Ideology, Just Four Nines

On the simulation route represented by 王鹤, 姚 leaves room for the approach. For some tasks, environments, and workpieces, simulation can handle extensive early development and validation. Technologies such as Genie Sim can reconstruct environments at high fidelity, alter elements, and test generalization. Humans also learn not only through direct contact with reality, but through classrooms, books, recordings, and films.
The boundary is clear. Some environments, tasks, and objects involve extremely fine physical interactions; only real contact and interaction can produce a data distribution that matches reality. Those cases require real-world collection.
The practical conclusion is blunt: 80% or 90% success in simulation or on a real robot may be impressive academically, but factories require at least 4 nines. At that point, teams need to do whatever is necessary to hit the standard rather than argue over technical schools.

44. Migo’s Name and Business Model: The Didi of Embodied Data

The name was chosen first for memorability and approachability, like Didi and Meituan; second, bees collect nectar just as Migo collects data, while also symbolizing search and exploration.
The deeper idea is distributed. “Data acquisition in the future will definitely be distributed and clustered.” Millions of robots will be spread throughout society, returning data every day and creating bee-swarm-style social coordination. The mission is “to put the world’s data to work for AI.”
Unlike Didi, embodied data is not a To C market. Demand is concentrated among KA and large enterprise customers. The company must first occupy their minds through a To B model, then develop upstream capacity.
The launch event highlighted prototype data completion within 48 hours and customer responses within 24 hours. 姚 explicitly confirmed “response within 24 hours”; he did not separately confirm 48 hours as a universal industry standard. The reason is that customization remains high and requires continuous communication with frontline R&D and algorithm teams.
Migo plans 10M hours of capacity in 2026. 姚 equates that to 10K people, or robots operating for a year at a normal workload of more than 10 hours per day. If achieved, it would be “absolutely leading.” Robot-collected data still accounts for a large share today, but embodiment-free and human-centric data is scaling quickly; the human workforce is expected to overtake robots.
2 Migo products have already launched. Five-finger wearables and full-body devices are next, with products defined around market demand.

45. Globalization and Compliance: Going Overseas From Day One

The launch event offered simultaneous interpretation in English, Japanese, Arabic, and Korean. More than 5,000 people registered to attend, while the source described attendance as more than 1,000 and potentially more than 2,000, constrained by venue capacity. The figures all point to Migo’s day-one global positioning.
Migo’s production nodes are not limited to China; it has laid out capacity in Southeast Asia, the US, Japan, South Korea, and the Middle East. Existing players such as Scale AI may also build capacity in lower-cost regions such as Southeast Asia, either directly or through partners.
Data production and labeling are fundamentally operations-heavy businesses. The economics ultimately come down to cost and efficiency.
Compliance systems are still being built. Cross-border transfer of Chinese data will eventually follow a formal, complete process. In regions with legal or regulatory requirements, data will not be sent back to China or another market; it will be stored locally.

46. Data Silos and Common Standards: End-Effector Abstraction

Hardware specifications and data requirements are not yet fully standardized; the industry is still working through individual benchmark cases. 姚 expects basic consensus within roughly 1 year on FOV, frame rate, resolution, camera layout, and other modalities.
Different hardware does not mean data is completely non-transferable. Different arms and humanoid configurations can represent actions in the same space, ultimately abstracting to the end effector—essentially the wrist trajectory.
A world model can fundamentally be a “predict next frame” system. A frame is a 2D pixel matrix, so the model can learn from different robots and human data. It may also extract general physical laws from how different robots manipulate objects.
The entry point for breaking data silos is putting idle capacity to work. Many collection and training sites are not professionally or efficiently operated; joining the Migo system can improve efficiency through training and demand allocation.
The data marketplace will accept data collected tightly to Migo’s standards while also welcoming third parties to upload their own shelf-ready datasets. The 3,000-square-meter Zhangjiang collection center is the birthplace of AgiBot World, but functions more like a showroom and command center: it designs workflows, runs pilot collection, refines the process, and then sends the playbook to regional sites for scaled execution.

47. The Open-Closed Spectrum: Get More People Using Robots

The governing principle is to get more people using robots in valuable scenarios. Open-source data lets more teams experience the model capabilities unlocked by high-value embodied data.
Platforms including Polar Studio, Lingxing, and Lingchuang target developers. Lingchuang can convert a dance video into a robot dance, while Lingxing allows users to customize a robot’s character and voice.
The closed loop depends on user authorization. With permission, platform data can feed back into the underlying models, creating a chain from commercial success to the data flywheel across task, motion-control, and interaction models.

48. The Standards Fight: G1–G5, Government Leadership, and a 2–3-Year Entry Barrier

In 2024, 姚 proposed a G1-to-G5 technology roadmap for embodied AI modeled on autonomous driving. The host said it had been incorporated into Shanghai’s grading guidelines for intelligent embodied-AI development. 姚 confirmed that the standards are led by the Shanghai Artificial Intelligence Industry Association, with AgiBot participating in their formulation. AgiBot is also a deputy-chairman member of the Ministry of Industry and Information Technology’s humanoid-robot standardization technical committee.
Migo wants to promote 2 types of industry consensus: data-quality standards, and commercial standards covering pricing and transaction models. The goal is to avoid inconsistent quality and an early price war that damages the industry.
Government standards may not be immediately visible, but large-scale commercialization will gradually create application thresholds. 姚 expects that process to unfold in roughly 2 or 3 years.
The current position is an intermediate state between G3 and G4. G3 means end-to-end autonomous execution in defined scenarios and tasks; G4 means end-to-end execution across scenarios and tasks.

49. The Eve Before the Eve: The 100M- and 1B-Hour Roadmaps

The 100M-hour threshold came from brainstorming with many users and may correspond to the scale of training-corpus time used by GPT-2 and GPT-3. Demand above 100M hours cannot yet be forecast accurately; the first step is to use 100M hours to validate the current thesis.
The host summarized the roadmap as “10M hours this year, with a chance of reaching 100M globally next year.” Asked about the 1B-hour level in the roadmap, 姚 said 2027–2028 was achievable—still within 2 or 3 years.
Where will emergence happen? 姚 answered: “It will definitely happen in the laboratory.” The industry will first collect enough data from factories and other scenarios, then trigger the technical breakthrough in the lab.
The moment he most wants to see is an “aha” moment: a robot appearing to have some form of autonomous awareness, planning for itself and responding to instructions in a complex environment instead of merely mapping images into movements.
Asked whether he had experienced anything similar in his career, he said that only the birth of large models in recent years may have produced that feeling in the history of technology. He has not experienced it himself yet and is looking forward to it.

50. Closing Notes: Valuation, Rumors, and Self-Positioning

The host put Migo’s current valuation at roughly RMB3B. 姚 said the latest round was already above that level. Judging by revenue, profit, and long-term scarcity, he does not believe the valuation is too high.
On the claim that every company will eventually collect its own data, he expects supply growth to lag demand growth. The future will require hundreds of millions of hours or more, with large numbers of devices and people entering the real world. Migo and the rest of the value chain therefore still have substantial room to grow.
He explicitly denied rumors that he would start an independent company: “That definitely will not happen.” Market information can be a smokescreen; the source says friends and candidates asked about it during recruiting, without specifically mentioning headhunters.
He does not fully accept the label “commercial operator.” He would rather be defined as “a pioneer actively advancing robot technology and commercial applications.”
His advice for building an embodied-AI company from 0 is that the company must go through the full path—from product to mass production, commercialization, and ecosystem building. Even OpenAI and Anthropic need to monetize through code models, tokens, or similar products; no company can raise indefinitely and prosper forever simply by saying, “I’m a foundation model.”
Data and compute are similar, but data may be harder to secure. Compute is a mature, standardized product that can be obtained quickly with money. Data requires demand planning, people to execute it, operations, and delivery—a long, systematic process.

Cerebras Early Investor 周楠 on the Nvidia Challenge and Baidu's US Lab Ep. 015 - DG Matrix on 800V DC vs Legacy AC in Data Centers