Sun Xinyang had been hearing variations of those warnings with growing frequency from recruiters and former colleagues. The growing urgency pushed him to seriously consider leaving Xpeng this spring.
In the three or four months before he resigned, Sun found himself reporting to a new direct supervisor almost every month. Desks would fill up, then empty again. Xu Zhou, who also worked at Xpeng, saw the same pattern on his team. Over six months, he watched nine colleagues leave. Almost every month, he found himself typing another “wish you all the best” in the group chat.
Almost all of them were going to the same place: robotics companies.
The talent rush into robotics began in 2024 and has continued for nearly two years, with smart driving engineers among the sector’s most sought-after recruits. But interviews conducted by 36Kr suggest the migration is nearing the end of its most frantic phase.
At its peak, stories circulated across the industry of robotics companies offering smart driving engineers triple their salaries. According to 36Kr, one robotics founder was said to have personally approved an offer to cover a prospective co-founder’s mortgage to persuade him to join.
Those stories are becoming rarer. These days, a 50% pay increase is more typical, according to people interviewed by 36Kr. The premium remains, but it is becoming more measured.
Hiring criteria have also gone from broad to specific. At first, experience in autonomous driving could be enough to get someone through the door. Engineers who had worked on perception or planning could leverage a stint in automotive R&D into double the salary and a package of stock options.
A year later, the bar has risen. Robotics companies increasingly want people who have worked on vision-language-action (VLA) models. Some want only engineers with hands-on experience working with physical robots.
Money and resources move in cycles. So do career opportunities. With Unitree Robotics briefly surpassing RMB 400 billion (USD 59.5 billion) in market capitalization on its first day of trading, and several leading robotics companies preparing for listings, capital in the sector appears to be concentrating among the leaders.
Chen Wei, a former chief artificial intelligence scientist at Li Auto who founded Xieyue Intelligence this year, told 36Kr that he believes the window will close in 2028.
“We’ve already moved past the stage where people only look at ideas and models. Everyone is now focused on products and real-world deployment,” Chen said. The embodied intelligence industry, he said, now has “everything ready except the model.”
Even if physical-world models develop more slowly than large language models, he added, they may need only another two or three years to mature. As robotics companies move into a new stage, the kinds of people they need are changing with them.
Should engineers switch jobs? Is moving into robotics just a way to make a quick buck? Will they end up walking into a mess?
Many people are still watching from the sidelines. The automotive sector is established and tangible. Robotics, by contrast, has yet to converge on a technical path and remains awash in capital attracted by the technology’s potential. Of the hundreds of companies operating today, industry participants expect many to disappear as the market consolidates.
To understand what the transition is actually like, 36Kr spoke with several automotive veterans who have already moved into robotics.
Those who left
Lu Yuan joined Li Auto straight out of college in 2023. That year, the company sold 370,000 vehicles. Founder Li Xiang had begun having senior executives study Huawei’s methods and was still talking extensively on social media about his ideas for the industry, organizational structures, and products.
At the time, many people in the automotive sector still believed the competitive landscape would be shaped by product managers who designed products, rather than procurement managers focused on squeezing costs from suppliers.
Automakers were also investing heavily in R&D and treating technology as a key competitive advantage. Lang Xianpeng, Jia Peng, and Xia Zhongpu were all still at Li Auto. Lu worked alongside them through the company’s 100-day end-to-end development sprint and the development of Li Auto’s VLA system.
Later, the team that had helped establish Li Auto as a serious contender in smart driving began to disperse. Lu also felt the automotive industry was gradually placing less emphasis on technology as the center of competition. Executives including Xia Zhongpu, Jia Peng, and Wang Jiajia left one after another to start companies. Lu’s colleagues followed. Some went with Jia to Simplexity Robotics, while others were recruited by embodied intelligence companies.
Every so often, another familiar name disappeared from the work chat.
Eventually, Lu and his colleagues stopped talking privately about smart driving. Instead, they began asking one another whether robotics companies needed people with VLA experience, whether embodied intelligence companies were still hiring smart driving engineers, and whether such companies were worth joining.
Nearly every leading Chinese automaker has explored robotics over the past two years. Last year, after Xpeng’s Iron robot spent just two minutes walking across a stage, the company’s shares rose about 20% over the following three days. Li Auto formally established a humanoid robotics business early this year, while BYD and Changan Automobile have also been conducting internal research.
Yet Lu, Xu, Sun, and others barely considered staying at automakers to work on robots.
“Smart driving is still the core business. The company isn’t going to let too many people transfer over,” Lu said.
At Li Auto, talent, accelerator cards, computing power, data, and other resources were concentrated in the automotive business. Its robotics team also had plenty of smart driving engineers, but according to Lu, most had moved there after struggling to advance within the smart driving organization.
Xu had a similar impression. He knew from the time he joined Xpeng that the company had long maintained a robotics team. In 2025, it internally demonstrated robots moving goods and tightening screws in factories. But when the time came for a public launch event, most of the demonstrations could not be performed live.
“A lot of people on the robotics team had transferred from smart driving, but most of them had also lost out in internal competition,” Xu said.
Automakers do not want to lose too much talent, either. According to Xu, Xpeng requires departing employees to observe a three-month cooling-off period.
“Robotics companies usually need people urgently. If you make them wait three months, it’s very easy for the offer to disappear,” Xu said, visibly frustrated by the requirement.
“You can’t stop it. There’s nothing you can do. This trend is irreversible,” he said.
In conversations with 36Kr, Xu, Lu, and Lin Yejun all made essentially the same point: the endpoint for smart driving had become clearer, making incremental work feel less meaningful to them.
“Two years ago, smart driving was moving from 60 to 80,” Xu said. “Now it’s moving from 90 to 95. The marginal return on every additional effort is falling. But embodied intelligence is only getting started. It’s going from 20 to 60. For someone who wants to build something meaningful, that has much more potential.”
During three years at Li Auto, Lu participated in many critical projects but remained an individual contributor. The likely path upward was to advance gradually through the promotion system.
At the robotics company he later joined, Lu was already leading a ten-person team and overseeing the full R&D stack for embodied intelligence.
Lu studied robotics as an undergraduate, focusing on robotic arm control. At the time, robots could largely be controlled through rules. Machine intelligence today, by contrast, is fundamentally an AI-centered systems problem. To work on the most important technical questions, he said, engineers need exposure to the entire stack.
For the first time, Lu felt he had returned to what he had originally wanted to do and could make a meaningful contribution.
“In a stable, mature industry, it’s hard for you to sit at the same table as those young prodigies,” Lu said. “But in a chaotic new industry, everyone gets a chance to reshuffle the deck.”
For some workers, embodied intelligence has created opportunities for rapid professional advancement.
Another algorithm engineer at an automaker told 36Kr that a former colleague had been a rank-and-file employee at Nio. After moving to Agibot, he was already leading a preliminary research team of more than ten people, working on foundation models and publishing papers.
“Working at an automaker is exhausting, and embodied intelligence is just as exhausting. But over there, they pay more, there are stock options, and there’s a future. You don’t need to think very hard to figure it out,” Lin Weichang, an HR professional in BYD’s smart driving business, told 36Kr.
“Going into embodied intelligence is making a big bet,” he said. “And even if the bet doesn’t pay off, you’re not losing money. Robotics companies are paying plenty of cash.”
That repricing of talent spread quickly. A headhunter who has long recruited for smart driving roles said that of more than 60 candidates he successfully placed in the first half of this year, 95% came from smart driving teams and 90% had considered embodied intelligence positions.
Sun saw the change directly in the jobs recruiters sent him.
Two years ago, when he was looking to switch jobs, roughly three-quarters of the opportunities recruiters recommended were in smart driving, while one-quarter were at embodied intelligence companies. This year, the ratio has reversed: three-quarters are in embodied intelligence, and one-quarter are automotive-related.
If Sun remained in smart driving, he estimated that he would probably get a pay increase of no more than 30%. Moving into embodied intelligence, by contrast, made a 50% raise the baseline, according to the opportunities he was seeing.
“People with relatively little experience in frontier areas are getting total compensation of around RMB 600,000–800,000 (USD 89,000–119,000), while senior engineers can reach around RMB 1 million (USD 149,000),” the headhunter said.
When a new industry takes off, wages can rise quickly as fundraising intensifies competition for a limited pool of talent.
A similar dynamic played out during China’s semiconductor boom in 2020, when a wave of domestic GPU startups emerged. Some engineers changed jobs every three months, more than tripling their salaries within a year. Five years later, as some of those GPU companies went public at high valuations, engineers who had made the jump stood to earn what could amount to decades of income for engineers in other sectors.
The headhunter gave the example of an algorithm engineer who had worked at Xiaomi for only a year before moving into embodied intelligence. His total compensation rose from RMB 700,000 (USD 104,000) to RMB 1.1 million (USD 164,000), plus stock options.
Even if the options ultimately became worthless, the headhunter said, the engineer would still have come out ahead.
According to ITjuzi, China’s embodied intelligence and robotics sectors recorded 288 funding deals in the first half of 2026, totaling more than RMB 46 billion (USD 6.8 billion).
Industry estimates reviewed by 36Kr suggest that more than 20 embodied intelligence companies are now valued above RMB 20 billion (USD 3.0 billion), and the number is still rising.
“A lot of institutions have people specifically watching Shanghai Jiao Tong University, Harbin Institute of Technology, and the Hong Kong University of Science and Technology,” an investor said. “Others keep a close eye on DJI and the automakers. As soon as someone leaves to start a company, we reach out immediately.”
A person familiar with Contemporary Amperex Technology’s corporate investment activities said two types of projects are especially likely to win internal approval: “young prodigies starting companies, and automotive veterans starting companies.”
Even people who were not the most senior business leaders can raise substantial sums in this cycle.
One industry source said the former head of Horizon Robotics’ core HSD algorithms attracted interest from several leading investors as soon as he left to start a company, quickly pushing its valuation into the eight-figure USD range.
A migration had begun. For many smart driving engineers, what they were leaving was an industry whose direction seemed increasingly clear. What they were heading toward was a field where there was still no settled answer.
From an apartment building to a thatched hut
“After I arrived, I realized there was nothing here.”
Less than a month into his new job, Xu discovered he had not simply moved into another version of the smart vehicle industry.
Another former automaker algorithm engineer put it this way: “We used to live in an apartment building. Now we’ve moved into a thatched hut.”
Data platforms, training frameworks, and computing platforms that automakers take for granted all have to be built piece by piece at robotics startups.
Xu’s current employer rents computing capacity from Alibaba Group, Baidu, ByteDance, Tencent, and other companies. The different computing platforms are not interconnected, making it cumbersome to move data around when he trains a shared model.
“Something like that can easily eat up a person’s entire day,” he said.
At Xpeng, the training accelerators Xu’s team used were supplied by Alibaba Group, while storage came from Alibaba Cloud. The compute cards could read training data directly.
“We always thought that was basic infrastructure,” Xu said.
Startups are smaller and can buy computing capacity only in small batches. Cloud providers also give them less preferential access to resources, forcing them to procure capacity piecemeal. Xu said his company’s founder has now assigned one person specifically to solve the problem.
“There’s no way around it. That’s what startups are like,” Xu said.
And it is not only data. At a startup, many things have to be built from scratch.
After Lu became a team leader, he realized that most of his time was not spent training models. He was recruiting people, sourcing robot hardware, and building processes.
At Li Auto, someone had already been assigned to handle all of those tasks. At the startup, everything reset to zero.
“You’ve seen the way Li Auto fights battles, with everyone responsible for breaking through on their own front,” Lu said. “It’s only after you leave that you realize how rare that kind of setup is. Here, there’s nothing. You have to manage everything yourself.”
Some people struggled to adjust.
Many of Xu’s colleagues came from Alibaba, Baidu, and ByteDance, and anxiety set in quickly. To them, the startup environment felt minuscule.
Some had been there for only two months before they started asking around about their next jobs. They were considering moving to startups such as Agibot or X Square Robot, which by comparison already had more developed systems. At least there, employees could return to being one cog in a larger machine.
Xu was less alarmed.
Perhaps, he said, it was because he had worked on mass production and was more accustomed to getting his hands dirty. Smart driving had also been built piece by piece in its early days: collect data, clean data, and run models. Plenty of tedious, grueling work had to be done manually.
Xu was still in the honeymoon period of his new job. Lu, after six months in embodied intelligence, had already reached a more fundamental conclusion: “Smart driving and robotics both fall under embodied intelligence, but they’re not the same thing.”
“Embodied intelligence really is a completely new industry. There are no existing roles you can use as a reference,” Lu said.
Robots have different degrees of freedom and use different motion algorithms from cars.
Algorithm engineer Li Liyang experienced the same gap.
Before joining, he thought anyone who had worked in smart driving could do the job. Once he actually started training models, he realized the work was largely unfamiliar.
“The overall code stack, the training methods, and a lot of the open-source frameworks are actually much closer to the way large models are trained,” he said.
“Everyone knows that eventually there will be one big model and a clear paradigm, and then you’ll just pile in data,” Li said. “But right now, nobody knows what that thing will be. Will it be the VLA route? The world model route? Or some other path nobody has taken yet?”
The direction remains unclear. Department heads do not necessarily know which way to go, either. No one knows which paradigm will ultimately work.
After moving from an automaker to an embodied intelligence company, Li noticed another difference.
At an automaker, even when someone is being marginalized, a rigorously enforced OKR (objectives and key result) system gives everyone a defined target. Even if the work is tedious or unpleasant, there is always something assigned to do.
At his new company, without a strict OKR structure, employees have to find work for themselves.
“Everyone is confused,” Li said. “But people also pretend that what they’re working on is important, and then tell everyone else that their own route has a future.”
Then they stopped being so sought-after
After Xu joined his company, he realized he had caught the last train.
The company was no longer using headhunters to poach employees. Lin Yejun, who works in Alibaba Group’s embodied intelligence division, also said his department had stopped hiring.
Over the past two months, Lu interviewed more than 100 people. Yet he said he would rather hire new graduates and interns who have actually worked on robots than former colleagues who have spent their entire careers in autonomous driving.
Embodied intelligence companies care more about how quickly someone can grow, he said. Younger candidates are more willing to make mistakes and learn new things.
He laughed after saying it, then added: “A lot of people hire this way. It’s not just me.”
After six months at the new company, Lu has come to believe that hiring someone who only wants to reproduce what worked in the past can slow progress dramatically. Eventually, he said, that person also discovers that the new field is full of problems their old methods cannot solve.
Another startup Lu knows hired large numbers of engineers from the smart driving sector last year. They reused their previous methods as the company pushed ahead. The result, according to Lu, was a tangled technical and code stack, with engineers often unable to determine where errors were coming from.
“One big problem with smart driving people is that they think doing AI means you should first build a large-scale data cluster and a closed-loop data system,” Lu said.
“But robotics doesn’t work that way at this stage. It’s a puzzle that has only just begun to be assembled, and it’s far more complex than smart driving.”
For Lu, the priority now is to choose a direction, move forward on a limited scale, and complete one small section of the picture first.
“Nobody knows what kind of ‘house’ embodied intelligence is supposed to become,” he said. “You shouldn’t start a massive ‘construction project’ before you even understand robotics, hammering away and piling up lumber. You might eventually discover it isn’t supposed to be made of wood at all.”
In just one year, embodied intelligence companies have markedly changed how they view automotive talent.
During the startup phase, many needed to assemble teams quickly. Even an algorithm engineer working on occupancy perception at BYD could land a job at a high-profile company such as X Square Robot.
Today, after repeated development experiments, companies have a clearer idea of the people they need.
Poaching has also changed direction. Rather than recruiting broadly from automakers, companies increasingly prefer to poach directly from competitors.
After Sun joined an embodied intelligence company, two colleagues around him were recruited away by competitors in less than a month. Some employees were talking with other companies even while still on probation.
Most had just received 30% raises when joining their new employers, then went back to the market for another round of negotiations. Expectations rose accordingly.
“Who’s going to move for only a 30% raise?” one of Sun’s colleagues said.
Within one or two months, one engineer used AI tools to get a physical robot running, recorded a video of it performing movements, and sent the footage to another company.
Before long, another offer arrived.
“There’s no moat in the algorithms. Everyone copies everyone else,” an algorithm engineer at a robotics company said.
According to Sun, the salaries of people with hands-on experience working with physical robots were rising rapidly.
Watching one colleague after another leave, Sun’s boss could do little more than say “we wish them well” at the weekly team meeting.
Capital is still flowing into robotics, but investors’ patience appears to be shortening. Robotics companies increasingly have to demonstrate progress through working prototypes and plans for mass production.
The segment is reshuffling amid uncertainty. Roughly every three months, industry participants say, a new technical narrative emerges, along with a new fundraising favorite. Companies know the window is still open. Nobody knows for how long.
That is changing demand for talent, too.
The first generation of robotics companies wanted people who knew autonomous driving.
Today, they increasingly want people who can use a demonstration to prove they can actually make a robot work.
For now, practical experience with physical robots appears to be carrying more weight.
The biggest bubble, and the biggest possibility
Lu told 36Kr he has never assumed his company will make it to an IPO.
“My expectation is simply that the company can stay alive for another year or two,” he said.
He knows there is a bubble in robotics and expects many companies to fail. Yet compared with staying in smart driving, he sees the latter as the bigger risk.
Sun had already begun to feel marginalized at the start of this year.
“The smart driving ‘cake’ has already been baked. How it gets divided depends on whoever is in charge,” he said. “I’m just a cog in that machine. There’s only so much cake I can get. If I don’t take my piece, it goes to someone else.”
As he became increasingly sidelined, Sun felt his value slipping away.
It is not only automotive workers. An algorithm engineer who moved from the internet industry told 36Kr that he, too, understood the enormous uncertainty in robotics.
“But as long as I can make money from the next wave, that’s enough,” he said.
As automakers focus more heavily on costs, profits, and price competition, while internet companies cut jobs and adopt AI tools that can automate some programming work, embodied intelligence has become one of the few emerging industries still willing to pay a substantial premium for technology and algorithm talent.
That has made it a natural destination for workers looking for their next opportunity.
One robotics entrepreneur told 36Kr that more than 200 companies are now exploring different technical approaches at the same time. In essence, he said, everyone is waiting for the answer that will eventually define the industry.
Until that answer emerges, companies need people from different backgrounds. Autonomous driving engineers, large-model researchers, and robotics specialists all have a chance to take a seat at the table.
But once foundation models begin to converge and competition shifts back toward computing power, data, and capital, today’s window could close quickly.
“Over the long run, technology itself is difficult to turn into a lasting competitive advantage,” the entrepreneur said. “I may be three months or six months ahead of you, but eventually it spreads.”
For that reason, he believes the real window for startups may last only from 2026 to 2028.
“After that, it becomes a game for the giants,” he said.
The automotive market is locked in intense price competition. Internet companies have reduced headcounts. Consumer industries remain sluggish. Among emerging sectors that are still expanding, AI model companies tend to need relatively small numbers of highly specialized workers.
That has left embodied intelligence as one of the few sectors still absorbing large numbers of technical workers.
“Embodied intelligence companies have raised so much money! Are they hiring?” a former employee of a major internet company asked 36Kr, hoping to make the switch.
After leaving a big tech company, he had joined a traditional industry, believing he had finally reached safe harbor. Whether because of bad luck or a harsher reality, he never managed to find his footing there.
When 36Kr warned him about the bubble and risks in embodied intelligence, he was unconcerned.
“That’s fine,” he said. “I’ll do it for two years. As long as the money is good.”
Many people interviewed by 36Kr believe there is a bubble in embodied intelligence and expect many companies to fail. But from where they stand today, the industries they came from appear to offer fewer opportunities than before.
Embodied intelligence offers no certainty. For many of them, though, continuing to move may feel less risky than standing still.
Note: Xu Zhou, Sun Xinyang, Lu Yuan, Lin Yejun, Li Liyang, and Lin Weichang are pseudonyms used at the interviewees’ request to preserve their anonymity.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Xiao Man for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.73 = USD 1 based on estimates as of September 10, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.