Back to Industry Dialogues

Industry Dialogues

Peter Chen: The Next Decade Will Produce World-Class AI and Robotics Companies

A full conversation with Alphaist Partners founder Peter Chen, Zhe Chen, on robotics investing, founders, data, deployment, company building and the new fund.

Zha Dami2025-02-1019 min readWeChat original
Contents

Investing in robotics has a history of more than a decade in China, but relatively few investors have built a record, developed deep understanding and stayed with the field over the long term. In my view, Peter Chen, whose Chinese name is Zhe Chen, is one of those who have done it particularly well. Against a difficult fundraising backdrop, he has just established a new fund, Alphaist Partners. I was delighted to speak with him at this moment about his investment experience, his perspective on embodied intelligence and his message to the founders the new fund hopes to support. This interview retains its February 2025 context; Peter is distinct from Covariant co-founder Peter Chen, whose Chinese name is Xi Chen.

Peter is an engineer turned investor. After high school, he went to Canada and completed undergraduate and graduate studies in electrical engineering. His career began at BlackBerry, working on smartphone communications. He then moved to Silicon Valley and joined Google's Google Glass team, focusing on embedded systems and signal processing.

In 2016, he went to Harvard for an MBA. During that period, he returned to China and joined Morningside Venture Capital, later renamed 5Y Capital, beginning his venture-investing career. He stayed there until the end of 2024, focusing on AI and robotics.

His interest in robotics began with autonomous driving, where he led investments in several companies and watched for technologies expanding into adjacent applications. Examples included Xiaomi-backed Roborock's use of low-cost planar lidar in robot vacuums and DJI's introduction of visual obstacle avoidance in consumer drones. From 2017 to 2018, he began investing more extensively across robotics.

He has now established Alphaist Partners, a new early-stage US-dollar fund, continuing to focus on AI and robotics.

Original article outline

Contents

↓

Part I: Experience investing in robotics

1. Observations from more than a decade of robotics investment

2. What leads to major success in entrepreneurship and investment?

3. What distinguishes successful business and consumer robotics founders?

4. How large are the opportunities in business and consumer robotics?

Part II: Investing in today's embodied-intelligence wave

5. How have the past two years felt?

6. What has changed technologically in this wave?

7. What should we expect from deployment?

8. How would you describe general-manipulation startups' technical approaches?

9. Robot data

10. Startups providing robot data

11. Startups focused only on embodied models

12. Potential deployment scenarios

13. Companion robots

14. Should investors back embodied-intelligence companies focused on specific scenarios?

15. What is the ultimate role of general-purpose robots?

Part III: Working methods and the new fund

16. How do you keep up with frontier research?

17. Recommended reading

18. A message from the new fund

Part I: Experience investing in robotics

1. Observations from more than a decade of robotics investment

Looking at the industry's development, I would divide it into several stages. Around 2014–2015, industrial automation and replacing imported equipment with domestic products were the main trends. Around 2018–2019, e-commerce growth stimulated mobile and warehouse robotics. After 2020, consumer robots—especially robot vacuums—entered the market more broadly, marking an important breakthrough.

A milestone for warehouse robotics was Amazon's acquisition of Kiva Systems in 2012, which helped establish the basis for large-scale deployment. Chinese companies such as Geek+ and HAI Robotics subsequently emerged. Navigation and localization technologies relevant to these deployments became increasingly mature around 2014–2015, but cost and product maturity held deployment back. E-commerce growth and demand for reduced reliance on manual labor, particularly after the pandemic, drove expansion around 2019–2020.

Robot vacuums followed a similar path. SLAM had made important advances by around 2010, but applying low-cost planar lidar in consumer products, as Roborock did, substantially improved navigation efficiency and the user experience. That transition required a long period of technical iteration. These dates describe developments relevant to the product cycle, not the invention of SLAM or the first lidar-equipped vacuum.

As investors, we continually follow technology development and product iteration. At the same time, a robot has to be manufactured at scale and remain stable and reliable, making technical maturity and manufacturability crucial. Frontier technology may not commercialize immediately, but commercialization can accelerate once it reaches sufficient maturity.

Looking back over seven or eight years, the core technologies behind robot vacuums, warehouse robots and drones were largely mature four or five years before their markets took off. We are now seeing another wave of rapid progress, particularly in intelligent and humanoid robots. I look forward to seeing whether these technologies can become mature, scalable industries within a relatively short period.

2. What leads to major success in robotics entrepreneurship and investment?

Successful robotics businesses address a sufficiently large need in a market that is repeatable or standardized enough to support scale. Drones, robot vacuums and warehouse robots are three clear examples. Their shipments have become substantial; consumer drones and vacuums have reached millions or tens of millions of units in the market scopes being discussed. Their markets are large enough, and their applications standardized enough, to replicate products rather than customize a new solution for every customer.

Market size and standardization are central to these companies' growth. Warehouse robotics illustrates the business-to-business case: one warehouse can deploy hundreds or even a thousand robots, making scale easier to achieve than in many fragmented, long-tail applications.

3. What distinguishes successful business and consumer robotics founders?

I think consumer-robotics founders often have strong first-principles thinking and an intense focus on user value. In competitive markets, they iterate and upgrade quickly. Their teams' ability to sustain that effort is another defining quality.

Business robotics demands a more comprehensive set of abilities, particularly leadership and building a sales organization. I do not think Chinese business-robotics startups have a quick shortcut. Founders need to address their weaknesses, seize opportunities, serve leading customers well and expand overseas. These are more complex challenges. Judging from the results so far, Chinese business-robotics companies still have a long road ahead.

Over the long term, An excellent business-robotics company’s ceiling will certainly not be lower than a consumer company’s. But its route will take more time and be harder. Companies that persist and succeed will develop strong competitive positions and defensive barriers.

4. How large are the opportunities in business and consumer robotics?

Consumer companies often reach larger revenue scales, which is unsurprising. Like consumer electronics or appliances, products bought widely by households can become very large businesses. But consumer companies do not always have the advantage. Some of the businesses that create lasting public-market value and become difficult to replace are industrial or business-facing companies. They accumulate domain knowledge and barriers over many years, supporting continued growth. Huawei, Hikvision and Inovance illustrate the kinds of business and industrial capabilities I have in mind, though their business mixes differ.

In the Chinese business-robotics companies I am discussing, annual revenue commonly ranges from several hundred million renminbi to around RMB 1 billion. Despite technical and automation progress, many remain constrained by their organizational and product capabilities. Until they can work more broadly across industries and applications, revenue is limited by the scenarios they serve, the size of those industries and their degree of standardization. Successful companies break through these limits. This is my February 2025 assessment of that company group, not a statement covering every global industrial-robotics supplier.

Part II: Investing in today's embodied-intelligence wave

5. How have the past two years felt as a leading robotics investor?

Robotics is at an important moment and attracting considerable attention. Two or three years ago, AI robotics was less widely discussed, and traditional startups focused on putting specific products into specific applications. Founders were thinking about marketing, distribution and team management. The previous technical paradigm had become reasonably mature, so the emphasis was deployment.

Over the past two or three years, large models and embodied intelligence have introduced a new perspective. People see the possibility of a new robot paradigm or industrial structure. That has raised expectations for a technological revolution, especially while it is still in its early stages.

The entrepreneurial wave around AI and robotics, particularly embodied intelligence, has therefore grown stronger. Many investors see the combination of AI and robotics as the most worthwhile area to invest in following large models and ChatGPT. New and established players are entering, and investors are actively positioning themselves in the field.

6. What has changed technologically in this robotics wave?

Looking at the development of this particular AI-robotics startup wave, its early form appeared around 2015–2016. Advances in autonomous driving and computer vision helped produce robots with stronger intelligent perception, including 3D vision, and better understanding of their surroundings. Compared with conventional industrial automation, that was a substantial leap. AI and robotics research itself began much earlier; these dates describe the commercialization wave I am discussing.

The first wave of these companies emerged in China and the United States around 2015–2017. AI mainly supplied perception, adaptation and closed-loop repetitive motion, resembling the capabilities of China's leading computer-vision startups at the time. This produced a group of robotics companies. By around 2021, deployment had revealed significant bottlenecks in that generation of vision technology.

The biggest change in the last year or two is the progress in embodied intelligence, particularly large models. ChatGPT's release at the end of 2022 helped prompt attempts to use similar large-scale methods to train more general robot models. Google, OpenAI and leading US startups have presented progress suggesting a route from specialized machines toward general-purpose systems. That potential has been a key driver of the recent enthusiasm.

7. What should we expect from deployment?

I am less optimistic about the timetable than some people. Using autonomous driving as an analogy, today's robots feel closer to 2010 than to 2015 or 2016. Google had already demonstrated promising self-driving prototypes, and US research competitions had showcased routes involving lidar, detailed maps, localization and SLAM. They demonstrated feasibility, but there remained a large gap to practical, scaled deployment in complex situations. The major DARPA vehicle competitions took place in 2004–2007; the 2010 analogy describes the stage of maturity, not their event dates.

Many robot capabilities and demonstrations today remind me of those promising early driving demos. Yet major engineering issues, algorithm architectures and training methods remain unresolved. Autonomous driving went through more than a decade of evolution: from modular localization and mapping approaches to 3D bird's-eye-view representations, occupancy networks and end-to-end learning. The interview associates those developments with roughly 2021, 2022 and 2023 and sees reliable scaled robotaxi deployment becoming clearer by 2025. These dates describe the speaker's shorthand for an adoption sequence, not the first publication of each technique or the first commercial robotaxi service.

I think general-purpose embodied intelligence is at least as difficult as driving on public roads and may be harder. I am therefore cautious about solving it quickly. Simpler settings can come first. In driving, ports, enclosed campuses and other constrained environments offered earlier deployment opportunities. The reference to office buildings concerns indoor autonomous mobility, rather than road vehicles driving through offices. Even successful constrained deployments are a long way from the more complex data-driven, end-to-end neural and predictive-control systems needed for open environments.

Efficient deployment in more complex settings may therefore require an algorithmic and technical framework quite different from what we have today.

8. How would you describe general-manipulation startups' technical approaches?

I see several possibilities, but I would organize the discussion around two main routes.

The first is imitation learning. Borrowing the driving analogy, a robot learns from good human demonstrations or trajectories to develop policies with some adaptability and resistance to disturbances. Mobile ALOHA and Diffusion Policy are representative examples. Relatively small demonstration datasets can support useful behavior in particular settings. The challenge is dependence on good human demonstrations: the system learns by reproducing what people show it.

The second uses a large model as a higher-level planning and coordination component, decomposing and arranging tasks. Lower-level manipulation and control are handled by learned algorithms or traditional controllers.

Representative work such as PI’s is, in the original account, an example of combining vision-language models and large language models through end-to-end training. It aims to combine a VLM's general and common-sense knowledge with motor-control capabilities in a more unified model. This direction has attracted considerable attention over the preceding year. Editorial clarification: π0 uses a pretrained VLM backbone and an action expert trained with flow matching; it does not simply combine two independent VLM/LLM systems. The hierarchical route above and PI’s end-to-end VLA are different architectures, and π0 itself also learns from robot demonstrations.

Given the scarcity of robot data and the difficulty of collecting or generating it, I think the first route may deploy more readily in the near term. Over a longer horizon, more data-driven end-to-end systems incorporating environmental common sense may have a higher ceiling and a longer iteration cycle. But PI's and Google's explorations leave theoretical and architectural uncertainties. We do not yet know whether today's structures are optimal. These categories overlap rather than excluding one another.

9. Robot data

One of the largest challenges for robot foundation models is insufficient data. When discussing development paths, the key questions are where that data will come from and how we progress before it becomes abundant.

Language models face a different situation. Once a model such as GPT-3.5 was trained, a product like ChatGPT could deliver value directly to consumers. Access was easy, and usage frequency and scale could grow quickly.

An embodied robot model cannot act usefully in reality without a physical robot and the ability to affect the world. It needs a suitable embodiment and product configuration for deployment and feedback. Only then can we turn model capabilities into user value, collect more experience through use, and support further development.

10. How do you view startups providing robot data?

This year I have spoken with data-service companies and prospective customers about a direction robotics founders readily consider. Major research organizations, including Google and PI, lack data and face high collection costs. There is therefore an opportunity for third-party providers.

But I think the market differs from the one Scale AI entered with AI data annotation.

Scale AI had clear commercial applications and customers willing to pay, and it grew rapidly. In robotics, everyone agrees data is scarce, but each company is still experimenting at its own scale. Numerous providers can supply initial robot-data services—close to ten in the United States, in my assessment, and many in China. For leading model-research companies, providers and their basic capabilities are not necessarily scarce. The challenge for those providers is finding the right industrial position and business route.

Collection is the main service today, but it requires choosing equipment and embodiments. Cross-platform generalization remains weak, and data collected with different arms or hardware may be difficult to reuse, particularly when sensors and viewpoints differ. Its general utility is therefore limited. Providers may have to customize collection for a particular model or major customer such as Google.

At this stage, large companies' budgets for these services are still relatively limited. Many providers want to enter, but model architectures have not converged and large-scale collection commitments have not formed. That leaves the market uncertain.

11. How do you view startups focused only on embodied models?

At today's stage, I am unsure how the business model for a foundation-model-only startup develops. OpenAI can serve over 100 million users through its models and generate over US$100 million through subscriptions or API use. Language-model companies have that direct distribution advantage. For an embodied-intelligence company without access to a deployment scenario or robot capability, it is hard to see how it sustains the flow of capital and experience needed to improve the model. That makes me question whether a startup should focus exclusively on this direction.

This resembles a difficulty faced by companies building Level 4 robotaxi algorithms. They developed advanced driving systems, but without sufficient control of the vehicle platform or a strong deployment partnership, they were unable to obtain both commercial and data advantages, in the argument made here when robotaxis reached practical use.

Similarly, an embodied model detached from physical deployment struggles to establish a clear business-and-data loop. The central question is not simply whether the company owns a particular robot or targets one scenario. It is whether it has a clear route to iterate in real applications and collect further experience, thereby improving generality and generalization.

12. Where might embodied intelligence deploy?

There are many possibilities, but home cleaning is especially interesting to me. At CES 2025, Roborock and Dreame showed developments in robot vacuums with small arm-like mechanisms. I think their practical functions are still limited at this stage, but they offer a promising starting point: for the first time, widely adopted robot vacuums had the ability to interact with, alter and receive feedback from their surroundings, in the sense of object manipulation discussed here. Editorial clarification: Roborock's Saros Z70 demonstrated a grasping arm; arm terminology across vacuum products can also refer to extending brushes or mops, so the mechanisms should not all be treated as equivalent manipulation systems. The opportunity being discussed is expansion from floor cleaning into interaction with objects.

If vacuum makers continue improving arm control, recognition and grasping, the resulting interaction data from real homes could support more general and intelligent robots.

In home cleaning, I think existing vacuum makers with established positions and continued R&D and iteration capabilities could be the first competitive players. Consumer-technology companies such as Huawei and Xiaomi also have opportunities. If they enter, I doubt they would skip floor cleaning, a frequent household need with clear willingness to pay and familiar purchasing expectations. They could build broader cleaning and environmental-interaction capabilities on that base. New startups will have a chance only if they develop effective R&D, manufacturing and sales capabilities.

13. How do you view companion robots?

I think companion robots benefit more directly from advances in large language and multimodal models than from models for physical manipulation. Most do not need to alter the environment extensively. Their central value is natural, emotionally expressive interaction through text, speech, images and other channels. Some biomimetic or anthropomorphic movement can make that interaction more expressive and make the robot feel more human or animal-like.

Technically, most companion robots do not primarily aim to improve physical intervention or environmental feedback. At this stage, I see them as a product opportunity combining large models with consumer electronics. I find it hard to imagine deep physical manipulation as the fundamental requirement of that category.

14. Should investors back embodied-intelligence companies focused on specific scenarios?

My investment thinking is not very narrowly constrained or linear. If general-purpose robots become possible, their creators will need strong, comprehensive capabilities across software, hardware, supply chains and production. A product with such broad applications demands top-level competence in many areas, and a team cannot afford any weaknesses.

If that holy grail is so large and distant, extrapolating from today to ten or fifteen years ahead does not give a simple answer about exactly what a company should build now. The more important question is whether its current products develop the capabilities needed when general robots take off. I often tell friends that, in my present judgment, only a handful of companies worldwide may have the capacity to realize that goal. In the United States, Tesla is one candidate; in China, Huawei, Xiaomi, DJI, smart-EV companies such as NIO, XPeng and Li Auto, and manufacturers such as BYD have relevant strengths. I think reaching the goal requires the R&D and product capabilities of major technology companies. It is difficult to imagine a successful general robot without them.

What interests me more is this: given my expectation that general robots will arrive within the next decade, how does a founder use those ten years to build a company like DJI or Xiaomi and become a leader?

15. What is the ultimate role of general-purpose robots?

I believe that in ten or twenty years, the ultimate form of intelligent robots will necessarily encompass all the possibilities and capabilities we imagine today for artificial general intelligence. Like many researchers, I think an AI without physical interaction and feedback cannot develop the kind of general intelligence we seek. Training entirely in the digital world lacks important feedback and understanding from reality, leaving reasoning and cognition incomplete. In my view, general intelligence becomes possible only when AI can interact with its environment, receive feedback and learn for itself. I therefore believe robots will ultimately be the vehicle for a perfect union of AI and the physical world.

A robot in that future should be highly intelligent and able to adapt. It might be humanoid or take other forms with different mobility and manipulation capabilities. Whatever its shape, I believe it will become important in society: a companion or partner helping us address lasting questions about growth, meaning, loneliness and value. Ultimately, I believe robots will necessarily exist in some form and profoundly influence our society and lives.

Part III: Working methods and the new fund

16. How do you stay closely informed about frontier progress in robotics?

I do not always approach a question purely as an investor, nor am I simply trying to invest quickly. Curiosity about the field is the main reason. I want to meet researchers and encounter early signals from the frontier so that they inform my thinking and judgment.

Over a longer horizon, my philosophy of building a fund or starting a business is to find the exceptional founders of the next era. Great founders need to connect with a major trend or paradigm shift, as in the internet and mobile-internet eras. I think AI, robotics and embodied intelligence may offer that kind of wave to technology entrepreneurs over the next decade, so I need to keep following them.

Meeting founders early and staying in touch lets me observe their growth and change, which is central to early-stage investing. Understanding them earlier gives me a longer observation window, something I value greatly. Historically, I have almost never invested in someone I had known for less than twelve months. My decisions rest on sustained acquaintance and interaction, giving me a clear view of their ability, character and acceleration. That is closely connected to my philosophy of building a fund and starting a business.

17. Recommended reading

I like biographies, especially of technology founders or companies, but also of people who achieved exceptional things outside conventional technology entrepreneurship. I am often most interested in the opening chapters: how someone starting with little finds purpose, discovers a direction through hardship, earns their first meaningful money or receives their first positive signal in a career.

That phase reveals how a person develops and can offer deep insight. I hope to bring those observations into how we recognize and understand young founders, particularly first-time entrepreneurs.

Recently, I read a fascinating biography of Alexander Hamilton, one of America's Founding Fathers. He inspired me. Unlike some other founders of the country, he did not begin with a privileged family background. He was born outside marriage in difficult circumstances and rose through effort and talent to become important in the American Revolution and a close aide to George Washington. His life was full of challenges and breakthroughs, and I see qualities shared with exceptional entrepreneurs. Editorial note: the original says he lacked a good education; more precisely, his early circumstances were disadvantaged, but he later attended King's College, now Columbia, without graduating before the Revolution. The interview does not name the biography or its author.

Such founders often encounter setbacks or even misfortune early in life. Comparing Hamilton's upbringing with those of exceptional entrepreneurs such as Steve Jobs and Elon Musk, I see early encounters with adversity. I think these experiences helped them become unafraid of conflict, strongly confident and resilient from a young age. Once formed, those traits can persist throughout their lives and become crucial to success when they face greater conflicts and challenges.

18. A message from the new fund

My fund is called Alphaist Partners, combining 'alpha' with the suffix '-ist.' The idea is an organization that seeks and creates alpha. I emphasize it because Chinese entrepreneurship and VC benefited greatly from market beta over the last twenty years, especially the internet paradigm shift. In the next phase of technology entrepreneurship, I think the larger opportunity—and the harder challenge—is creating and capturing differentiated value within the broad beta of technological change.

I see founders who go beyond consensus and create fundamental value as having that alpha quality. I want us to build long relationships with them, becoming their earliest, most enduring and influential partners, not just investors. I also try to take the founder's perspective: if I were a twenty-seven-, twenty-eight- or thirty-year-old engineer starting a company today, what would I do? Which direction would I choose, and how would I design the path? Understanding their perspective better could make me a more lasting and useful partner, sharing in the returns from growing alongside them.

That purpose and my personal interest make me willing to research and invest in robotics over the long term. I believe China will produce more world-class technology companies like Huawei and DJI in the next decade. Younger founders have greater sophistication, broader perspectives and better training, and I believe they will certainly create technology companies even greater than the previous generation's.

My conversation with Peter left a deep impression. I admire his determination, ability and execution. As someone who also follows and invests in robotics, I am recording this interview as encouragement to work harder myself.

Smart Note · Industry Dialogues