| By Fu Chong

Editor｜Su Jianxun

Quick and pragmatic are the work styles of Liu Fang, the former head of Xiaomi's intelligent driving and founder of AGM Robotics. In September 2024, he founded AGM, and just one year later, the first batch of AGM robots has been deployed to customer

According to exclusive information from "Intelligent Emergence", since 2025, Amio has successively completed financing for seed rounds and angel rounds. Among them, the seed round was jointly led by Anke Innovation and Xinglian Capital, and Xinwangda

In the domestic market, there is no secret in technology, and what finally matters is how to continuously meet customer needs and form customer loyalty, Liu Fang told Smart Trends.

Liu Fang joined Xiaomi in its entrepreneurial stage in 2012. Over the past 13 years, he has been responsible for multiple sectors including mobile phone systems, AI hardware (such as the Xiao'ai speaker, translation devices, and in-car rear-view mirr

Liu Fang also brings this business mindset to his entrepreneurial projects, such as the core concept of robot deployment scenarios.

Before the establishment of Amiotai, Liu Fang had conducted extensive market research. He found that when analyzing in detail, three criteria are important: there is a clear demand; AI technology can bring significant improvement; and there is a clea

According to Liu Fang, embodied intelligence in industrial scenarios is not intended to replace automation, but rather to meet the demand where human labor costs are too high or where automation is difficult to implement effectively.

Although Chinese workers have a relatively low monthly salary, it is difficult to recruit them and the turnover rate is high, resulting in a significant labor gap. Meanwhile, with the increase of small and rapidly-iterating orders on factory assembly

Ami's plan is to take on the manufacturing processes of sorting, assembly, and testing in the 3C industry, which are not well done in traditional assembly lines and have high labor costs.

The account is calculated like this: in factories along the southeast coast, a worker earns a monthly salary of six to seven thousand yuan, with annual total costs usually at 80,000 to 100,000 yuan. With three shifts working in a day, the cost of a w

Based on this math problem, Liu Fang estimated the single-unit price of the robot to be around 200,000 yuan. According to his understanding, if the payback period of the robot can be controlled within one to one and a half years, customers will consi

Liu Fang, born in the post-80s, has a deeper understanding of the competition between giants and start-ups.

Some people believed that in the end, large factories equipped with AI will take over all market shares. However, Liu Fang believed that since factory business does not have high net profit margins, tech giants might not be attracted to it, leaving r

Looking ahead, Liu Fang believes that the geopolitical relocation of production capacity from China's factories also brings future opportunities for the going-global of domestically produced embodied intelligence.

Recently, The Emerge of Intelligence interviewed Liu Fang about his knowledge of searching for embodied intelligent scenarios, as well as his views on future technology and industry development. He shared his observations with us. The following is a

△Amio founder Liu Fang, photo: provided by the interviewer

We've considered entering the ToC market, but we can't see a clear demand.

Why have you chosen the manufacturing industry for your embodied entrepreneurship in 《The Emergence of Intelligence》?

Before starting my own business last year, Liu Fang spent a lot of time conducting market research. After reviewing everything, he realized that he had to find a scenario with a clear demand and a calculable ROI (return on investment).

Frankly speaking, everyone wants to do To C business, which sounds attractive with promising imagination. However, after research, we found it is not achievable at present due to technology and cost constraints.

For example, in the context of household services, Chinese families often have elderly members and part-time workers are not expensive. In addition, I always believe that beyond technology, in order to make users truly accept robots in human-machine

For example, during my trip to Japan, I found that although Japan's high-end service industry is well-developed, in scenarios where efficiency and ROI are more sought after, Japanese restaurants would adopt automatic ordering machines instead of manu

If we turn the lens to industry, the logic becomes very clear. The labor cost of a job position is about RMB 100,000 per year, and operating in three shifts will cost RMB 300,000. If we can sell a robot with a price of around RMB 200,000, which can r

Emerging Intelligence: Why is 3C Electronic Manufacturing Selected for Multiple Industrial Scenarios? Liu Fang: We have several core criteria for selecting scenarios: 1) clear demand; 2) significant improvement brought by AI technology; 3) clear retu

The 3C manufacturing industry is labor-intensive, with tens of thousands of workers in common factories; workstations are concentrated, facilitating deployment. Moreover, labor costs account for 12%-15% in this scenario, which is a significant expens

Intelligent emergence: You believe that the application of embodied intelligence in factory production-related scenarios will achieve daily repetitive tasks by around 2026, while complex tasks will take another year or two. How did you calculate this

It is mainly calculated based on data accumulation and the time required for deployment, debugging, and adaptation.

I believe that a total amount of data reaching tens of thousands of hours can enable robots to perform complex tasks such as flexible assembly.

This year, our data collection plan can reach thousands of hours, and our goal by the end of next year is about tens of thousands of hours. This amount of data is the foundation for training embodied intelligence for tasks such as rigid assembly.

△Amio Robot, Image: Provided by the interviewer

Be a quick learner, not a jack-of-all-trades

What are your thoughts on the technology barriers of embodied intelligence in terms of its emergence of intelligence?

In China, there are no secrets in technology. Ultimately, it's about how we can solve real customer problems and establish long-term associations with corresponding customer scenarios. Our barrier lies in making the right decisions as quickly as poss

We prefer robots to be fast learners rather than general robots that can do everything right out of the factory. Our core is to enable robots to learn quickly in one station.

We established the data strategy focusing on first-person perspective videos very early. In simple terms, it means allowing workers to wear cameras while working, and robots learn from these videos to imitate human operations. This approach minimizes

In Amiro, the ratio of video data to real machine data is approximately 6:1, with video comprising the majority of our training data, supplemented with a small amount of real machine data for calibration and fine-tuning.

It's like a new employee can quickly take over a position by watching videos of experienced teachers' operations and doing a little hands-on practice. Our goal now is to reduce the deployment time of a new workstation from several months to within a

Smart emergence: so it's not about pursuing comprehensive robot capabilities at the beginning?

Liu Fang believes in the power of data revolution, but not blindly pursuing big data. We are more confident in the strength of the revolution in action.

In the vertical field, embodied intelligence requires expertise and quickness on specific issues, to quickly master a task and learn from practice.

When a robot learns multiple job functions, it can develop relatively generalizable capabilities within a factory environment.

Intelligence emergence: how is the brain of AlphaGo and model training achieved? In terms of reinforcement learning, what specific thoughts does AlphaGo have?

For major decisions and path planning, we follow the VLA paradigm, while for fine-grained strategies, we adopt reinforcement learning for the final touch.

It is worth mentioning that our reinforcement learning focuses on "real machine reinforcement learning" rather than reinforcement learning in a simulated environment, because simulation cannot accurately replicate the feedback of force and all the de

Real machine reinforcement is used to solve two problems: fine grasping and assembly in the last few millimeters, and autonomous error correction when anomalies occur.

In simulation data, embodied intelligence can roughly understand how actions are executed, but it is only through actual operation that we can clearly see whether the actions are properly performed. Therefore, allowing robots to perform real-machine

Smart emergence: Some people believe that the implementation of embodied intelligence in factory scenarios has difficulties such as data security (difficult to transmit customer data back) and lower fault tolerance. How do you address these challenge

Liu Fang said: First of all, most of the factories we are currently cooperating with have a workforce of around 10,000 to 20,000. They can provide a significant amount of worker data and are more willing to cooperate than many large factories. In oth

If the basic model is well-trained in the future and we receive orders from companies involving data confidentiality, we can then collect and supplement data specific to the factory.

It should also be noted that the fault tolerance rate is similar in all embodied intelligent work scenarios, and it is not the case that only factories have a low fault tolerance rate.

For example, a milk tea shop always operates in a fast-paced environment, which could be faster than a factory. However, factories have fixed workspaces with similar environmental structures, making them more friendly for the entry of embodied intell

Emerging Intelligence: How to View More Advanced Technologies Such as Haptic Sensing and World Models

Liu Fang said that embodied intelligence is already a frontier technology trend, and at this stage, everyone is looking for stable and reliable technologies as well as engineering solutions.

The evolution of embodied intelligence itself is a development of cutting-edge technology. Now, VLA, world models, and multi-modal sensing are all directions we are focusing on. In addition to the pursuit of advanced technology itself, we also attach

How is the current order and commercialization progress of AI emergence?

Liu Fang: We currently have three key account customers, and our robot products have been running on their factory lines for some time. They are also considering increasing their purchasing volume. Overall, the progress is faster than we had previous

△Amio Robot, Photo: Provided by the interviewer

Will the physical retail bubble pop next year? Focus on the commercialization development of the industry

Emerging Intelligence: Why Choose to Start a Business Now and Enter the Embodied Intelligence Field?

Liu Fang said that both robots and automobiles are traditional industries with many years of foundation and have been developing linearly in the past. However, the injection of AI has given intelligent driving exponential growth potential, and I beli

In Dachang, although the technical trends are clearly visible, the company's decision-making path may be more conservative. This may miss opportunities for those who have ambitions in the industry. Therefore, I chose to start my own business.

Intelligence emerged: Why did the company name "Amio"?

Liu Fang: The name comes from the Spanish word 'Amigos' (friends). We are a robotics company, and our ultimate vision is to make robots partners of humans. The homophony of 'Amiuo' sounds better.

As an experienced player in the field of embodied intelligence, what do you feel about the emergence of many young founders in the industry who are born in the 90s or even later in the 95s?

Liu Fang: (laugh) Yes, I am indeed on the older side, but that is not a disadvantage. Having experienced several technology waves, I have a deeper understanding and awe of industrial changes.

My observation is that in every technological wave, the companies that can really survive are founded by people who have accumulated experience in the previous technological wave. In complex fields that require industrial chain collaboration, the fou

In the recruitment of technicians, we will employ young people who can directly learn the latest technologies and form a mutual cooperation with them.

Smart emergence: Does Amigo have plans to go abroad?

Liu Fang: Yes.

The motivation comes from two aspects: one is that our clients have the need to transfer production capacity, and the other is that the ROI of overseas markets is better.

For example, a Hungarian customer mentioned previously that they would consider purchasing an embodied intelligent robot priced up to 150,000 euros, which is approximately equivalent to a unit price of 1.5 million Chinese yuan.

As the largest manufacturing country, China has the most diverse demands and scenarios. Our strategy is to hone our technology, product, and service capabilities in China, and initially accompany Chinese customers as they expand overseas, and then gr

What is your view on the "bubble" in the embodied intelligence industry? What are the key indicators to predict the industry trend next year?

Liu Fang said that embodied intelligence is an emerging industry that is developing rapidly and that capital has high expectations for it, which may lead to some bubbles. To evaluate the industry, we can look at the situation of the top companies in

But the industry will always move forward, and we are confident about this. The industry will also deposit valuable things. Rather than pursuing high valuation, we prefer our company to rely on healthy business cash flow to operate smoothly.

The trend of next year's market mainly depends on the progress of commercialization in the industry.

Cover Source | Official Enterprise