news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Text｜Wang Qin

Edited by Qiu Xiaofen and Su Jianxun

Although the story of Ivy League students dropping out of school to start a business is not new, this time it is the turn of a post-00s generation, and a Chinese academic who has published in top journals and conferences and received a direct doctorate from Yale. The hot competition in the field of AI big models and embodied intelligence has given people a sense of urgency that "if they don't start a business now, they will be old."

Fred Yang, a native of Jiangsu Province, was born in 2000 and received a full scholarship to study at Yale University. He is the founder of the embodied intelligence company UniX AI, which has developed a household service robot that can wash clothes, make hamburgers, and wash dishes.

Fred Yang, founder of humanoid robotics company UniX AI

Although Yang Fengyu started his own business at such a young age, he did not drop out of his studies out of impulse. First of all, he wanted to start a business when he was in high school, and he had some "small-scale" entrepreneurial projects when he was an undergraduate. Secondly, he said that the number of papers he published was enough for a doctorate graduate, and dropping out of school was not a risky decision. Moreover, he has been accumulating entrepreneurial resources. Since his undergraduate studies, he has been accumulating academic contacts in the field of embodied intelligence in China. After starting his own business, he has been looking for senior people in the robotics industry in China to form a team.

Now, the UniX AI company he founded has invited Wang Hesheng, distinguished professor at Shanghai Jiao Tong University and chairman of the top robotics conference IROS2025, as its chief scientist.

Although he has published papers in top international journals and won the title of Outstanding Undergraduate Scientist of the Association for Computing Machinery, Fred Yang, as a young entrepreneur who has just returned to China, often encountered setbacks when recruiting people. When he first started building a team and invited some senior experts who were 20 years older than him to join, he had to visit them three times, and even talked for more than 8 hours at a time.

Currently, the wheeled + two-arm humanoid robot developed by UniX AI, as a home service robot, can automatically identify the clothes you throw around in various corners of the house, automatically pick them up and take them to the washing machine to wash, and can also collect plates and wash dishes for you after you finish your meal, and clean the floor.

UniX AI robot can pick up tofu

For home use, the robot's hand operation ability is critical. UniX AI's self-developed three-finger gripper can hold a ballpoint pen or tofu. UniX AI said that its first batch of 100 humanoid robots will be mass-produced in September.

The following is a conversation between Intelligence Emergence and Fred Yang, founder of embodied intelligence company UniX AI.

Intelligence Emergence: There is no unified definition of humanoid robots in the industry, and different companies have different technical paths. Some companies started out making bipedal humanoid robots, while others made humanoid-like robots, such as your company's wheeled + dual-armed robots. How do you view different technical paths?

Fred Yang：This market is large enough, and different companies are rooted in different scenarios. Everyone designs their own mechanical structure and overall software and hardware solutions based on the application scenario. At this stage, it is not difficult to find an application scenario for each technical solution. I think there is no right or wrong, and everyone considers it from the perspective of demand.

Intelligence Emergence: Why did you choose the wheel + double-arm configuration?

Fred Yang: First of all, let’s talk about the arms. In the home scene, the most important thing is the operation ability of the hands. The spatial height of many things in the home space is designed for people. Our bionic humanoid arm has 7 degrees of freedom (which can be understood as the arm has 7 joints) and can perform many human-like operations.

Regarding the wheeled model, considering the mobile accuracy and safety, we first chose the wheeled model to enter the home scene. Everyone should reverse design their own hardware solutions according to different usage scenarios.

UniX AI Robot's operational capabilities

Intelligence Emergence: Are you currently working on both hardware and software?

Fred Yang: Our core logic is to start from the scenario. Hardware and software are equally important to us.

Intelligence Emergence: Generally, companies working on embodied intelligence will have their own focus between the brain, cerebellum and the hardware itself. Do you have your own focus?

Fred Yang: We do both software and hardware, but if we talk about the focus among the brain, cerebellum and hardware, we focus more on the hardware and cerebellum levels.

Hardware is the foundation of the algorithm. The software and hardware are highly coupled, and the hardware must follow the scene. For example, in the home scene, the three-finger gripper we designed has two modes (three-finger mode and two-finger mode). The three fingers can be rotated into two fingers to hold small ballpoint pens or even tofu.

In addition, it is also important to lay a solid foundation for the cerebellum at this stage. Regarding the "generalizability" issue that everyone is concerned about, the generalization process can be divided into three stages: from limited scenarios of a single task, to open scenarios of a single task, and then to open scenarios of open tasks.

(Note: For example, from being able to do laundry in a single home environment, to being able to do laundry in a home environment in different spatial environments, to being able to do laundry, cook, tutor children, and other tasks in different spatial environments)

Now everyone in the industry is still working hard from the first stage to the second stage, which mainly tests the cerebellum. If we do open scenarios later (the third stage), it is essentially planning (task planning) at the brain level, but the first task now is to lay a solid foundation at the cerebellum level and solve the operational issues first.

UniX AI humanoid robot grabs clothes

Intelligence Emergence: One of your research results is the tactile multimodal large model UniTouch. What is visual touch and what is a tactile large model? What is the significance of visual touch in promoting the technology of humanoid robots?

Fred Yang: Touch is very important in the robot operating system, and even in the human operating system. From a human perspective, touch is the most instinctive perception modality, and feedback is obtained through real interaction with the physical world. For example, when looking for keys in a bag, generally speaking, people do not rely on vision, but mainly on touch.

From the perspective of the robot, due to the limitations of mechanical structure and sensor selection, vision alone is often not enough. For example, when a robot is asked to grab a bottle cap, the bottle cap is very small and when the robot arm grabs it, it is already blocked by the robot arm itself.

At this time, we can only rely on tactile feedback to complete the final grasping and verification. This situation is particularly prominent in the operation of deformable objects. When a deformable object is touched, the shape of the object changes, and the information that visual priors can provide is very limited. We must rely on very local but highly sensitive tactile information to perceive and complete the task.

At the same time, touch provides other information that vision cannot, such as force. Vision can provide the robot with the grasping position, but cannot tell the grasping force. Simple force sensors also have limitations. For example, before crushing a cup, no change can be seen from the perspective of force, even if there are tiny cracks on the wall of the cup. But at this time, the tactile sensor can capture the tiny cracks and determine whether the next operation will crush it.

I published a paper before, and created the world's first large tactile multimodal model, which blends tactile information with visual information and other language modal information. Each modality has its own limitations, and vision also has its own limitations, but after adding tactile information, each modality can complement each other.

Many other companies are also working on touch, but they are still focusing on the hardware level, such as tactile sensors. However, UniX AI is aimed at home users, so the hardware must have a lifespan of 3-5 years. At present, many high-precision tactile sensors cannot meet commercial needs in terms of lifespan.

UniX AI robot opens the washing machine

Intelligence Emergence: Your first batch of 100 humanoid robots will be put into mass production from September. This number is relatively large for a humanoid robot manufacturer. How does Unix AI do this? How is the supply chain managed?

Fred Yang: Mass production is mainly focused on the supply chain. We have some supply chain experts from Mercedes-Benz and Haier who have experience in supply chain manufacturing and cost control. We also have a group of supply chain members with rich mass production experience from traditional robotics, consumer electronics, automotive, and aerospace industries.

Intelligence Emergence: How do you build a team?

Fred Yang: The field of robotics cannot be solved by just one technology stack. It requires the overall coordination of hardware and software, and also requires a diverse team background. The entire embodied intelligence industry is very new, and our algorithm team is very young, basically composed of PhDs and postdoctoral fellows working in robotics or artificial intelligence at home and abroad.

In terms of hardware, we are currently focusing on home scenarios, and safety is definitely the most important factor. Our team has members from home service robots who are responsible for obstacle avoidance tasks in complex environments of different families, such as winding wires and suspended obstacles. For hardware related to embodied intelligence, including the arms and legs of humanoid robots, we also have highly capable scientists as the heads of hardware development.

Intelligence Emergence: You are very young, born in the 2000s. Such a large team requires strong operation capabilities. How do you recruit people?

Fred Yang: It was really difficult to recruit people at the beginning. I received my undergraduate degree from the University of Michigan and my doctorate from Yale. I am well-known in the embodied intelligence circle abroad, so I quickly found some friends who were working on algorithms abroad; but robots need to be a combination of software and hardware. When I returned to China, everyone was not familiar with us, and I encountered many obstacles. Before each important member joined, I talked to them for a very long time. For talents, we should have the spirit of "visiting the thatched cottage three times".

Emergence of Intelligence: It is not easy to implement humanoid robots. Currently, most of them are still in the stage of being sold to research institutes, and it is quite difficult to expand the scale. You said that you can go directly to C in one step, and it is consumer-grade, and you will mass-produce 100 units this year. How did you do it?

Fred Yang: The mass production of wheeled humanoid robots is not as difficult as that of bipedal humanoid robots. There are actually differences in the definition of humanoid robots. For bipedal humanoid robots, I think there is still a long way to go and they are not yet fully ready for the market because there are some safety issues that have not been resolved. But the most important thing in the home is the ability to operate the hands, so the first generation we launched is a wheeled + dual-arm robot, and the wheeled one is a more mature technology.

The speed of product iteration is very important. It is difficult to produce a high-quality product in the first generation. We first quickly produce a relatively stable, reliable, and cost-effective robot, which is widely available in the market, and then iterate quickly and repeatedly. We have always said that it takes three generations to produce a high-quality product.

Our initial 100 units are mainly seed users who are willing to be the “first to try something new”.

UniX AI robot cleans the table

Intelligence Emergence: Do you have TO C customers now? Did you have contact with them in the early stage?

Fred Yang: The initial individual users are mainly some of our friends, some friends in the technology industry, and technology enthusiasts who are very interested in cutting-edge products. They are similar to the target group willing to buy Tesla when it was first launched, and they are also the class that likes to try out new technologies. We will also do some focus group interviews to understand the specific needs of the target group.

Intelligence Emergence: General humanoid robots in home scenarios are much more difficult to implement than robots in vertical scenarios (such as commercial services, warehousing and logistics, security inspection, etc.), and require much higher generalization capabilities of robots. Your information says that the UniX AI humanoid robot has the functions of washing clothes, dining assistant, cleaning, and tutoring children's homework. How is the implementation?

Fred Yang: The ultimate goal of humanoid robots is, of course, to make them capable of doing everything, but meals must be eaten one bite at a time and things must be done one thing at a time. In the home scenario, it is also implemented step by step, from a limited scenario with a single task, to an open scenario with a single task, and finally to an open scenario with an open task.

Our products currently have several mature scenarios, such as washing clothes, helping to collect dishes after meals, and 3D cleaning functions. Take the laundry function as an example. You can throw clothes in various places. The robot Wanda will first find objects and build a three-dimensional map independently. The second step is to grab clothes. Wanda's UniX AI self-developed gripper can universally grab flexible objects. The third step is to use the washing machine.

It is not difficult to generalize the use of washing machines (the robot can operate different types of washing machines). We can call different small models to complete this task.

Intelligence Emergence: Why did you choose to return to China to start a business? Have you ever thought about starting a business in the United States?

Fred Yang: I don’t define myself as starting a business in China or in the United States. We have teams in both the United States and China, but we all take advantage of different locations. The United States has a top algorithm team, and China has colleagues in structure and hardware. We also have a team in Shenzhen and a research and development center in Shanghai, where many colleagues work.