Exclusive Interview with Wang He of Galaxy Universal: Robots That Can Truly "Get the Job Done" Aren't Afraid of Price Wars | 36 Emerging Figures

 Emergence is a key phenomenon in the wave of generative AI: when the model scale expands to a critical point, AI will exhibit human-like intelligence, being able to understand, learn, and even create.

"Emergence" is also happening in the real world - the silicon-based civilization is on the verge of breaking out. Entrepreneurs and creators in the AI field are using their wisdom and brains to light up the long journey towards achieving AGI.

At the moment of the alternation of old and new productive forces, "Intelligent Emergence" launches a new column called "36 People of Emergence". Through dialogues with key figures in the industry, we will record the new thinking at this stage.

Text by | Wang Fangyu

Edited by | Su Jianxun

"Apart from dancing and somersaulting, what else can humanoid robots do?"

Wang He, an assistant professor at Peking University, the founder, and CTO of Galaxy Universal Robotics, may be the most qualified person to answer the above "soul - searching question" from Zhu Xiaohu. Galaxy Universal, founded by him, is a unicorn in the domestic embodied intelligence track and is also the most focused on the "brain" of robots among the first - tier players in the industry.

Since its establishment in May 2023, Galaxy Universal has only released one robot body product, Galbot (G1), but has released multiple embodied large models. The company has invested most of its resources and funds in the research and development of embodied large models, aiming to improve the versatility and generalization ability of robots.

Wang He, the founder of Galaxy Universal       Image source: Enterprise authorization

In Dr. Wang He's view, the result of the competition in humanoid robot bodies is to sell robots at the price of steel raw materials, and there are already signs of a price war in the current market. Only by improving the capabilities of embodied intelligence models can humanoid robots be given higher value.

The general embodied large model is an "uncharted area" of human frontier technology. When talking about the current development of embodied models, Wang He, who shoulders this grand proposition, is unexpectedly "conservative" and pragmatic:

"I particularly don't recommend talking about embodied AGI. Many companies hope to achieve embodied AGI in one step, which I don't agree with."

"There are still many immature aspects of embodied intelligence models. It may take five to ten years to be able to do all kinds of jobs."

"A large number of scientific research results have emerged over such a long period of time, but products that can be mass - produced have never been launched."

Currently, many domestic embodied intelligence model manufacturers are keen on "showing off their muscles": demonstrating the generalization ability of their models through demos of complex operations such as "folding clothes, shaving, and zipping". Galaxy Universal, on the other hand, is focused on conquering the "less complex" skills of moving, grasping, and placing. The name of its embodied operation model is not at all "sexy" - the basic large grasping model, GraspVLA.

Wang He told "Intelligent Emergence" straightforwardly that Galaxy Universal is also researching the skill of hanging clothes with a hanger, but such complex operations are just a scientific research result and are still far from being launched and commercialized.

Currently, the skills of embodied intelligence models that are closest to real - world applications are the relatively "simple" Mobile, Pick and Place. Galaxy Universal is committed to applying the Mobile, Pick and Place skills in some scenarios such as pharmacies and retail first.

It is reported that Galaxy Universal, in cooperation with its partners, has launched the world's first intelligent retail solution for humanoid robots. Currently, nearly 10 pharmacies in Beijing are unattended 24/7 by robots. Galaxy Universal's humanoid robots automatically and accurately pick medicines continuously and deliver them to riders.

The company plans to open 100 unattended retail stores in cities such as Beijing, Shanghai, and Shenzhen this year. Currently, this application scenario has been commercialized, and it is expected to bring nearly 100 million yuan in revenue to Galaxy Universal this year.

At the opening ceremony of the recent Beijing Academy of Artificial Intelligence Conference, Galaxy Universal's robot Galbot conducted a live - streamed real - machine demonstration on the main forum stage. Under Wang He's voice command, the robot autonomously and accurately moved to the correct position, picked up a drink from the shelf, and achieved the full - process autonomous execution of complex shelf grasping and delivery without remote control and without pre - collecting scene data.

Galaxy Universal demonstrates grasping and delivery at the Beijing Academy of Artificial Intelligence Conference. Image source: Enterprise authorization

Wang He admitted that for embodied intelligence to enter any scenario, some data preparation is required to create a 100% successful product. The Mobile, Pick and Place skills are still being continuously updated. Galaxy Universal chooses to start from the retail shelf scenario and gradually improve the generalization ability of operations.

In Wang He's view, completely solving the generalization problem of "simple" operations like Mobile, Pick and Place is already an important milestone in the history of human embodied intelligence and robotics. According to his calculation, the maturity of this skill can open up a new market worth hundreds of billions of yuan, helping humans complete arduous tasks in multiple scenarios such as retail, front - warehouses, and car factory SPS sorting.

From the perspective of generalization, if a omnipotent humanoid robot is rated as 100, a robot that masters the Mobile, Pick and Place skills is rated as 10, and the implementation of Mobile, Pick and Place in the retail shelf scenario is only "1".

Currently, Galaxy Universal has achieved the breakthrough from "0 to 1" and is moving towards the ultimate goal of general embodied intelligence.

The following is a dialogue between "Intelligent Emergence" and Wang He, the founder and CTO of Galaxy Universal. The content has been slightly edited:

The market for welcome performances is just a flash in the pan. Robots should be trained to do high - value work

"Intelligent Emergence": What is the current size of your company's staff?

Wang He: We currently have more than a hundred employees.

"Intelligent Emergence": It seems to be fewer than that of your peers in the same tier.

Wang He: At this stage, we still focus on the R & D and production team. Currently, Galaxy Universal has launched a humanoid robot product, Galbot G1. Starting from the core needs of scenarios such as industry, retail, and services, its main skills are moving, grasping, and placing.

I believe that this skill can build a complete closed - loop Skill Set in various broad scenarios such as industry, business, and services, rather than developing many scattered small skills or a wide variety of full - category robot products, as this would lead to a much larger workforce.

"Intelligent Emergence": Galaxy Universal has only made one robot body but has released multiple models. Are you allocating more resources to the models?

Wang He: Actually, there are more people working on "hardware" in the company than those working on "software", which may be different from what the outside world imagines. The outside world may think that since Galaxy Universal only makes one product, it doesn't need many hardware engineers. In fact, our robot standards are different from those of many peers.

If a robot is only used for scientific research, as a hardware platform, or for a 5 - minute demo, this form of presentation doesn't require high reliability for the product. This is very different from the reliability of a robot that can truly work 24/7 because it can't be applied in real - world scenarios.

Galaxy Universal's hardware has undergone multiple rounds of intensive iterative upgrades around one product so that we can truly achieve 24/7 operation of robots in the unattended pharmacy scenario. If there is a hardware problem and an engineer needs to come on - site for repairs, the cost will be very high. Therefore, the original design intention of our product is to build humanoid robot products according to automotive - grade standards, or even higher.

"Intelligent Emergence": What about the capital investment?

Wang He: As an embodied large - model company, our largest investment is still in model research and development. However, this doesn't rely on simply increasing the number of people, because no company can build a good model by just piling up employees for model training. Instead, a complete closed - loop team from data infrastructure to model training and testing needs to be established. The computing power cost accounts for a large proportion. In fact, the number of top - notch geniuses in model - making is small in any company.

"Intelligent Emergence": Emphasizing synthetic data is a very distinct label of Galaxy Universal. However, many peers also say that they will use simulation data, combined with some Internet videos and real - machine data. What's the difference between you?

Wang He: Those who don't know how to use synthetic data well will say that simulation is "toxic" and so on. Synthetic data has played a very important role in Galaxy Universal's current achievements. Based on our self - developed synthetic data technology, the training cost of our embodied large model has been greatly reduced. At the same time, we also emphasize the integration of virtual and real, which enables our embodied large model to truly lead the world. This shows that we can really make good use of synthetic data.

For example, anyone can download Internet video data. A bit more challenging is remote operation. The embodied robots we currently deploy in supermarket and retail environments use real remote - operation data, but its proportion is much lower than that of simulated synthetic data.

The method of generating synthetic data requires manufacturers to have a good set of infrastructure including graphics, physical simulation, physical rendering, and automatic motion synthesis pipelines, as well as a verification closed - loop. These require long - term accumulation and core technical know - how. These accumulations are also the key reasons why Galaxy Universal can make better and more generalized models.

Galaxy Universal's Galbot robot   Image source: Enterprise authorization

"Intelligent Emergence": Your robot body has a wheeled chassis. Can it be understood that Galaxy Universal pays more attention to developing the upper - limb operation ability of robots?

Wang He: It depends on which aspect we focus on. On the product side, we are guided by the actual application requirements.

Currently, most customers, considering their own enterprise needs, such as doing some moving, grasping, and placing work in factories and supermarket retail scenarios, require a chassis - type robot. Bipedal robots are noisy and have short battery life. Our wheeled - chassis robot only needs to be charged every 6 - 8 hours, which has a natural advantage over bipedal robots.

From the perspective of R & D, Galaxy Universal has a full - stack layout for embodied intelligence and also has a layout for bipedal humanoid robots. However, at this stage, they are not products that can be widely applied on the product side.

"Intelligent Emergence": The market for welcome performances has emerged this year, and many peers are vigorously entering this field. Why hasn't Galaxy Universal seized this market?

Wang He: In my opinion, these flashy scenarios are just a flash in the pan. The market will ultimately not be won by a wave of traffic. What really matters is a good user experience.

Galaxy Universal always attaches great importance to the user experience. For example, there are many welcome robots in lobbies, but they mainly do some insignificant work. What we are doing is a next - generation reception robot product that can make customers willing to use it and truly assist people in work. As long as we can achieve this, I believe there will be a vast market for us.

So, it's not that we won't do it. Instead, we are already in the process of planning. Currently, we are in the stage of integrating many technical points into a line and then into a surface.

The market space for "Moving, Grasping, and Placing" is large, but the technology is not fully mature

"Intelligent Emergence": Do investors put a lot of pressure on your company for commercialization?

Wang He: Investors have given us great support, not only in financial investment but also in strategic cooperation resources. Currently, we have solid implementation results, and the company should have an impressive revenue scale this year.

"Intelligent Emergence": What about the education and scientific research markets? Have you made any plans for them?

Wang He: I think it's a matter of different understandings of priorities. How profitable is the education market? What is its ceiling in terms of the number of products? In fact, many bipedal robot companies have joined the competition for the education market. Galaxy Universal will focus on the fields where it has advantages. Starting from the perspective of demand, we aim to make robots truly meet the existing pain points in the market.

Galaxy Universal doesn't focus on selling humanoid robot bodies like selling steel raw materials. If the competition in humanoid robots continues like this, in the future, everyone will tend to price according to the material cost. What we expect is that humanoid robots can do valuable work and that embodied intelligence can truly create intelligent value.

"Intelligent Emergence": Do you see the trend of a price war in the current market?

Wang He: Yes, the prices are dropping crazily now. They have dropped to a few tens of thousands of yuan, and someone may offer an even lower price in the future. In fact, we are glad to see the entire industry rapidly reducing hardware costs through rapid hardware iteration. The cost - reduction of the supply chain is also beneficial to Galaxy Universal.

The problem is what problems the robots at this price can actually solve. What we are currently focusing on is high - value work. Our robots are sold at a price of several hundred thousand yuan, and customers are still very willing to use them because they significantly relieve the labor cost pressure of employees working in three shifts. This is also the reason why we (expect) to achieve hundreds of millions of yuan in revenue.

"Intelligent Emergence": Why can customers still accept the price of several hundred thousand yuan for your robots?

Wang He: As I mentioned above, the psychological expectations of users for our robots used in real - world scenarios are different from those for the cheap humanoid robots sold by other companies. The requirements for product maturity and reliability are different.

Galaxy Universal's robots can work continuously for a month without a single error, which is our core competitive advantage. I call our robots "real - world application robots", while those used for scientific research and mall performances in the market are called "R & D platform - type robots".

"Intelligent Emergence": Your main skills revolve around moving, grasping, and placing. However, some people think that these so - called "PPT operations" (Pick, Place, and Transfer) can solve very limited practical problems and are applicable to very limited scenarios.