The tycoon with a 100-billion-yuan net worth takes action, and Wang Qibin raises hundreds of millions of US dollars in a new round of financing.

It is exclusively learned from ChinaVenture that Lingchu Intelligence has completed a new round of financing of several hundred million US dollars, making it one of the 10-billion-yuan unicorns in the embodied brain track. The main capital of this round comes from the real industry side. Three leading manufacturing giants, Top Group, Ruicheng Fund under Chery Holdings, and Lens Technology, jointly participated in this round, with the support of institutions such as 37 Interactive Entertainment, Wuhu Investment Holding Group, and Fosun Chuangfu, while the old shareholder Huajin Capital continued to make excess additional investment.

In early February, I had a conversation with Wang Qibin. Half a year has passed in a blink, and this company has obtained financing from three leading manufacturing giants including Lens Technology. Zhou Qunfei, the 100-billion-yuan wealthiest entrepreneur, has also made investments, and has placed investments in more than one enterprise in this track. As an Apple supplier, Lens Technology itself also has a robot business layout, which fully proves once again how high the capital attention of the embodied intelligence industry is and how many practitioners are eager to get a share of the market.

Specifically, the three enterprises of Top Group, Ruicheng Fund under Chery Holdings, and Lens Technology cover three high-end manufacturing fields respectively: complete automobile manufacturing, core robot actuators, and 3C precision processing. They are not only the core demand side for intelligent manufacturing upgrading, but also the direct holders of industrial scenario resources. This year, the embodied intelligence industry has collectively entered the phased commercial implementation stage, and Lingchu has also achieved small-scale commercial verification in logistics scenarios. With the injection of industrial resources in this round, the cycle from human data supply, model iteration, to scenario implementation and data return will be further accelerated.

In previous articles, I introduced that Lingchu Intelligence is an embodied intelligence enterprise positioned as a "small full-stack" company. The reason why it is called "small full-stack" is that it has made trade-offs in R&D, focusing on building a software and data acquisition toolchain system with the end-to-end WAM model as the core. At present, Lingchu adopts a dual-system architecture in terms of models, with the Psi-R2 model paired with the Psi-W0 action-conditioned world model, which greatly improves the operation robustness and generalization ability in unstructured scenarios.

Data is a key link in model training. High-quality human operation data is the core fuel of embodied intelligence, and it is also a common bottleneck for the global industry at present. Because it trains models by itself, Lingchu knows exactly what high-quality data it needs. Based on this logic, Lingchu has independently developed a set of data acquisition systems to accumulate human operation data, and has completed data precipitation of 100,000-hour magnitude. Wang Qibin revealed that the company will hit the target of collecting 1 million hours of human data within the year, and will iterate two new models in the second half of the year.

Around this round of financing, the company's new progress and changes in the industry's development trend, I had a conversation with Wang Qibin. After the almost crazy financing boom in the first half of the year, he believes that the embodied intelligence industry has entered the deep water zone, and all enterprises are shifting from demo to improving model pipeline and scenario implementation. Wang Qibin is an industry veteran with more than 20 years of product and commercialization experience in the robot and consumer electronics industries. Having gone through industry cycles, he is sensitive to the ups and downs of the market and speaks openly, so the following short conversation is worth reading.

The following is the content of the conversation, slightly edited:

Embodied intelligence enterprises have not yet formed a clear gap in competitiveness

ChinaVenture: The last round of financing was led by state-owned capital. Why are the core participants of this round of financing mostly industrial parties? How did the three parties of Top Group, Chery and Lens Technology finalize the investment?

Wang Qibin: In fact, Lingchu already had industrial shareholders before, namely Zhongji Xuchuang and Yangtze Optical Fibre. This time we mapped all industrial parties to select the partners we value most. Now we believe that three major industries, namely automobile, 3C and optical industry, will be the absolute advantageous industries in China. Later we talked with several investors and invited them to invest in the company.

ChinaVenture: Does this also indicate that at the current stage of embodied intelligence, ordinary financial investors are no longer able to support such large-scale investment?

Wang Qibin: To be honest, at this point in time this year, we can feel that some financial investors are under pressure. But as an enterprise that needs continuous development, we do need to consider why industrial investment is very important at present. Lens Technology and Top Group have overwhelming advantages in the standards of the entire vehicle factories, the service of T1 suppliers to OEMs, and mass production advantages. Now we have products that can be OEMed and carry out in-depth mass production R&D in the two enterprises respectively. In turn, Lens Technology will open up some scenarios this year, and we are also cooperating with Top Group to develop some scenarios, and both sides have good cooperation points.

ChinaVenture: The last time we talked was in February. It has been almost half a year since February this year. What do you think are the particularly big changes in the whole market?

Wang Qibin: I think a lot of greater changes have taken place in the part from data to models in the past six months. The first is the problem of data format. From the terminal of data collection, to the construction of data pipeline, and then to model training, several rounds of cycles have been completed respectively. The main format of data is also gradually converging to some consensus. For example, ego-centric binocular and monocular data are now expanding rapidly; when we develop dexterous hands, the hardware is also iterating rapidly. Several leading enterprises should have built data pipelines that can support large-scale data output, and start to use such data to train models step by step. Leading foreign enterprises are also constantly releasing new models, shifting their focus from the generalization of models to the success rate of specific tasks. So the overall development is quite fast.

ChinaVenture: Is your current main business still the dexterous hand algorithm and data glove?

Wang Qibin: Yes. We have been collecting ego-centric video data, including monocular data, and we are also working on binocular data now. In terms of hand-related data, we are also iterating the glove, the new version is lighter, fits the hand better, and collects data more accurately.

ChinaVenture: So according to what you said, the current data collection route of the whole industry is converging?

Wang Qibin: The data collection route is converging, and Ego data itself has become a very strong consensus. For parts other than Ego data, the hand-related data, some enterprises are working on UMI, while we are still working on human hand data. In these two routes, the technical solution for the hand part is currently undergoing rapid iteration. For gloves, we have exoskeleton products, and we are also working on parts related to electronic positioning.

ChinaVenture: What does your data pyramid probably look like?

Wang Qibin: We use a set of equipment to collect all data, including the ego-centric video data on the head, and the joint angle data on the hand. All data can be collected in one workstation project.

ChinaVenture: What upgrades have been made to the models? The world model is very popular this year. What adjustments have you made in terms of models? Are the two models you are developing now working in coordination?

Wang Qibin: We have newly trained a model, which is the World Action Model, a policy model. Its input is the whole vision plus the state of the robot, plus some text to generate output. But what is more important is the Simulator, which is mainly used to evaluate and optimize all the generated policies to judge which ones are high-quality. If a generated policy is used to perform some operations, the Simulator will check whether the future frames are accurate and whether the expected effect can be achieved after the policy is put in. If not, reinforcement learning optimization will be carried out in this model. The two models are connected in series for joint training.

ChinaVenture: There are many technical paths in the field of world model now. How do you make trade-offs among them?

Wang Qibin: The focus at the current stage is more on training the action modality. No matter what kind of large VLA and World Model people talk about, the essential problem to be solved is the generalization on the action dimension. On the basis of this generalization, we will further improve the success rate. Following this path, we trained the World Action Model, and then put it into the Action-conditioned World Model for optimization.

Li Feifei once divided the model into three core modules: Rendering, Planner and Simulator. We have actually developed the second and third modules, and connected them in series. Some enterprises may develop vision plus action, or these modules can be combined arbitrarily.

ChinaVenture: Back to the topic, do you think enterprises have really formed a clear gap in model technical capabilities now?

Wang Qibin: I don't think there is an essential gap at present. There are several such enterprises around the world, and I think there are about 5 enterprises in China that are really devoting time, data and computing power to develop models. Generally speaking, we are still in the first stage, all enterprises are making great efforts to develop core technologies, and there is no absolute gap between them, only different development focuses. For example, we are concentrating on developing the model of general-purpose dexterous hands, and do not involve industrial grippers.

ChinaVenture: So you are one of the five enterprises?

Wang Qibin: That's for sure.

ChinaVenture: A number of emerging world model enterprises have risen this year. They claim that they have not taken detours in technology, and their products sound similar to yours. So do you think Lingchu has a first-mover advantage?

Wang Qibin: I personally think we have a first-mover advantage. These first-mover enterprises have already had their own relatively mature data pipeline teams, and the data flywheel has at least started to operate on a small scale. These latecomer enterprises have very strong algorithm capabilities, but they will be slower to build engineering systems. In the end, to realize the operation of the data flywheel, many engineering teams need to be built as soon as possible. All enterprises need to complete the transformation from academic research to algorithm engineering.

Demo is only 1% of the whole journey

ChinaVenture: You said the embodied intelligence industry has entered the deep water zone. Does that mean it is time to hand over the phased answer sheet?

Wang Qibin: I don't think it's time to hand over the answer sheet. I was browsing YouTube yesterday and saw a former CEO of Waymo talking, he said demo is only 1% of the whole journey.

This year, it is very typical that all enterprises are shifting from making demos to developing model pipelines seriously, or starting to promote scenario implementation. In any case, looking from the top down, we need to check whether there is a complete data pipeline? Where does the data come from? Which operator provides the operation service? How much does the data efficiency decay? Going down further, the stability of data acquisition hardware, such as the binocular headband, the binocular data of some manufacturers will be distorted after two hours, and the accuracy will deviate; now many people use Pico to collect binocular data, but it will overheat and shut down after working for 4 hours generally, of course it is not designed to work continuously for 10 hours; then when it comes to the complete machine, there are many problems to consider, which requires a lot of time.

ChinaVenture: Those leading embodied brain enterprises need data to train models. In addition to accumulating data by themselves, will they also buy data from you?

Wang Qibin: Surely all enterprises are building such capabilities by themselves, including the capabilities of teleoperation, ego data acquisition, large-scale collection, stable hardware, and building their own data pipelines to train models.

ChinaVenture: Looking ahead, what are your business expectations for this year?

Wang Qibin: The core goal this year is to make the effect of the whole model reach the expected level. We will continue to release new models in the follow-up, hoping to improve the generalization and the success rate of some typical downstream tasks.

ChinaVenture: What changes do you think will happen to this industry by the end of the year?

Wang Qibin: It's hard to say. I think some enterprises may fall behind in the first half of next year, and some enterprises will face great operational risks.

ChinaVenture: Speaking of which, the financing of embodied intelligence was so crazy in the first half of this year. What do you think is the deep reason behind it?

Wang Qibin: I personally think there are two reasons. First, from the perspective of the track, last year everyone completely followed the venture capital logic of the primary market, thinking that this track is very promising. This year, this industry is positioned as one of the six future core industries, so the capital structure has changed accordingly.

Second, there is a new narrative logic this year: the world model concept has been put forward, and new talent teams have emerged. The combination of the two factors will promote the new technology main line and new teams to be superimposed with capital. Some traditional VCs that have invested in this track last year will make new investments in the world model track this year. The superposition of the two logics makes the market very hot in the first half of this year.

But I personally think that the market is gradually moving towards a more rational and calm direction. I think the real benign growth is bound to be like a heartbeat, with ups and downs. This track has been developed for more than two and a half years, and a rational correction is normal. All long-term tracks cannot keep rising all the time, they all rise spirally with adjustments. All enterprises focusing on production and research need to return to the iteration of their own products, and create value no matter from the model side or the application side.

ChinaVenture: Listing is also the main theme of the embodied intelligence industry this year, but if two or three dozen enterprises are listed on the Hong Kong Stock Exchange, can all of them achieve their expected goals?

Wang Qibin: I think there will be K-line differentiation. Just like the intelligent driving track in those days, the performance of different listed enterprises varies greatly. It is no longer a simple logic of a booming track. People will see that even if the track fluctuates, there will be obvious differences between different enterprises. Without strong technical strength and business model support, the listed enterprises will face great pressure after listing.

This article is from the WeChat official account "ChinaVenture", author: Liu Yanqiu, published with authorization from 36Kr.