Top-tier VC has made four consecutive rounds of investment, and Liu Songming has raised hundreds of millions of yuan in new financing.

LiberAI, which I reported on in June, announced today that it has completed a new round of Pre-A+ financing worth hundreds of millions of yuan. The investors include 360 Strategic Investment, China Development Bank Capital, Sinovation Ventures, CMC Capital, Binfu Capital, etc. HSG, an old shareholder, has made successive additional investments in four rounds since the angel round. Within three months, LiberAI has received continuous bets from top-tier institutions such as HSG, ZhenFund, Meituan Dragon Ball, and Shunwei Capital.

In the first half of the year, new companies in the world model track emerged in endlessly, creating a scene where "dazzling varieties are gradually overwhelming the senses". Entering the second half of the year, capital will converge on leading projects. It is no longer feasible to only present conceptual stories, and solid business progress is required.

Last time, the article Exclusive | Top VCs Have Invested Three Consecutive Rounds, Tsinghua Special Award Winner Born After 2000 Secured Another Hundreds of Millions of Financing introduced Liu Songming's entrepreneurial journey and LiberAI's choice of technical route — the team placed heavy bets on native multimodal pre-training in the model, and built its own data collection and hardware system at the same time. This time, I focused on discussing with him the importance of pre-training, human data and dexterous hands, as well as the progress of self-developed human data gloves.

This post-2000 founder believes that pre-training is the root of intelligence. The world model is essentially an extension of the multimodal model, and the key lies in how to implement the multimodal methodology after introducing the physical modality. Specifically for human data, he believes that the three modalities of text, video and physics must be aligned in the latent space to solve the sampling and decision-making problems of dexterous hands in high-dimensional space.

When talking about progress and plans, Liu Songming revealed that LiberAI has fully run through the entire process from model training to deployment on high-degree-of-freedom dexterous hands, and its self-developed human data gloves have also entered the mass production stage. By the end of this year, the company will focus on promoting two things:

The first is to fully run through the human data process, derive the pre-training Scaling Law for human data, achieve emergence in generalization, so as to no longer rely on simulation and real machine data, and make human data the most core pillar.

"For grippers, you can get a pretty good demo within the first 100 or 1000 hours, and every step is usable. When we develop dexterous hands, there will be a 'silent period' before we reproduce human intelligence, and the risk is indeed higher. But once we get through this stage, we can connect all human data and internet videos, and the upside is huge." This is Liu Songming's bet.

The second is that in terms of model architecture, LiberAI aims to build a world model with agent capabilities. This is no longer a model for single-point tasks, but a model capable of reasoning, just like the "physical version of Claude", which can perform task planning, reasoning, execution and error correction in the real world.

Gripper or Dexterous Hand

The embodied industry has been developing for two or three years, but divergences and unconsensus views on data collection methods and model technical routes have always existed.

Generalist AI just announced a new capability of its model in July — the same pre-trained model can drive a wide range of end effectors, including not only common grippers, but also special tools such as electric screwdrivers and egg beaters. Its cross-embodiment capability is amazing, but it clearly stated that it does not favor the five-finger dexterous hand route, arguing that robot hands will not look like human hands.

"The five-finger hand is just one tool in the toolbox. Limiting robots to this form is a failure of imagination." Generalist AI expressed its distinct attitude in a blog post.

Liu Songming holds a completely different view on the choice of end effectors. He believes that as long as human intelligence is fully distilled, this intelligence will naturally have the ability to adapt to different embodiments. Contrary to Generalist AI, LiberAI has never developed any grippers, but focuses on dexterous hand algorithms and the collection and pre-training of human data.

During data collection, whether to "let humans imitate robots" or "let robots imitate humans" is a philosophical-level question in Liu Songming's view. The former corresponds to collecting data through teleoperation, UMI grippers and exoskeletons, while the latter corresponds to collecting human data with data collection gloves.

"One idea is to let humans imitate robots, believing that the data collected in this way can be transferred from humans to robots. But our idea is the opposite: let humans retain their natural characteristics as much as possible, and pursue non-intrusive data collection. The cost of this approach is that the gap between humans and robots may widen in the short term, but the advantages are a higher ceiling, lower operation and maintenance costs, and easier scaling up." Liu Songming analyzed the trade-offs in this choice.

Of course, from the perspective of algorithm implementation, grippers and dexterous hands are not at the same level of difficulty at all. For example, a gripper has only one degree of freedom, and its state space is discrete 0 and 1, with simple control and low algorithm difficulty, so a demo can be produced in a short period of time. The dexterous hand deployed by LiberAI has 20 degrees of freedom, which is more difficult in observation, data sampling, underlying infrastructure and control of some details.

The R&D of models deployable on high-degree-of-freedom dexterous hands is widely recognized as a hard nut to crack, but it is very meaningful in Liu Songming's eyes. "It's like climbing a snow mountain. If you accept this level of difficulty, once you reach the summit, you will have huge advantages and can connect to all human intelligence. We have verified the gripper route at the earliest stage, and it is very difficult to generalize from grippers to dexterous hands, but our model can easily transfer from dexterous hands to grippers. In the end, we are more convinced of the human data and dexterous hand route we have chosen."

The logic supporting this judgment is: since the ultimate goal is to enable robots to have human-level general operation capabilities, human data is irreplaceable "oil". Human data is naturally generated in dexterous operations. No matter how much gripper data there is, it cannot be aligned with human videos, and it is difficult to scale up to general intelligence.

The "Impossible Trinity" and Data Pyramid

To achieve this goal, LiberAI has mainly carried out work in three aspects:

The first is data infrastructure and engineering algorithms, including the construction of the entire data pipeline — how to clean data, cut internet videos frame by frame, detect hands from the videos and optimize the speed of the hand detection operator, so as to quickly segment hand data. The entire link must run fast, and the segmentation boundaries must be clear. For example, in the action of drinking water, when to grab the cup and when to unscrew the bottle cap must be accurately identified and segmented from the video.

At present, this set of data infrastructure and data pipeline has been fully systematized, with an automation rate of over 90%. Automatic labeling is basically completed through APIs and self-developed fine-tuned models, and the pipeline itself is monitored and optimized by agents, which will become smarter with more labeling tasks.

The second is the training of the model itself. The key is to improve the model's capability to enable it to sample efficiently in complex spaces. The three modalities of text, video, and dexterous hand motion need to be aligned to the same latent space for sampling, so as to improve the efficiency of hand perception, sampling and control. The LiberAI team has completed a large amount of human data and multimodal training, and aligned them to a unified space.

To make this route work, it is also necessary to overcome the "impossible trinity" of data collection gloves: high precision, wearability, electromagnetic interference resistance and occlusion resistance. Liu Songming introduced that LiberAI's self-developed human data gloves have overcome the above impossible trinity, and have entered the stage of mass production and large-scale in-the-wild data collection.

Every embodied company has its own data recipe. Now the data types have generally converged to several categories such as internet videos, real machine teleoperation, UMI, and human data, but the proportions of different types of data still vary among different companies.

In LiberAI's data pyramid, first-person and third-person internet human video data occupy the absolute base of the tower, which is 1 to 2 orders of magnitude larger than glove data; glove data comes next, which is 1 to 2 orders of magnitude larger than real machine data; real machine data is the least. The company's vision is that human data can become the most important pillar, and there is no need to rely heavily on real machine data which has high cost and limited scale.

This seems like a rather radical bet. It should be noted that the mainstream view in the industry last year was still to collect real machine data through teleoperation.

The company is also trying to deeply integrate its data collection team with the agent system. For example, when model reasoning finds that "the model performs well in folding clothes but poorly in garbage collection", the agent will automatically issue tasks to schedule more people to collect garbage-related data.

Using AI to manage AI's data sounds very cool, which is a brand new organizational idea in the AI for AI era.

Apart from dexterous hands and data collection, we also talked about industry competition and corporate culture. The following is part of our conversation.

Liu Yanqiu: Dexterous hands are so difficult that many people in the industry choose to give up five-finger designs and develop two-finger grippers instead. Why do you insist on developing algorithms for five-finger dexterous hands?

Liu Songming: This is a problem about short-term and long-term trade-offs. For grippers, you can get a pretty good demo within the first 100 or 1000 hours, and every step is usable. When we develop dexterous hands, there will be a "silent period" before we reproduce human intelligence, and the risk is indeed higher. But once we get through this stage, we can connect all human data and internet videos, and the upside is huge.

From the perspective of first principles, we prefer to pursue things that are difficult but correct.

Liu Yanqiu: What do you mean by "kill real machines for subsequent training"?

Liu Songming: As we mentioned earlier, on the one hand, we have sufficient human intelligence; on the other hand, by self-developing the ontology, we can design it to be closer to humans themselves. The goal is to eliminate the need to collect real machine data and directly complete migration, so as to fundamentally solve the problem of high cost and difficulty in scaling real machine data. At present, our humanoid ontology has started mass production, and the hardware OS system is developed in an agent-based manner. In the past, when the hardware broke down, people needed to troubleshoot it, but now agents can help with this work, which can automatically locate the damage and even provide customer service, making the use more convenient. We now connect all these processes to agents. As they encounter more and more various errors and interactions, they will become smarter and smarter.

Liu Yanqiu: This idea sounds very AI-native.

Liu Songming: Yes, we are also using new ideas, using AI ideas to develop hardware.

The data pipeline mentioned earlier is an agent itself, which is like a living body that monitors what problems occur in the pipeline. Since our automation rate is 90%, there will always be some scenarios that agents cannot cover. It can summarize the uncovered problems on its own, and then iterate its own models and algorithms. I think this is the idea of AI for AI, making it a self-improvement system.

Including the data collection link, on the one hand we cooperate with outsourcing vendors, and on the other hand the entire management system is agent-based. Task issuance, verification, simple communication and customer service are all handled by agents. After the data is collected, it is used to train the model, and the model will perform reasoning. The agent will observe the overall effect of model reasoning, and then modify the issued tasks accordingly. For example, if it is found that the model performs very well in folding clothes but poorly in garbage collection, the agent will issue tasks to let more people collect garbage-related data.

Liu Yanqiu: As a new generation of embodied companies, how do you view earlier established companies? If new technologies and data collection routes emerge, earlier established embodied companies may not have first-mover advantages?

Liu Songming: Suppose you have collected 50,000 or 100,000 hours of gripper data, and now you want to develop dexterous hands, what will you do with these data? For another example, if you go to households to collect teleoperation data, you need to build an operation system, schedule delivery vehicles, and arrange maintenance engineers to follow up. If you stop collecting teleoperation data now, do you still need to keep that system? There are indeed some practical problems.

Many larger companies run multiple routes at the same time, and it is difficult to bet on a single one, which is precisely a sign of lack of confidence. Startups must focus and iterate quickly. The founder only has 24 hours a day. If they start multiple routes at the same time, they may not be able to catch up with others in any of them.

Liu Yanqiu: Aren't you worried that if you go all in on this one route, you might bet wrong?

Liu Songming: First, this is derived from first principles. My previous research on the pre-training paper of the embodied foundation model and the verification of the scaling effect of no-ontology data have proved our capabilities, and I have also seen the limitations of the previous routes. Second, we are continuously verifying step by step, and we are more convinced of the path we have chosen in the process. I have insisted on betting on this path since the middle of last year, and now it is gradually getting rid of unconsensus views.

Liu Yanqiu: But after all, first movers have accumulated technology and data, can't these accumulations be fully integrated into the new path?

Liu Songming: The unity of knowledge and action is very critical. Strong historical path dependence may lead to a mismatch in recruited personnel. Many companies recruit people from other fields, but we find that the most suitable people are either very young embodied researchers or people who have experience in video and multimodal research. When the organization grows larger, mismatched backgrounds will become a burden that affects the iteration speed.

The main task of a number of first movers at present may be to sprint for IPO, and their goals from the first principles perspective are inconsistent with ours. Our bet is to sprint for the arrival of AGI, which is more pure, more focused, more agile, iterates faster, and has a steeper growth slope.

Liu Yanqiu: You bet that there will be a breakthrough within one or two years. What is the basis for that?

Liu Songming: The basis is that we think we have already solved the data problem, and the development of large models will be very fast in the future. In principle, this problem is still a multimodal problem. Since it is a multimodal problem, there are multimodal solutions, and multimodal has been proven to succeed in many fields, and this field is no exception.

Liu Yanqiu: Compared with the top embodied companies abroad, what are the gaps and advantages of your team?

Liu Songming: I think the gap in models is very small, because our latest results have not been released yet. The emergence capabilities we have observed ourselves are already on the same starting line as Silicon Valley companies, and we are running side by side. China should have taken the lead in hardware, but for various reasons, many domestic companies are still stuck in the era of teleoperation.

Liu Yanqiu: You are a post-2000 founder. Will investors pay attention to your management capabilities?

Liu Songming: Investors asked me many questions related to culture and management. My answers focus on three core points: benefits, information, and values.

In terms of benefits, the company is a concentric circle, with vision and business as the core. From the CEO to senior executives and then to frontline employees, the distribution must be fair, and everyone can benefit from the development of the company. In terms of information, we are highly aligned. I also write memos to share my thoughts internally. I think writing is more profound and less emotional than verbal communication. In terms of values, we advocate humility and empathy, and we assess whether candidates are open enough to put aside their ego and cooperate with others. We always adhere to the principle of one team, one goal, and oppose factional culture. We will not deliberately screen out people without impressive resumes, and we pay more attention to whether there is direct evidence to prove their engineering capabilities. This is also a very core