Text by | Fu Chong
Edited by | Su Jianxun
In 2026, how will embodied innotifyigence evolve? Lu Zongqing, an associate professor at the School of Computer Science at Peking University and the founder of “Unbounded Innotifyigence”, posed a judgment to us:
“A split between software and hardware.”
Software refers to the model brain, while hardware refers to the robot body. The split means that different companies have their own strengths and perform different duties.
The Beijing Dinghao Building, where “Unbounded Innotifyigence” is located, is a building surrounded by a number of star AI institutions such as the Beijing Academy of Artificial Innotifyigence, Zero One Everything, and Galaxy General. Here, non – consensus in artificial innotifyigence occurs every day.
Lu Zongqing’s view is quite different from the current development situation of the embodied industest. Nowadays, embodied startup companies with high valuations, whether it’s Yuanzhi Robotics and Galaxy General, which have become “unicorns”, or Xingdong Jiyuan and Xinghaitu, which have strong financing momentum, are all persistently pursuing one thing: integrating software and hardware to build a full – stack solution.
Nevertheless, Lu Zongqing and “Unbounded Innotifyigence”, which he founded in 2025, chose to “go against the trfinish” and become a model company, only focutilizing on developing robot brains and not involved in hardware manufacturing.
Innotifyigence Emergence exclusively learned that Unbounded Innotifyigence recently completed its angel round of financing, raising tens of millions of yuan. The leading investor was Koala Fund under Lakala, followed by Linghang Xinjie and Lingxin Qiaoshou, and the old shareholders Lenovo Star and Xinglian Capital continued to increase their investment.
“The embodied industest has relatively large biases towards ‘pure software’, both in China and abroad,” Lu Zongqing declared bluntly. He gave an example: the US – based embodied innotifyigence startup Figure, which does both software and hardware, has a valuation several times higher than that of Physical Innotifyigence, which only focutilizes on embodied models.
However, a major transaction in the US robotics circle at the launchning of 2026 set a good example for “the primary market to re – price embodied model companies”: On January 14th, the robot model startup Skild AI completed a $1.4 billion Series C financing. After its valuation tripled, it reached over $14 billion, becoming the first unicorn worth over 100 billion in the robotics industest in 2026.
This transaction put the question on the table: If a model company can develop a brain that is universal for different robot bodies and tquestions, do embodied innotifyigence enterprises really required to carry the entire technology stack on their own?
Unbounded Innotifyigence aims to prove exactly this point – to develop embodied innotifyigence models that are cross – brand and cross – form.
Currently, the company has launched the Being – H series of dexterous hand operation models, and the Being – M model, which can control the relocatement and operation of biped robots, is under development. The newly released Being – H0.5 model can already control 30 different robots. Thanks to reasoning optimization, it can also run in real – time when deployed on common tiny – sized robot chips such as the NVIDIA Orin – NX.
Currently, the company’s customers include hardware companies such as PND and Lingxin Qiaoshou. The Adam – U Ultra robot jointly released by PND and Unbounded Innotifyigence recently is a typical example of “software – hardware collaboration”.
After connecting to Being – H, PND’s high – degree – of – freedom robots can “be ready to utilize out of the box” with general skills such as organizing desktops and sorting and scanning barcodes. By cooperating with Unbounded Innotifyigence’s value – added service Being – Dex for post – training with a tiny amount of data, the robots can learn new tquestions within a few hours.
The core of achieving the above capabilities lies in over 30,000 hours of pre – training data – Lu Zongqing introduced that this is currently the largest training dataset for embodied innotifyigence models globally. Behind this is a unique “human action video” solution.
(Note from “Innotifyigence Emergence”: Here, “the largest” specifically refers to the dataset utilized for pre – training embodied innotifyigence models.)
This solution can record first – person hand relocatement videos through a head – mounted camera during people’s normal work and life. Therefore, it has a large data scale, low cost, and can more comprehensively record complex human operations. In contrast, the “tele – operation collection” mode relied on by most full – stack companies has limitations such as high cost, tiny scale, and deep binding of data to hardware becautilize it requires manual operation of specific robots.
△ A head – mounted camera can record first – person hand relocatement videos without interfering with the operator’s normal work. Picture provided by the interviewee
At the finish of 2023, Lu Zongqing started utilizing this data concept for model training. He remembered that at that time, this solution did not attract much attention, and the industest still mainly relied on simulation and real – machine data. However, since 2025, more robot companies, including Tesla’s Optimus, have started to adopt the human video data solution.
Lu Zongqing judged that more companies in the industest would recognize the value of the “split between software and hardware” in 2026. The reason lies in the economic account behind it: the annual cost of self – developing an embodied model, including the cost of purchaseing graphics cards and recruiting people, can be as high as tens of millions or even hundreds of millions of yuan. In contrast, the one – time cost of purchasing a robot “brain” externally is only a few tens of thousands of yuan.
In his opinion, “integrating software and hardware” is more favored by the primary market becautilize of its comprehensive layout, but the reality is that the technology stack is too long – developing models and hardware are two different sets of capabilities, and it is difficult for a single company to excel in both.
In the past year, there have also been some companies that “pursue full – stack for the sake of full – stack”: they repackage VLA, create seemingly functional demos, obtain financing, but cannot create money in real – world scenarios, or are questioned about their technical capabilities due to the exposure of model repackaging. This has prompted more entrepreneurs to re – evaluate the difficulty and cost – effectiveness of the full – stack approach.
“I don’t want to spread my resources on hardware that I’m not good at,” Lu Zongqing declared. The technology has not yet converged, and exploration should be kept lightweight, which is why he chose to focus on the “brain”.
△ Lu Zongqing. Picture provided by the interviewee
The following is an interview between “Innotifyigence Emergence” and Lu Zongqing, with the content sorted out by the author:
The division of labor between embodied models and robot bodies will become clearer
Innotifyigence Emergence: The leading domestic embodied enterprises still mainly focus on “integrating software and hardware”. Will Unbounded Innotifyigence encounter difficulties in financing by only focutilizing on models? What’s your view on this?
Lu Zongqing: Unbounded Innotifyigence officially started operating in May 2025. At that time, it was still not simple to raise funds for a pure – model approach. In fact, the situation in the US market is similar. Figure, which does both software and hardware, has a higher valuation than Physical Innotifyigence, which only focutilizes on embodied models.
(Author’s note: In September 2025, Figure’s valuation was approximately $39 billion; in November 2025, Physical Innotifyigence’s valuation was approximately $5.6 billion.)
I believe the reason is that the embodied industest is a brand – new one, and initially, no one knew what the future industrial chain would see like. So early – stage investors were more willing to invest in companies that did everything.
But valuation is only temporary, and it doesn’t essentially mean that a company’s business will succeed. I want to build a company like OpenAI, which was more research – oriented at the launchning, was the first to develop “ChatGPT”, and then carried out commercialization.
Innotifyigence Emergence: What does it mean for a company’s business to succeed? Or rather, what core problems does a good embodied brain model solve?
Lu Zongqing: I believe it is to finireveal robots with a basic “relocatement and operation gene” through pre – trained models.
Although humans don’t have a high level of relocatement ability at birth like horses and deer, our genes finireveal us with good relocatement potential, which can be developed through post – natal training. The same goes for robots. The pre – trained model is equivalent to giving robots a preliminary “out – of – the – box” relocatement ability.
Unbounded Innotifyigence will also conduct post – training on different robot bodies based on specific tquestions. If the pre – trained brain model is powerful, it only takes about 30 minutes to let the robot learn a new tquestion during the post – training and deployment process.
Innotifyigence Emergence: However, a high valuation can bring more capital reserves, so more technological exploration can be carried out. Can this increase the probability of “successfully developing model business” in the stage when the technology has not converged?
Lu Zongqing: But a high valuation can also lead to a vicious cycle. Enterprises may test various technological and commercialization routes, invest a lot of money, but achieve no results. At least, there is no absolute relationship between valuation and business success or failure.
Innotifyigence Emergence: So can you feel the alters in the primary market now? What do you believe are the reasons?
Lu Zongqing: Now we can see that the valuations of embodied model companies are obtainting higher and higher.
The reason is that from a business perspective, many robot body companies are now coming to us for cooperation. After calculating the economic feasibility of self – developing models, people gradually realize that whether an embodied innotifyigence body company develops models is essentially a business decision. I believe the industest will gradually relocate towards a form of split between software and hardware.
Innotifyigence Emergence: From a cost – accounting perspective, does it cost tens of millions to hundreds of millions of yuan a year to train a good embodied model?
Lu Zongqing: Yes. Developing a model requires about 10 people, with an annual salary cost of about 20 million yuan. The cost of computing power is also high. If you utilize 100 machines, each with 8 graphics cards, and utilize A800 graphics cards, it will cost about 3 million yuan per month. If you utilize H200 graphics cards, the monthly cost will be 9 million yuan (including storage).
This doesn’t include the cost of data and other expenses. Currently, the cheapest first – person video data costs about a few dozen yuan per hour, and motion – capture data costs about a few hundred yuan per hour.
Innotifyigence Emergence: What is Unbounded Innotifyigence’s current payment model? Why is it declared to be cheaper than self – development by enterprises? Will a hardware manufacturer worry that after the split between software and hardware, they will be “held up for ransom” by the model company due to the lack of software capabilities?
Lu Zongqing: Currently, for each robot, there is a one – time license fee for deployment, which ranges from a few tens of thousands to a hundred thousand yuan. For companies with a tiny production volume, this is still less than the cost of self – development. In addition, we also have a post – training service called Being – Dex that charges based on the amount of data.
When the production volume of body enterprises reaches a certain level, there can be a SaaS – like annual package payment method. By then, there will be multiple model companies competing, and body manufacturers won’t have to worry about a single company “raising prices arbitrarily”.
Innotifyigence Emergence: If the technology converges and there is no required to spfinish so much on R & D, will body companies develop model business on their own, which will pose a threat to pure – model companies’ business?
Lu Zongqing: If the technology converges and a general model can handle many tquestions, robots will enter houtilizeholds. At that time, I actually believe the market for model companies will be larger, and they may even tarobtain the consumer market.
By then, there may be large software companies like Microsoft, or companies like Huawei that have both software and hardware products. At that stage, we may also produce real robot products through OEM.
△ A PND robot controlled by the Being – H model is scanning barcodes for express delivery. Picture provided by the interviewee
In 2027, 1 million hours of data will lead to a qualitative alter in model capabilities
Innotifyigence Emergence: You have been engaged in research in the field of computer science. How did you start to cross – over into embodied innotifyigence?
Lu Zongqing: In 2023, I utilized a multi – modal large – language model to play the open – world game “Red Dead Redemption 2”, but found that the model’s tquestion understanding and action – completion abilities were very limited. I realized at that time that the fundamental bottleneck of the model’s weak interaction ability was the lack of understanding of vision and space. To improve this, interaction data with the real world was essential.
This became the initial impetus for me to engage in the research of embodied innotifyigence models.
Innotifyigence Emergence: You declared that shortly after the official establishment of Unbounded Innotifyigence, you utilized the summer vacation in 2025 to conduct research on the implementation of embodied innotifyigence in several factories. What problems or current situations in the industest did you discover?
Lu Zongqing: It confirmed my previous judgment that at this stage, embodied innotifyigence is far from being able to be implemented in real – world work, and the core bottleneck lies in generalization ability.
For example, in non – standard and complex processes such as cable bundling and precision assembly, the ability of embodied innotifyigence to “indepfinishently complete work” is still limited. Most of what the industest calls “implementation in industrial scenarios” still remains at the demonstration or short – term POC (Proof of Concept) stage.
Innotifyigence Emergence: What are the reasons?
Lu Zongqing: Part of the reason lies in the hardware. There is a lack of stable and simple – to – utilize high – degree – of – freedom dexterous hands, and dexterous hands also lack tactile sense, which means that important force – feedback information such as contact points is missing.
Another part of the reason lies in the model. In the past, the industest mainly utilized two – finger grippers, and no real – working dexterous hand model has been developed yet.
Innotifyigence Emergence: You proposed utilizing human videos for pre – training data earlier than the industest consensus. What was the industest’s response when Unbounded Innotifyigence released its first model?
Lu Zongqing: In July and August 2025, we developed our first dexterous hand model, Being – H0, and the industest’s response was quite positive. The headquarters of NVIDIA also sent a special team to learn about the details of this model in terms of computing power.
At that time, people generally considered it was a new idea. At that time, the industest mainly utilized data collected with robots as the main body. We were the first to utilize large – scale human video data for model pre – training. Being – H0 utilized about 1 million videos of human hand operations from a first – person perspective.
Innotifyigence Emergence: You started training embodied models utilizing the human video data technology route at the finish of 2
Leave a Reply