Wang Xingxing enters the Odyssey period and calls for Liang Wenfeng

On the day Unitree Robotics was listed on the A-share market, the World Robot Conference (WRC) was held in Beijing, and both sides were extremely bustling. On one side, Unitree's share price rose more than 6 times from its issue price, with a total market cap exceeding 440 billion yuan (closing at 341.8 billion yuan). At the scene, some investors pulled Wang Xingxing to take group photos and toast with him.

On the other side, at Unitree Robotics' booth at WRC, seven or eight robots started dancing with their best motion control capabilities to the background music "Love Bird".

Wang Xingxing's decade of relentless dedication to robotics has finally borne fruit today. We stayed at Unitree Robotics' booth at WRC for a long time. In addition to dancing, there was a Unitree robot playing table tennis, and a large screen next to it was playing a video of Unitree robots performing yangko dance on the Spring Festival Gala.

We also talked to Hu Likang from Wulian Technology, an ecological partner of Unitree Robotics, who was happy for Unitree's listing. "Unitree has brought robots to the consumer grade. The R1 series only costs more than 20,000 yuan, which is definitely worthy of praise from everyone."

Today should have been the happiest day for Wang Xingxing. However, from the photos spread from the listing ceremony scene and some media reports, Wang Xingxing looked serious without a trace of a smile.

It is perhaps understandable that after ten years of sharpening the sword, the successful listing is a market reward for Unitree, but it is far from a relaxing moment for the founder.

Unitree's non-recurring profit and loss net profit in Q1 and Q2 of 2026 has declined year-on-year for two consecutive quarters. Wang Xingxing needs to constantly step out of his technical comfort zone and face the pressure from capital and the market directly. He has to prove that Unitree has the best "body" in the industry, and find a sufficiently smart "brain" and sufficiently specific work scenarios for this body before the capital market loses patience.

At the celebration banquet on the first day of listing, Wang Xingxing has already learned to tell capital stories. On the core issue of Unitree Robotics, which has always valued the ontology and neglected the "brain", Wang Xingxing described a narrative that caters to the Silicon Valley technology boom.

He believes that the next stage of robotics is to let AI fully participate in the whole R&D process of robots, and realize autonomous learning and self-evolution capabilities in all-round scientific research processes such as design, training, and iteration.

This narrative is logically valid, but it fails to answer the most urgent problem that Unitree is facing right now. Unitree's humanoid robots are mostly used in scenarios such as scientific research, leasing and personal development, and truly productive application scenarios are not yet clear.

The generalization of the "brain" is required to realize the large-scale deployment of productive application scenarios. The "brain" is exactly the short board of Unitree. Unitree's previous choice to cooperate with DeepSeek may be exactly to make up for this short board, but this cooperation process is still full of uncertainties.

Unitree's Core Problem

The rise in share price on the first day of listing cannot answer the most core problem of Unitree.

At the WRC conference, most companies with robot ontologies are demonstrating their progress in specific vertical scenarios such as logistics, household, and industry. In the past two years, no matter at the Beijing Robot Conference or the WAIC in Shanghai, the style of Unitree Robotics' booth has always been the same: either dancing or boxing.

On the eve of listing, Unitree even launched the "Superman" robot, which can jump 2 meters high in place and reach an ultimate running speed of 12.66 m/s. Both figures exceed the limit level of human beings.

An investor in Nanjing spoke highly of Unitree, telling us, "In terms of motion control capability and 'cerebellum', Unitree is almost second to none among humanoid robots." A clear indicator to evaluate the quality of a humanoid robot product is consistency: for example, when you give an instruction to one Unitree robot, several other robots of the same model can all execute it.

No matter performing Tai Chi or other shows, every movement and posture of Unitree robots will not be deformed, and the motion trajectories are consistent.

Many robotics companies, including Galaxy General, will buy Unitree's humanoid robots for research and reference. The investor gave an example that this is very normal, just like Lei Jun bought a Ferrari for reference before making cars.

It is precisely for this reason that Unitree's humanoid robot sales increased to 5,511 units in 2025, a 10-fold year-on-year increase. According to Unitree Robotics' prospectus, it ranks first in the world in terms of sales of humanoid robots (pure humanoid robots, excluding wheeled humanoid robots).

These facts do explain why Unitree can arouse public enthusiasm and become the focus of various activities and even public opinion. But for insiders who are used to thinking about commercialization, these actions are like "useless vases", with no visible landing value.

Unitree's continuous breakthroughs in robotics are more like carrying Wang Xingxing's ultimate imagination of physical carriers, which is essentially driven by personal interest. It is just that the era has picked him now.

In a media interview, Wang Xingxing mentioned that he agreed with a viewpoint: "When what you do conforms to the main theme of the era, many high-quality resources will tilt towards you."

Despite the huge resource tilt, in terms of actual productivity, in addition to the certain applications of quadruped robots in scenarios such as industrial inspection, Unitree's humanoid robots have rarely disclosed specific deployment cases in industrial or service scenarios so far, and the large-scale landing situation is also unclear.

Unitree sold more than 5,000 humanoid robots last year, but almost all of them were purchased by university scientific research institutions, leasing parties, and secondary development groups, rather than other embodied intelligence companies that build industrial solutions.

While ontology robot companies such as Agibot and Galaxy General announced their cooperation with industrial parties one after another last year, they accelerated the deployment of humanoid robots (wheeled) in industrial scenarios such as 3C manufacturing through large order layouts, hoping to obtain real industrial scenario data to feed back model training and iteration.

Industrial scenarios are indeed not Unitree's area of expertise. It is better at selling products with light delivery. Industrial cooperation requires deep involvement on site and undertaking a large number of non-standard, heavy workloads that need to be polished in details with industrial parties, which probably does not match Wang Xingxing's personal preference.

Completing a highly difficult action and repeating a low-difficulty action 100,000 times are two different kinds of engineering capabilities. Wang Xingxing chose the former and pushed it to the extreme.

In the prospectus, Unitree did not elaborate on the uncertainty of application scenarios and future commercialization, but just briefly mentioned in a general sentence: "In some specific usage scenarios or segmented application fields, the company has limited development resources invested in a targeted manner, so there is a risk of failing to obtain leading advantages."

However, Unitree's prospectus did respond to the problem of insufficient "brain", stating that it was due to insufficient R&D investment in real machine data collection and scenario training when the early technical route and scenario requirements were not yet clear.

Deeply Bound with DeepSeek

In the embodied intelligence industry, the most imaginative direction now is no longer the ontology, but the "brain", which is another problem of Unitree's robots.

A large part of the raised funds from Unitree's this listing will be used for brain R&D. The original planned raised amount was 4.202 billion yuan, of which 2.022 billion yuan (accounting for 48.13%, nearly half) will be invested in the "Intelligent Robot Model R&D Project", while only 1.11 billion yuan (accounting for 26.41%) will be used for ontology R&D.

However, this model R&D project is not just the so-called "brain". The prospectus clearly states that it will carry out R&D on embodied large models, embodied ontology intelligent models, data collection, simulation and synthesis, as well as supporting technologies for both "brain" and "cerebellum".

In addition to independent R&D of embodied models, the cooperation between Unitree and DeepSeek is also quite notable. Before Unitree's listing, these two star companies both based in Hangzhou officially announced their cooperation.

On August 7, Wang Xingxing confirmed at the roadshow site that the two parties had signed the "Strategic Cooperation Memorandum", covering three major directions: joint R&D of general artificial intelligence, high-performance general robots and AI large models.

When DeepSeek carries out embodied intelligence business, it will give priority to purchasing Unitree's robots under the same conditions. When Unitree trains its embodied large model, it will give priority to calling DeepSeek's models and computing power under the same conditions, and DeepSeek will provide technical support such as model architecture scheme and intelligent computing cluster construction.

The two parties further bound their interests through equity. The night before, Unitree disclosed the strategic placement list: DeepSeek was allocated 933,400 shares at a cost of 141 million yuan. The lock-up period of CNPC, China Southern Power Grid and Tencent's capital in the list is 12 months, while only DeepSeek has a 36-month lock-up period.

141 million yuan is not a large amount in strategic placements that usually cost hundreds of millions of yuan. What is really valuable is the three-year lock-up period, which means DeepSeek voluntarily gave up all the liquidity premium in the early stage of listing.

This is almost a two-way lock-up, which means neither party can exit within three years. The logic behind why these two parties come together is not obscure: Unitree Robotics provides top robot hardware ontology as the carrier, while DeepSeek can obtain priority access to physical world data feedback and test scenarios.

Liang Wenfeng once described DeepSeek's understanding of AGI as a ladder in an investor communication: after the language model comes CoT, then Agent, then solving continuous learning, then moving towards model self-iteration, and finally reaching embodied intelligence.

In other words, to reach the final destination, large model companies need real machine data and hardware carriers.

In March 2026, Unitree open-sourced the full-body teleoperation real machine dataset of humanoid robots in open scenarios, and announced that it will continue to update it at high frequency. This kind of data asset cannot be created by model companies themselves.

Conversely, Wang Xingxing also needs Liang Wenfeng. Unitree Robotics clearly explained the strategic consideration of introducing DeepSeek in its announcement: "Focusing on the R&D and product development of AI large model and embodied intelligence related technologies, we aim to integrate the technical advantages of both parties, improve the understanding ability and generalization ability of embodied intelligent robots for complex scenarios, and jointly promote the technological development of general artificial intelligence."

This means that DeepSeek may participate in the underlying technical route and computing power infrastructure of Unitree's next-generation embodied intelligent model. In short, Liang Wenfeng is starting to find "bodies" for his models, while Wang Xingxing is starting to find "brains" for his robots.

Wang Xingxing once said at the prospectus roadshow site that listing is just a new starting point. In the next 2 to 3 years, the "ChatGPT moment" of embodied intelligence is expected to arrive. When robots can complete 80% of tasks in 80% of unfamiliar scenarios only through language or text instructions, the industry will usher in a real explosion.

2 to 3 years is exactly the period covered by DeepSeek's 36-month lock-up, which is a bet against the time window.

But there was a precedent before that did not end well. In 2024, Figure AI cooperated with OpenAI, with one providing the body and the other providing the model. Only one year later, the two parties broke up, and Figure turned to independent R&D.

Embodied large models are actually very different from general large models. To make robot models, you must constantly run the models on real robots to see where they fail, then collect data, train, deploy, and continue testing.

At that time, it was difficult for the two parties to truly form such high-frequency joint testing and iteration, and the OpenAI team did not deeply participate in Figure's robot testing.

This is also the most noteworthy part of the cooperation between Unitree and DeepSeek this time: are they following the path that Figure once took, or can they take a different path through deeper binding of models, data, computing power and ontology.

This article is from the WeChat Official Account "White Whale Lab", written by Liu Jia, edited by Wu Xun, and published with authorization from 36Kr.