Over the course of 16 trading days, with 11 consecutive 20% daily limit up sessions, the stock price surged 10.83 times and the market value surged by over 34 billion yuan. This is the record set by Shangwei New Materials (688585. SH), a listed company on the Science and Technology Innovation Board, in July. As a result, Shangwei New Materials has become the first and only "tenfold stock" this year.
Behind this capital frenzy is a embodied intelligence company that has been established for less than 3 years - Zhiyuan Robotics.
On July 8th, Shangwei New Materials announced that Shanghai Zhiyuan New Technology Co., Ltd. [Zhiyuan Robot Operation Subject, now renamed Zhiyuan Innovation (Shanghai) Technology Co., Ltd.] plans to acquire no more than 63.62% of Shangwei New Materials' shares through a shareholding platform jointly funded by the company and its core team for a price of 2.1 billion yuan.
After the acquisition is completed, the controlling shareholder of Shangwei New Materials will be changed to a joint shareholder of Zhiyuan Robotics and its management team, and the actual controller will be changed to Deng Taihua. The core team includes Peng Zhihui and others.
Zhiyuan Robotics first gained attention from capital and the industry due to former Huawei "genius youth" Peng Zhihui. Peng Zhihui was born in 1993, with the online name "Zhihui Jun". He is one of the top 100 UP owners on Bilibili, with 2.82 million followers. He has worked at Huawei in the field of Ascend AI chips and AI algorithms.
During his time at Huawei, Peng Zhihui was praised by Ren Zhengfei in a speech, calling him the "driving force of Huawei's innovation". But at the end of 2022, Peng Zhihui announced in a post that he would leave Huawei and participate in the founding of Zhiyuan Robotics as CTO.
Subsequently, the founding team of Zhiyuan Robotics continued to expand. In March 2025, the legal representative of Zhiyuan Innovation was changed to Deng Taihua. Deng was once the Vice President of Huawei, leading the Huawei 5G campaign and building the Ascend AI ecosystem. Deng was not only Peng Zhihui's former boss at Huawei, but also an alumnus of his University of Electronic Science and Technology of China; In April 2025, Luo Jianlan will be appointed as the Chief Scientist. Luo was a research scientist at Google DeepMind and a core member of Sergey Levine's team at the University of Berkeley, where he led the development of the world's first superhuman robot reinforcement learning system; In 2024, Yao Maoqing joined as a partner and president of the business department. Yao had previously been responsible for autonomous driving algorithms at Waymo and NIO.
The financing speed and amount of Zhiyuan Robot have reached the highest level in the industry. Less than a year after its establishment, its valuation quickly reached $1 billion, making it the fastest physical intelligence company in the world to enter the unicorn market. Investors include well-known institutions such as Hillhouse Capital, Jingwei Capital, Dinghui Capital, and BYD.
As of now, Zhiyuan Robotics has completed 11 rounds of financing. In March 2025, Zhiyuan Robotics received a new round of financing led by Tencent; In May 2025, JD.com and Shanghai Guishen Intelligent Fund participated in its new round of financing, while SAIC and other old shareholders increased their capital. After the completion of this round of financing, third-party research institutions have valued Zhiyuan at as high as 15 billion yuan; On August 1, 2025, Zhiyuan Robotics announced once again that it has obtained strategic investment from an international group, led by LG Electronics and Mirae Asset of South Korea.
Recently, due to jointly winning the first 124 million yuan order in the field of humanoid robots with Yushu, Zhiyuan Robotics has once again been pushed into the spotlight.
According to China Mobile Procurement and Tendering Network, Zhiyuan Robot won the bid for the humanoid robot procurement project of a wholly-owned subsidiary of China Mobile for 78 million yuan. A large order of 78 million yuan is not only a breakthrough in short-term revenue for Zhiyuan Robotics, but also a key catalyst for promoting its comprehensive upgrade in technology, supply chain, and market strategy.
Specifically, it is a bipedal humanoid robot customized by China Mobile, mainly used in the stores and business halls of China Mobile operators to do interactive work such as reception and explanation. In the future, similar service reception scenes will be seen in operators, hotels, banks and other places, and there is actually a huge demand for them. "Regarding order details, Yao Maoqing told" Chinese Entrepreneur ".
At present, Zhiyuan robots have entered the "centralized commercial delivery stage". Zhiyuan has already produced over 2000 robots in offline production, and it is expected that the shipment volume will reach thousands of units this year. Breakthroughs have been achieved in four major scenarios: industrial manufacturing, warehousing and logistics, power inspection, and interactive guidance, "said Yao Maoqing. In addition, he revealed that Zhiyuan has launched overseas plans and has achieved layout in markets such as North America, Europe, the Middle East, Japan, South Korea, and Southeast Asia.
Full stack layout
When some entrepreneurs in the field of robotics seek a single breakthrough, Zhiyuan has chosen a different path, like Apple and Tesla, to thoroughly optimize the entire software and hardware stack and layout in a closed-loop manner.
In terms of product layout, Zhiyuan Robotics has three major robot families: Expedition, Elf, and Rhinoceros, covering various commercial scenarios. Among them, the Expedition series is positioned for industrial grade applications; Elf series, aimed at commercial and light industrial scenarios; The Lingxi series focuses on scientific research and cutting-edge exploration.
In terms of technological layout, Zhiyuan Robot has chosen to independently develop the robot body, cerebellum, and brain full stack, building the core capabilities of the robot in interaction, operation, and motion intelligence. This is not an easy path, on the one hand, there is a huge amount of capital investment; On the other hand, any bottleneck in any link can become a constraint for the entire company.
Regarding the reason for choosing a full stack layout, Yao Maoqing told "Chinese Entrepreneur": "The landing of intelligent robots is a tightly closed flywheel, and it is difficult to completely outsource one of them. In the practical process, it is a cyclical process, from ontology design, to data collection, to iterative models. After the deployment is completed, you will find many new areas that need to be improved. Therefore, the team is currently investing in saturation in various directions
In terms of capital operation, Zhiyuan robots have also taken a completely different path. On the one hand, Zhiyuan Robotics has received 11 rounds of external investment within 2 years; On the other hand, Zhiyuan Robotics is also actively expanding its upstream supply chain through investment.
According to statistics from "Chinese Entrepreneur", Zhiyuan Robotics has currently invested in more than 10 companies related to body intelligent robots, including Digital Huaxia, Lingchu Intelligence, Qianjue Robotics, and Fuxing Electromechanical. This includes upstream companies such as robot components, embodied intelligent systems, and biomimetic robots, as well as joint ventures between Zhiyuan and listed companies such as Bozhong Precision, Dafeng Industry, Wolong Electric Drive, Softtek, and Fulin Precision.
We often invest more in upstream supply chains, such as sensors, joint modules, and so on. However, some of the investors and shareholders we introduce are actually scene parties, including automotive, 3C electronics, and so on. ”Yao Maoqing told 'Chinese Entrepreneur'.
Regarding the progress of keeping up with downstream partners' collaboration, Yao Maoqing said, "We have already imported some upstream supply chain enterprises' components into our entire machine, and the downstream scene has also been opened up for us. Both parties have jointly built some POC (proof of concept) projects, and there is an opportunity to make some deliveries in the second half of the year. ”
Self built grain silo
Currently, the bottleneck in the development of embodied intelligent robot technology lies in the extreme lack of high-quality scene data, which is also the key reason why humanoid robots exhibit significant differences in performance in different scenarios.
For this reason, Zhiyuan Robot chose to "build its own granary".
In September 2024, Zhiyuan Robot built a data collection factory of over 3000 square meters, containing more than 3000 real objects. At present, through this factory, Zhiyuan has formed the world's largest dataset AgiBot World and opened it up, covering hundreds of scenarios such as home furnishings, laboratories, and retail stores.
Faced with the desert of embodied intelligent data, we chose to plant the first tree, hoping it could become a forest, "Yao Maoqing said when talking about the original intention of creating datasets and open source.
In March 2025, Zhiyuan Robot released the first universal incarnate base model - "Zhiyuan Qiyuan Grand Model GO1", which can combine Internet video and real human demonstration to learn and enhance the model's understanding of human behavior. In the design concept of Zhiyuan, this model is a universal robot strategy model that can migrate between different robot forms and seamlessly adapt to heterogeneous robot bodies such as bipeds, wheels, and robotic arms, greatly reducing the cost of intelligent migration.
Embodied intelligence refers to giving AI models a "body" to perceive, make decisions, and act in the real physical world. Therefore, while humanoid robots are embroidering their legs with flower fists, many people are also concerned about when the "ChatGPT moment" belonging to the embodied intelligence industry will come.
In Luo Jianlan's opinion, achieving the 'ChatGPT moment' is meaningless, and the goal is to achieve the 'DeepSeeker R1 moment' after post training optimization. Enabling robots to have both generalization and reduced illusion rates without compromising performance, resulting in a success rate of nearly 100% for each task, is the only way to truly make robots useful in the real physical world. ”
I only know the principle of tying shoelaces, but the actual success rate is only 10%, so I can't leave the house for three hours, "Luo Jianlan explained to" Chinese Entrepreneur ".
To this end, Zhiyuan Robotics released Genie Envisioner, the industry's first open-source platform for real-world dual arm robots, at the 2025 World Artificial Intelligence Conference. The platform is based on the collected AgiBot World dataset and utilizes over 1 million synchronized video streams of the head and wrists for nearly 3000 hours to capture spatial layout, motion evolution, and semantic intent in robot operation tasks. It can help robots complete cross modal transfer from visual perception to robot action execution.
If the GO1 model solves the problems of 'what to do' and 'how to do it', the world model solves the problems of 'how the environment changes' and' what are the consequences of actions'. The former helps robots achieve multitasking decision-making and action execution, while the latter helps robots build dynamic inference capabilities for physical environments. ”A embodied intelligence practitioner told China Entrepreneur.
Go off to work
In the embodied intelligent robot industry, evaluating the quality of a robot mainly depends on two indicators: generalization and performance. Generalization refers to an individual's ability to apply skills in different contexts, while performance refers to a robot's success rate and speed in completing tasks.
Unlike large models, robots have very high requirements for performance. If the success rate of task execution is only 50% or 60%, such robots cannot be applied in the real world. It's like a robot pouring water for you, spilling it twice, which is definitely not accepted by everyone, "Luo Jianlan told" Chinese Entrepreneur ".
In 2025, many embodied intelligence practitioners refer to it as the "first year of mass production" and "first year of commercialization" for humanoid robots. Luo Jianlan also believes that the robotics industry has passed the stage of showcasing demos and is evolving from showcasing technology to a closed-loop industry model.
At present, the entire application scenario and market direction of Zhiyuan robots are mainly focused on the B-end, and have achieved breakthrough implementation in four major scenarios: industrial manufacturing, warehousing and logistics, power inspection, and interactive guidance.
During the 2025 World Artificial Intelligence Conference, Zhiyuan Robot also collaborated with Dema Technology to conduct a global live broadcast, where Zhiyuan Robot transformed into a courier and sorted packages on site. According to Yao Maoqing, in less than a month of cooperation between Zhiyuan Robot and Dema Technology, the sorting speed of the Elf G1 has been increased to 6 seconds per piece, which can meet the needs of some clothing e-commerce enterprises.
However, although there have been some practical applications of humanoid robots working in the field, there are still certain limitations in the scenarios.
Over the past year, Yao Maoqing has clearly felt a change in customer attitudes. Previously, they thought that simply plugging in the plug would allow them to use it directly. Today, they realize that it actually requires bidirectional deployment, which is a data-driven process that requires robots to be trained and iterated in actual scenarios. Additionally, they need to cooperate with some production line modifications to better connect the robots. At the same time, from the perspective of customer funds, they are willing to invest a lot of money in scenario verification with us
Wang Chuang, President of the General Business Department of Zhiyuan Robotics, is responsible for the commercial implementation of Zhiyuan Robotics. He said, "In terms of scenario implementation, the implementation of intelligent robots follows the principle of 'easy first, difficult later'. We first solve scenarios where traditional intelligence cannot be achieved but embodied intelligence is easy to implement, such as material box transfer, and then tackle more complex scenarios such as material sorting that require generalization ability
|
16
0
