IDG Capital führt die Angel-Runde an. Projekte mit Beteiligung von Alibaba und Tsinghua University im Bereich der Embodied AI haben in sechs Monaten über 100 Millionen Yuan an Kapital beschafft | Exklusivbericht von 36Kr

36Kr has learned that the industrial Embodied AI company Guangxiang Technology has completed several rounds of financing, including the Seed round, the Angel round, and the Angel+ round. The cumulative financing amount exceeds 100 million yuan. The financing rounds were jointly led by the financial investment institutions IDG Capital and Orient Fortune Capital, while the robotics industry capital providers EVATEC, Zero One Capital, Datai Capital, and the L2F Entrepreneurs Fund participated as co - investors.

It is known that the company's financing funds are mainly used for the research and development of the core technologies of Embodied robots, product development, and commercial delivery.

Guangxiang Technology was jointly founded in April 2025 by Zhang Tao, the former CTO of Alibaba's Gaode, and Li Shengbo, a professor at Tsinghua University and an expert in the field of artificial intelligence. Currently, Guangxiang Technology is already a strategic partner for Embodied AI for several global automobile manufacturers. Based on the Embodied model that enables robots to learn by themselves and the tool platform that enables the mass introduction of Embodied AI, Guangxiang Technology hopes to "create a universal industrial Embodied brain for industrial manufacturing scenarios such as the automobile and 3C industries".

Transition from automobile manufacturing to general robotics

At the time when Zhang Tao decided to enter the Embodied - AI startup scene, there was an opinion in the industry that robotics companies that initially focused on vertical scenarios would be replaced in the future by companies that directly developed general robots.

However, Zhang Tao held a different view. He compared robots in vertical scenarios such as the industry to L2 in automated driving mode and general robots to L4. "If the technology develops fast enough, L4 can actually cover all L2 scenarios," he said. "But we believe that the robotics industry, like automated driving, will go through a long development cycle. Therefore, it is a more realistic business approach to gradually transition from vertical scenarios to universal robots for all scenarios."

Based on this consideration, Zhang Tao initially targeted cyclists - type industrial robots when founding the company.

In his opinion, industrial operation is a scenario with a "standardized environment and complex operations", which is currently challenging but also quickly implementable. In the industry, automobile manufacturing is the most established sector with sufficient market potential. Guangxiang Technology estimates that just the automation of the final assembly process in automobile manufacturing would have a market volume of hundreds of billions of yuan and could quickly expand to almost all industrial manufacturing scenarios.

After selecting the application case, the form of the robot can also be determined.

Zhang Tao explained to 36Kr why he chose cyclists - type robots: "The greatest advantage of two - legged humanoid robots lies in their ability to overcome terrain obstacles. However, in a standardized factory environment, these advantages cannot be fully utilized, while disadvantages such as high energy consumption and inaccurate positioning are more prominent. Cyclists - type robots, on the other hand, consume less energy and have more accurate positioning, which is better suited to the factory environment and requirements."

The market for automobile manufacturing robots has great potential, but accessing this market is not easy.

Zhang Tao told 36Kr that industrial robots are different from the demonstration robots on the CCTV Spring Festival Gala. For demonstration robots, people mainly care about whether they can perform the movement. In industrial scenarios, however, the standards for robot operation are stricter. "For example, industrial robots must consider both the accuracy of movement, the time rhythm, and the uniformity of movement."

In addition, there is a strong interaction between industrial robots and their environment. "The robot must perceive the state of the environment and the object to be operated in real - time, plan and execute the movement in real - time, and avoid collisions during the operation." All this poses a challenge to the creation of operation models.

Strive to make industrial robots self - learn and develop

One strategy of Guangxiang Technology is to "develop an industry - capable self - learning intelligence model".

When building the model, Guangxiang Technology has developed a highly smooth neural network structure specifically designed for industrial operations to ensure that the robot can perform highly precise, reliable, and uniform movements. In model training, instead of the easier - to - implement imitation learning, Guangxiang Technology uses the more promising but also more challenging method of reinforcement learning.

Zhang Tao said that although imitation learning "can quickly achieve an apparently good operation result with a small amount of data, such as a success rate of 90% to 95% in simple pick - and - place tasks", it is not able to guarantee the almost 100% success rate required in the industry, nor can it meet the multi - layered performance requirements such as efficiency and accuracy at the same time. However, these requirements are the key to high - quality automobile manufacturing.

Therefore, the Guangxiang team hopes that through the model training method of reinforcement learning, robots can "acquire the ability of sustainable self - learning and further development", thus opening up a technological path on which robots can continuously improve their performance and finally meet the strict requirements of the automobile manufacturing scenario for robots.

The training of the model depends on data. However, in the Embodied - AI industry, the shortage of real data is a problem that plagues most players in the industry.

Therefore, Guangxiang Technology proposes to increase the proportion of simulation data in model training and rely on the ability of high - precision scenario modeling and the resources of high - precision digital models of industrial customers to reduce the gap between simulation data and real data and thus close the model training chain from simulation to real robots.

Zhang Tao explained why the proportion of simulation data needs to be increased: "If we only want to make a demo project, we may be able to set up a fake workplace and collect some real data. But if we want to implement real applications in the future, the huge amounts of real data required for an extremely high success rate may not be achievable."

The GOPS platform is another preparation of Guangxiang Technology for the mass introduction of robots.

According to Zhang Tao, this platform has modularized the design, development, training, and even fine - tuning of Embodied - AI models for industrial scenarios. It can build a stable and efficient chain and enable high - quality end - to - end model development for each industrial scenario in clearly defined scenario tasks, so that the company obtains the "ability for mass delivery".

Currently, Guangxiang has already cooperated with several automobile companies and completed the first POC validation for real production workplaces. For the future, Guangxiang Technology has set the goal of reaching at least ten automobile manufacturers within the next three years and installing thousands of intelligent robots that meet the requirements of the factories. At the same time, Zhang Tao hopes that the company's products can also be used in other industrial manufacturing scenarios.

In Zhang Tao's long - term planning, Guangxiang Technology also wants to enter other large industrial and commercial scenarios outside of manufacturing and gradually approach universal Embodied AI.

Author's WhatsApp: luckg17305264638