news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Author｜Chang Minxiao

Editor: Qiu Xiaofen

Embodied intelligence is the most competitive field in the world this year. Musk launched the second generation of Optimus, and a number of major domestic companies have invested in humanoid robots. Various players are busy laying out the embodied field.

36Kr learned that the embodied intelligent robot company "Qianxun Intelligence" has completed nearly 200 million yuan in seed round + angel round financing. "Qianxun Intelligence" was founded by Han Fengtao, the former CTO of Luoshi Robotics. The company was founded in February 2024. In the six months since its establishment, the company has completed two rounds of financing in just four months.

The angel round of financing was led by Honghui Fund, followed by Dachen and Qiancheng, and seed round investors Shunwei Capital and Oasis Capital continued to invest. The funds raised in this round of financing will be mainly used for technology research and development and team expansion.

At present, humanoid robots generally face the common problems of weak generalization and limited interactivity. When reflected in work, this means that humanoid robots can only work in relatively fixed environments such as workshops and factories, have limited interaction with the surrounding environment, and their responses are not accurate or sensitive enough.

If humanoid robots are to be as smart as humans, they need to have an embodied macromodel as a smart brain.

In order to achieve the interactivity and generalization of robots, one of the core barriers of Qianxun Intelligence is to build a highly generalized and versatile robot brain. The visual language model is used in the brain.ViLa（The Vision-and-Language Models) and the component constraint model CoPa (Constraints of Parts) model serve as the underlying multimodal large model of the embodied intelligent robot.

It is worth noting that Figure AI has previously used the ViLa model to enable robots to understand daily scenarios and have common sense of life. The proposer of this model architecture is Gao Yang, the co-founder of "Qianxun Intelligence".

Image source: Qianxun Intelligence

In addition to building a basic large model suitable for embodied intelligence, another technical advantage of "Qianxun Intelligence" is that it solves the problem of data acquisition for the embodied intelligence large model.

Generally speaking, in order to teach robots how to work in real environments, scientists need to collect a large amount of data on people moving in real environments and pre-train humanoid robots to help them better understand the surrounding environment and complex scenes.

However, the current difficulty in obtaining pre-training data is the greatest.

Han Fengtao, founder and CEO of Qianxun Intelligence, told 36Kr that data, computing power and algorithms are the three elements for building a large model of embodied intelligence. Currently, most players have similar levels of computing power and cannot constitute an absolute technical barrier; the current technical route of the algorithm has not yet fully converged.Therefore, robot players can currently only widen the gap in existing data and the ability to collect new data.

The difficulty in obtaining training data is based on the following two reasons - first, the high-performance robot industry is still in its infancy, and it is difficult to obtain data from the robot itself.

Secondly, although training through simulation and synthetic data is also a solution, there is still a gap between virtual synthetic data and real-world data.

In order to overcome the problem of data shortage, Qianxun Intelligence's solution is to use general high-performance hardware systems, based on massive Internet data pre-training, high-sample efficiency imitation learning and reinforcement learning, to allow AI technology to better unleash hardware performance.

To this end, Gao Yang, co-founder of Qianxun Intelligence,Proposed the world's most sample-efficient reinforcement learning algorithm(EfficientZero and EfficientZero v2), by improving sample efficiency, solve the problem of data shortage from the bottom of the model.

In terms of imitation learning, Gao Yang proposed the EfficientImitate high-performance imitation learning algorithm, which can help robots learn to draw inferences during practical operations.It is reported that this algorithm has improved learning efficiency six times compared to Stanford's VMAIL algorithm.

In terms of products and commercialization, Qianxun Intelligence plans to use commercial, service and household applications as initial scenarios in the future.

In terms of the team, all members of Qianxun Intelligence have rich backgrounds in robotics research and development. Founder and CEO Dr. Han Fengtao has more than ten years of experience in the robotics industry, focusing on the research and development of high-performance light industrial robots. He has led the team to commercialize the results of 20+ industries, 100+ scenarios, and 1,000+ customers.

Professor Gao Yang from Tsinghua University, who is also a co-founder, has ten years of experience in embodied intelligence, machine vision, and machine learning research. He focuses on building a universal embodied intelligent brain that can perform any task in any scenario. He has worked closely with Sergey Levine, the founder of Physical Intelligence, a leading company in the field of embodied intelligence in the United States.

Investor comments:

Shunwei Capital Investment Director said" The Shunwei team has long been paying attention to the innovative opportunities in the field of robotics and embodied intelligence. We are optimistic about the combined background and industry experience of the founding team of Qianxun Intelligence and expect Qianxun to actively explore scenarios and business models and promote the implementation of embodied intelligence technology. "

The investment team of Honghui Fund said:"Embodied intelligence is an important application scenario of general artificial intelligence (AGI), and the market space is extremely broad. In the past, the control of robots relied on a large number of manual programming processes and had many restrictions on the scenarios. We are very optimistic about the improvement of traditional robots in terms of task generalization by the intelligent agent formed by the combination of embodied large model algorithms and hardware. This type of intelligent agent is a bridge between the virtual world and the physical world, and is also the best path to spatial intelligence.

China has a leading advantage in the robot hardware industry chain. The support of leading embodied models such as ViLa and CoPa is expected to help achieve overtaking. Qianxun Intelligence is our first target in this field. The team has extremely rich technical accumulation and product implementation experience. I believe that under the leadership of the founding team, the mass production of a new generation of intelligent robots will be just around the corner, setting off a new industrial revolution. Honghui Fund will combine its own relevant resources and work with the Qianxun team to promote the development of intelligent robots. "

Welcome to communicate, please add WeChat and indicate your name and purpose

end