The universal brain for robots has been put into industrialization, and "Qianjue Technology" has successively completed the Angel+ and Angel++ rounds of financing | 36Kr exclusive.

Written by Wang Fangyu

Edited by Su Jianxun

36Kr has learned that Beijing Qianjue Technology Co., Ltd. (hereinafter referred to as "Qianjue Technology") has recently completed consecutive rounds of financing of tens of millions of yuan for Angel+ and Angel++, with a cumulative financing reaching a scale of hundreds of millions of yuan.

The two rounds of investment were led by Huaye Tiancheng and Ruijiang Investment, funded by Gong Hongjia, the co-founder of Hikvision. Zero2IPO Ventures, Beijing Haisheng Group, Xinshi Era, Renai Capital followed the investment, as well as the old shareholders Shuimu Tsinghua Alumni Seed Fund and Jiushang Capital made additional investments. Maple Pledge Fengcheng Capital has been providing long-term financing advisory services. The funds raised will be mainly used for R & D investment, product iteration and market expansion.

Qianjue Technology is a high-tech company focused on the research and development of the embodied brain, benchmarking the American startup Physical Intelligence, and is committed to building an autonomous decision-making robot agent to help robots move towards general scenarios.

In April 2024, Qianjue Technology released a product-level robot perception and decision-making large model. Up to now, Qianjue's large model has been implemented in multiple scenarios, including home, commercial and industrial service scenarios, and has collaborated with several Internet giants and 3C giants.

"The company's second-generation embodied brain has completed training, and the developer version model has been opened; the commercial version model product will soon be open to the entire industry to help robot users in more extensive forms and richer application scenarios to use," Qianjue Technology founder and CEO Gao Haichuan told 36Kr.

The embodied intelligent brain is a key element to promote robots to enter general application scenarios. Robot expert Wang Tianmiao once said in an interview that the key to the development of embodied intelligence lies in the emergence of a general basic large model for robots. When the "general brain" appears, robots can be efficiently generalized and transferred for application. Currently, technology companies including OpenAI, Google DeepMind, and Physical Intelligence are actively deploying in this field.

However, the embodied intelligent brain is not simply a "shell" and optimization of a multimodal large model - autonomy and universality are the core elements of robot applications, and generative AI has obvious defects in these two aspects.

Gao Haichuan said that the vast majority of AI large models use the base Transformer training method, and due to the limitations of the relevance reasoning model, it is difficult to perform accurate causal reasoning and task decision-making.

The solution of Qianjue Technology is to use the decision-making mathematical framework of general robots and the computer mechanism inspired by the brain to create a large robot decision-making model, and to design a decision-making model architecture that is different from the one directly constructed using the base Transformer.

"Only by solving the common hallucination problem of the large model can it be possible to create a product-level robot perception and decision-making large model, achieve continuous decision-making and continuous observation of the environment, and realize closed-loop rolling decision-making, instead of always relying on human instructions," Gao Haichuan said.

It is reported that the core technology research and development of Qianjue Technology is supported by the Brain-like Center of Tsinghua University and the VIPLAB of the Department of Automation, applying the "brain-like" technology route to the decision-making process of robots.

Since 2018, the founder Gao Haichuan has been the leader of the Double-Arm Robot Research Group of Tsinghua University, leading the Brain-like Double-Arm Robot Research Group team; the chief technical advisor, Professor Chen Feng, is one of the first batch of brain-like computing research experts in China, undertaking several major projects of the China Brain Project. The main body of the technical team is composed of masters and doctoral students from the Department of Automation of Tsinghua University, with many years of first-line scientific and technological research and development and difficult technical project攻克 experiences in the field of artificial intelligence and robotics. At the same time, Qianjue is also responsible for an important sub-project of a national-level brain-like large project.

In terms of hardware, Qianjue Technology has created a domestically produced robot-exclusive end-side software and hardware integrated computing solution "Brain Dock", which supports the reasoning of the 13B embodied large model, can be compatible with both the Qianjue Brain and the cerebellum model of partners, and also achieves low cost and low power consumption.

The related research results were recently officially released at the top global AI conference ICLR 2024. At the same time, this achievement is also one of the core achievements of the "China Brain Project". According to Gao Haichuan, the spatio-temporal approximation (STA) method proposed in the paper realizes for the first time the direct conversion of the pre-trained Transformer into a brain-like network without the need for re-training.

It is worth mentioning that the embodied intelligent brain of Qianjue Technology has achieved generalization across forms, environments, tasks and objects, and is not only applicable to humanoid robots.

"The scenarios where Qianjue first landed and the customers we collaborated with are mostly robots with simple skills and can be quickly implemented, such as sweeping robots, delivery robots, industrial robots, etc. But even the previous generation of operational skills based on planning algorithms, after applying the Qianjue Brain, can show task generalization that was not previously available, and the product value is enhanced. We are currently in-depth cooperation with several industrial shareholders to accelerate the commercialization of the product," Gao Haichuan told 36Kr.

In the next step, Gao Haichuan said that Qianjue Technology will focus on accelerating market expansion, promoting the large-scale application of downstream cooperative customers, actively expanding new application scenarios, enabling robot manufacturers to create greater value, and will release some product-level industrial cooperation results in the near future.