The following article is from Intelligence Emergence, author Yang Xiao

Since 2023, embodied intelligence has been a hot topic that has attracted much attention.

In the view of Sun Teng, CEO of Ruoyu Technology, robots are the most ideal landing carriers and application platforms for large models. The introduction of large models has completely changed the dilemma that robots can only perform a single task in a single scenario, significantly improving the intelligence level of robots, making them smarter and able to undertake more complex tasks.

At present, some robot companies choose to load external large models (such as calling Chatgpt) directly on the robot. However, these large models are not trained with specific robot-related data for robot embodied intelligence, which may cause the large model to be incompatible with the robot execution code and the external environment, thereby reducing the instruction decomposition ability and execution success rate. At the same time, large models loaded in the cloud will also bring about latency problems, restricting the domain adaptability of multi-type robots driven by large models.

Ruoyu Technology is committed to the research of universal robot brains, and hopes to cooperate with robot manufacturers such as industrial robotic arms and service robots to provide advanced robot brain products.

Ruoyu Technology's research on the general robot brain based on multimodal large models mainly focuses on two directions: one is the perception model responsible for the perception level, and the other is the control model at the control level. The perception model obtains external environment information through voice interaction, visual information sensor signal extraction, etc.; after perceiving the environment, the control model converts the received demands into control instructions that the robot can operate and execute, such as instruction disassembly, code generation and other control-level tuning work.

Ruoyu Technology and Harbin Institute of Technology jointly developed the language large model base - Lizhi and the multimodal large model base - Jiutian, with a total parameter scale of 13 billion and training data exceeding 1500B tokens. It can complete multiple types of instruction data such as multi-task, multi-round dialogue, Chinese-English translation, thinking chain, tool use, etc., and can show good results in logical reasoning, relational reasoning, and perception ability.

Based on the self-developed language-based big model and multimodal big model, Ruoyu Technology has built a multimodal embodied decision-making big model. With the universal knowledge of the human world and the powerful proprietary code generation capabilities in the field of robotics, it can achieve a deep understanding of the external environment and intelligent control of various types of robots. In terms of product form, Ruoyu Technology provides two types of standardized products for robot manufacturers: one is the software form of SDK, which is loaded on the public machine on the robot side or deployed locally; the other is to combine the robot brain with the computing chip to provide a hardware and software integrated board solution.

The robot brain developed by Ruoyu Technology can be directly deployed at the internal development end of the robot architecture, and communicate directly with the robot's application layer and middleware. By automatically generating code, it reduces the development workload of traditional robot manufacturers and simplifies the development difficulty, which helps traditional robot manufacturers develop various different application functions.

△ Ruoyu Jiutian robot brain architecture

Ruoyu Technology has chosen to cooperate with leading server manufacturers to cover 70% to 80% of scenarios in different fields through the core needs of server manufacturers. In the future, the company plans to further increase the functions required by customers when cooperating with mid- and long-tail customers.

Ruoyu Technology has completed the research and development of a base model with 13 billion parameters, and has made considerable progress in the design of the robot brain architecture and product design and development. The company is developing functions for robots covering different fields and different forms, and expects to complete the development of the robot brain in the first half of 2024.
In addition, the company will open a free robot brain SDK in the second quarter of 2024 to better understand market demand and promote iterative products.

At the hardware level, Ruoyu Technology will also strengthen the connection between different forms and interfaces in the robotics field, realizing the transition from executing instructions at the data interface to collaborative work with different hardware manufacturers.

As for the founding team, the founding team of Ruoyu Technology comes from the School of Computer Science and Technology of Harbin Institute of Technology (Shenzhen). The team has dozens of national high-level talents in multimedia computing, natural language processing, embodied intelligence and large model quantization compression, and relies on Shenzhen Ha Shen Asset Management Co., Ltd. to transform scientific and technological achievements. Sun Teng, the company's co-founder and CEO, is a doctor in artificial intelligence and has continuous entrepreneurial experience.

Message from Professor Zhang Min, Chief Scientist of Ruoyu Technology:

In the development of large models, the importance of data cannot be ignored. The distribution, quality and input order of data will affect performance. Secondly, he emphasized the importance of personally participating in the development of large models, believing that only by trying it yourself can you better grasp the original innovation. Therefore, he suggested that every university teacher should personally train a large model, even if it is small in scale, with only 100 million parameters. Analogous to the principle of manufacturing an engine, even if others make it well, it is crucial to personally experience the process.

Modified on December 27, 2023. If there is any infringement, please contact us to make modifications.