What are the future applications of general-purpose humanoid robots? How far is it from industrial application? The reporter interviewed a number of experts. In their view, this kind of intelligent terminal will become a hot track in the artificial intelligence and robotics industry, which requires collaborative innovation of government, industry, university and research. The government can rely on new research and development institutions to establish a humanoid robot "atomic action" training platform and data platform, and promote industrial ecological innovation and product application through open source sharing.
Build a developed "cerebellum" and "brain"
Why are robots designed to be humanoid? Gu Jie, founder and CEO of Fourier Intelligence, explained that compared with other forms of robots, humanoid robots can directly use tools invented by humans; Human forms, on the other hand, appear amiable and confer high "emotional value" on the user.
The industry generally believes that "embodied intelligence" will become the next wave of artificial intelligence. The so-called "embodied intelligence" is to combine multi-modal large models such as language, vision and speech with robots, so that the artificial intelligence "brain" has a disposable, perceivable, interactive and actionable "body", so as to provide more intelligent services for humans.
The preferred carrier to realize "embodied intelligence" is humanoid robot. Gu Jie believes that they are expected to be used in areas such as elderly care, domestic service, welcome reception, and dangerous place patrol, covering many application scenarios in life and work.
"We are improving the level of motor intelligence of the humanoid robot, making its' cerebellum 'more developed, and learning to run and jump earlier." Gu Jie said, "The next step is to deploy a large model 'brain' to achieve the robot's vision, hearing, touch and other multi-sensory reasoning capabilities."
Zhiyuan robot's "Expedition A1" was also unveiled at the Shanghai Science and Technology Innovation Achievement Exhibition. Peng Zhihui, co-founder of the company, revealed that the next generation of commercial-oriented "Expedition A2" is under development, and will use a number of independently developed key core components, which will eliminate the stability and reliability defects of "Expedition A1" and greatly improve the movement ability of the robot.
The challenge of industrial application
On the journey towards industrial application, what challenges are Fourier Intelligent, Zhiyuan robot and other enterprises facing? Gu Jie said that in the humanoid robot industry chain, Fourier belongs to the intelligent terminal machine enterprise, its upstream is a parts supplier, the downstream is the use of GR-1 and other machine open hardware and software interface, the development of special robots for various application scenarios. In the future, the downstream ecosystem of the industrial chain will become larger and larger. To this end, he suggested that the government, on the one hand, promote industry-university-research cooperation to achieve breakthroughs in a series of key core technologies for humanoid robots as soon as possible; On the other hand, drawing on the experience of China's new energy automobile industry support policy, research and formulate policies to cultivate the humanoid robot industry and support the application of products, so that such robots are easy to use, durable and affordable for the common people.
"At present, the domestic industry chain is not perfect, for example, the robot radar is used for car radar, the camera is used for household products, and the robot arm is from industrial robots." "These parts will have to be customized in the future to better meet the development needs of humanoid robots," Gu said.
Talking about the technical challenges of industrial application, Peng Zhihui believes that the system design of the whole machine, the core components such as joint motors and dexterous hands, the whole body motion control algorithm, the real-time operating system, the general generalization of perception ability, the dialogue understanding and the action arrangement based on the large language model... These technical points are all directions to be explored, which require enterprises to have a full-stack technical perspective and a firm spirit of exploration, as well as government project support and industry-university-research cooperation.
Build a universal robot platform
In terms of the integration of large models and robots, Song Haitao, chairman of the General Robot Industry Research Institute, introduced that the general robot model can be divided into three types - multi-modal perception model, motion control model and autonomous task execution model. "Whether these models can be successfully deployed to the robot side, so that they have developed brains and strong executive power, requires a lot of pre-training research."
He further explained that the various actions of humanoid robots when performing tasks can be broken down into a series of "atomic actions", such as grabbing tools, pressing buttons, and moving items, and after the robot has mastered these "atomic actions", it can complete the task under the command of the "brain" through arrangement and combination. Therefore, new research and development institutions can establish an "atomic action" training platform with the support of the government, conduct large-sample training for humanoid robots in various application scenarios, and use the big data generated by training to establish a data platform.
In terms of common technology research and development, Yan Weixin, deputy director of scientific research of General Robot Industry Research Institute and doctoral supervisor of Robot Research Institute of Shanghai Jiao Tong University, said that the research Institute will collaborate with other innovative subjects to target key core components such as robot end computing platform, operating system, high burst actuators, and haptic sensors. To provide technical support for domestic general humanoid robots.