Qizhi Research Institute aims at embodied universal artificial intelligence | Enters new research and development institutions, humanoid robot "Little Star" emerges institution | Chen Jianyu | Research Institute

Editor's note: The new research and development institution is a product of various provinces and cities in China in promoting the layout of scientific and technological forces and the reform of scientific and technological systems and mechanisms. It has the characteristics and advantages of new conceptual ideas, new management methods, and new funding models. In recent years, Shanghai has promoted the construction of a number of strategic and forward-looking new research and development institutions based on national and local scientific and technological innovation strategies and task requirements. The goal proposed in the "Opinions on Promoting the High Quality Development of New R&D Institutions in Our City" is to focus on cultivating 20 high-level new R&D institutions with international influence by 2025, and guide social forces to build 200 new R&D institutions.

Where are the new research and development institutions, such as Shanghai Qizhi Research Institute, Shanghai Tsinghua International Innovation Center, and Shanghai Shutu Blockchain Research Institute? What technological innovation achievements have emerged through institutional innovation? Jiefang Daily's Shangguan News reporter recently visited multiple institutions to understand their reform and exploration processes.

"Hello, Xiaoxing, I'm at the gate of the second school. Can you help me get a bachelor's hat? Hurry up," Chen Jianyu, an assistant professor at the Institute of Cross Information Studies at Tsinghua University and a PI at Shanghai Institute of Intelligence, showed reporters in a video. The cute humanoid robot "Xiaoxing" can walk on both feet and in the future, it can help people complete various tasks such as picking up items and doing barbecues.

Recently, this robot made its debut at the 2023 World Artificial Intelligence Conference, attracting the attention of many professionals in the technology and investment sectors. Turing Award winner and Academician Yao Qizhi, President of Shanghai Qizhi Research Institute, stated at the conference that the next major challenge in the field of artificial intelligence will be to achieve "embodied universal artificial intelligence", which is to study how to build high-end robots that can master various skills through self-learning and perform universal tasks in real life.

As a new type of research and development institution, Shanghai Qizhi Research Institute provides funding, talent, and achievement transformation support to young scientists such as Chen Jianyu, creating a better environment for them to engage in basic research such as artificial intelligence. "I have teams at both Tsinghua University and Qizhi Research Institute, and the full support of the research institute will accelerate our research and development process," said Chen Jianyu.

"Xiaoxing" has attracted onlookers at the World Artificial Intelligence Conference.

Trust scientists and grant research autonomy

Shanghai Qizhi Research Institute is a new type of research and development institution with the nature of a public institution. It implements the "three no's mechanism", which means that the institution specifications are not clear, the establishment of the institution is not verified, and it is not limited by job positions and total wages. Compared to traditional public institutions, this research institute has greater flexibility in the use of research funds. "We implement negative list management to support the Jizhi Research Institute in focusing on established goals and tasks. According to the actual needs of scientific research activities, financial funds will be used for personnel, equipment, experiments, exchanges, and other aspects. The specific usage will be decided by the research institute independently." said Song Yang, Director of the Basic Research Department of the Municipal Science and Technology Commission.

This autonomy gives scientists full trust and also reflects the government's support for exploring basic research paths beyond competitive technology projects.

Chen Jianyu is a beneficiary of this institutional innovation. At the Shanghai Qizhi Research Institute, his team has received stable project funding support, which basically guarantees scientific research needs. Therefore, the time spent applying for competitive projects can be used for basic research. The project application process of Shanghai Qizhi Research Institute is streamlined, with a focus on assessing whether the research interests of the PI team are consistent with the main development direction of the institute, and whether they have a certain research foundation. The application and approval period for the project is relatively short, and the use of funds adopts a lump sum system, allowing the PI to adjust the funds for each subject based on the actual research situation. In the evaluation process, the research institute fully considers the characteristics of basic research, conducts differentiated performance evaluations of various PI teams, and pays more attention to the innovation and international influence of scientific research achievements.

The humanoid robots in Shanghai Qizhi Research Institute

Speaking of his own topic, Chen Jianyu said, "Universal artificial intelligence is a human dream. It can do many things like humans, no longer limited to certain specific tasks. ChatGPT can only process textual information, and we believe that the future of universal artificial intelligence requires embodied physical entities with eyes, ears, hands, and feet to better interact with humans in the physical world."

Software and hardware breakthroughs make robots flexible and powerful

This young scientist believes that the best carrier for "embodied universal artificial intelligence" is humanoid robots, because human society is built by and for humans. After possessing the form and basic functions of the human body, robots can easily enter various human work and life scenes, walk or run with both legs, use tools with both hands, and perform fine operations with fingers.

In the future, when humanoid robots equipped with artificial intelligence large models are put into commercial use, our society will undergo tremendous changes - robots can perform tasks that only humans could do in the past, they are not fatigued, and their focus will not decrease; They will also enter households and engage in services such as housekeeping, nursing, and education.

Humanoid robots will enter homes.

In the face of a bright future dream, the research and development plan for the humanoid robot "Xiaoxing" has emerged. Chen Jianyu introduced that this type of robot has very high technical requirements at both the hardware and software levels, integrating almost all robot functions such as walking, mobile navigation, and hand operation together, and they need to cooperate with each other.

For example, in terms of hand and leg coordination, when an artificial neural network issues control commands for both hands and legs simultaneously, the robot's hand movements may affect the walking of both legs, leading to falls. To this end, he led the team to tackle software issues and developed a decentralized robot control method. It can automatically analyze commands issued by neural networks, decouple hand and leg motion commands, but does not decouple left and right leg movements that require coordination. In this way, the robot can perform hand operations while walking.

At the hardware level, he led the team to adopt a new technological route of "ontology perception driver", combining high torque density motors with low reduction ratio reducers, current loop force control, and dynamic bipedal walking algorithms, making the robot have high performance such as strong force and flexible movement, and lower manufacturing costs. "The motor torque density is an important indicator for evaluating the hardware performance of humanoid robots. This high indicator indicates that the robot has great power and can move flexibly." Chen Jianyu said, "Currently, the academic community has done the best in this indicator, with UCLA and MIT reaching 38 Nm/kg. The torque density of 'Xiaoxing' is 30 Nm/kg, and the second generation 'Xiaoxing' we are developing will reach 34 Nm/kg."

Medium - and long-term evaluations allow scientists to devote themselves to research

At the 2023 World Artificial Intelligence Conference, the appearance of "Little Star" attracted the attention of many attendees. With a height of 1.2 meters, a white body, and a walking speed of 0.7 meters per second, this robot is cute and adorable.

Academician Yao Qizhi and "Xiaoxing" take a group photo.

This is the second-generation "Little Star" humanoid robot. The Chen Jianyu team is rapidly iterating software and hardware, and the future "Little Star" will be taller and walk faster. With the support of computing power from Shanghai Qizhi Research Institute, they conducted physical simulation experiments, which showed that the running speed limit of "Xiaoxing" could reach 4 meters per second.

By imitating human actions through reinforcement learning, this robot can not only learn to walk, but also learn to cooperate with both hands to complete tasks, such as grabbing a ball with the left hand and handing it over to the right hand, and lifting a pot with both hands.

Reinforcement learning solves the execution layer problem of humanoid robots, while task planning as the decision-making layer is handled by a large language model. After combining humanoid robots with large language models, a challenge arises - tasks planned by large models may not be executed well. To this end, the Chen Jianyu team has added an automatic monitoring system between the decision-making and execution levels. This system consists of a visual language model that can automatically monitor and infer task execution status. For example, when performing a box moving task, if the box in the robot's hand falls, automatic monitoring can detect it and issue a "pick up box" command to the robot.

They also improved the large language model itself through training, enabling it to learn active questioning to improve the efficiency of robot task execution. For example, when a robot is asked to look for something in an unfamiliar room, it will ask the owner of the room, "Do you know where this thing is?" instead of just going into the room and flipping through boxes and cabinets.

It is reported that Shanghai Qizhi Research Institute adopts a medium to long-term comprehensive evaluation mechanism with innovation performance as the core, and the mid-term assessment period of the project is 3 years. This evaluation mechanism has given young scientists such as Chen Jianyu ample time to devote themselves to research, laying a solid scientific foundation for the future entry of humanoid robots into homes and workplaces.