Are robots necessarily humanoid? Experts and scholars discuss the development path of embodied intelligence

On August 8th, the "2024 Technology Innovator Conference" hosted by Zhongguancun Zhiyou Research Institute and jointly organized by Zhongguancun Dongsheng Science and Technology Park and Lide Robot Platform was held at Zhongguancun Dongsheng Science and Technology Park in Beijing. Scientists, entrepreneurs, entrepreneurs, and investors both domestically and internationally are discussing the future development path of embodied intelligence.

"At present, the emergence of intelligence is just a 'static integration', waiting for the 'dynamic emergence' ability of the brain to achieve real AGI." Huang Tiejun, president of Beijing Zhiyuan Artificial Intelligence Research Institute and director of the State Key Laboratory of Multimedia Information Processing of Peking University, believes that in the next three years, the big model will be expanded to general intelligence such as vision, hearing, embodiment, and action, with larger scale and smaller size, and is expected to become the "brain" of embodiment intelligence; In the next 10 years, humanity will fully enter the era of intelligence, just like the industrial revolution liberates physical strength, just like the electricity revolution solves energy circulation, and the construction of technology, economy, and social ecology centered on intelligence will be completed; In the next 20 years, behaviorism (reinforcement learning) and neuromorphic intelligence (thorough connectionism) will play a greater role. In 2045, there will be the emergence of structure mimicking brains, functional brains, and performance super brains, and the arrival of physical AGI.

Peng Fangyu, Vice Dean of the School of Mechanical Science and Engineering at Huazhong University of Science and Technology and Chairman and General Manager of the National Digital Design and Manufacturing Innovation Center, said, "One of the important directions for future manufacturing is to achieve ubiquitous manufacturing through ubiquitous robots, ubiquitous sensors, and ubiquitous intelligence. The future trend of manufacturing is human-machine integration, which includes three levels: human-machine space sharing, human-machine skill transfer, and human-machine hybrid intelligence. The latter two levels will break through the level of robot intelligence and make robots become 'skilled craftsmen'

In the keynote report "Intelligent Control of Robots Based on Generative AI," Tencent Chief Scientist Zhang Zhengyou stated, "The current embodied intelligence is a combination of 'big models+robots,' which is artificially forced intelligence into the robot's body. In the future, we hope that intelligence and ontology can be organically combined, allowing robots to emerge with intelligence in their interaction with the environment

Turning to the evolution path of high-quality development of embodied intelligence, Wang Tianmiao, honorary director of Beihang Robotics Research Institute and president of Zhongguancun Zhiyou Research Institute, believed that "the path of sustainable development of embodied intelligence needs to iterate in the underlying innovation and application scenarios." There are two paths for innovation and entrepreneurship to cross the foam and adjust the cycle. One is for small companies with limited resources, starting from specific business needs, defining products with new technologies, focusing on a specific segment, and constantly improving their core competitiveness through the development path of "specialization, refinement, and novelty". This idea is easier to survive, and can also cooperate with large companies, or even sell to large companies, which provide application scenarios and financial support; Another approach is for platform companies with abundant resources to achieve breakthroughs in universal technology, which can be widely applied in several industries, generate revolutionary technological progress and industrial expansion, continuously raise funds, integrate upstream and downstream resources, achieve greater economies of scale and synergies, and thus occupy a favorable position as a leader or chain owner in market competition, but the mortality rate is also the highest.

Does the form of a robot necessarily have to be humanoid? Do we have to have limbs? Wang Yu, founding dean of the Robotics Research Institute at the Hong Kong University of Science and Technology and co-founder and chief scientist of Dai Meng Robotics, said, "In some industrial scenarios, we see that workers are basically sitting in fixed positions, and workpieces are (automatically) given to workers, so the requirements for 'legs' are quite low. More importantly, the arms, hands, and eyes can be used in conjunction with a pair of nimble hands to complete grasping, movement, and other actions, which may not necessarily be completely accomplished through 100% vision

Ultimately, when we need to consider every process and operation in the scene as a model, that is the cerebellar model. I narrowly call it embodied skills, which means it has the ability to independently complete precise operations and can acquire these skills through certain learning methods, "said Wang Yu.

In the theme sharing that embodied intelligence is the most effective way to achieve AGI, Wang Xingxing, founder and CEO of Yushu Technology, believes that "the deployment and data collection of physical robots are the key to embodied intelligence technology leading to AGI." In his view, there are four directions in which deep reinforcement learning can be done better in the present and near future, namely: imitating deep reinforcement learning for arbitrary body postures or dance movements; Under complex terrain, the whole body can move steadily and freely; Hand parts, complex items or components, requiring more complex and agile operations; Reinforcement learning navigation and obstacle avoidance based on perceptual data such as images or depth data.

From the perspective of technological exploration and global business practice, Shao Tianlan, founder and CEO of Mecamand, shared that "there are still three challenges for the mass implementation of intelligent industrial robots: new technology, high requirements, and long chain." However, technologies such as sensors and artificial intelligence are developing rapidly, and technology needs to be updated faster. Enterprises should form a positive cycle of leading technology products, leading business, and leading capital. Deeply cultivate core technologies, continuously iterate products, and expand applications with partners.

Dong Xiaojian, founder of Beijing Weijing Intelligent Technology Co., Ltd., believes that the development of embodied intelligence is a process of evolution from machines to humans, and from humans to intelligent agents. It emphasizes the combination of continuous learning and practice, especially through the extensive information collection and screening of multiple sensors (such as vision, touch, hearing, etc.), to build the perception and cognitive abilities of intelligent agents. This process is not limited to learning from digital samples, but also involves comprehensive processing, understanding, and decision-making of information, enabling intelligent agents to respond flexibly in practical environments.

In the view of Zhou Yong, CTO of Lingxin Clever Hand, "the next direction of embodied intelligence is multimodal perception and interactive algorithms. The 'nimble operation' in daily life will become a true symbol of the arrival of the embodied intelligence era

As a platform based technology, embodied intelligence is regarded as a potential key to solving the problem of efficient production of multiple varieties and small batches in the industrial field due to its strong universality. Experts attending the meeting stated that in order to achieve embodied intelligence technology from the laboratory to a wide market, especially in deploying billions of devices in "super scenarios", two core conditions need to be met: first, technological achievements must be oriented towards consumers (To C), ensuring the close integration of technology and daily life, thereby stimulating huge market demand; Secondly, it is necessary to achieve a comprehensive leap in accuracy and performance, ensuring that accurate and efficient services can be provided in various complex environments. This is the only way to achieve "super scenarios".

At the meeting, the preparatory launching ceremony for the "50 Young Scientists Forum" led by Zhongguancun Zhiyou Research Institute was officially held. The forum aims to build a platform for co creation, co construction, mutual trust, and win-win, providing strong support for the construction of the country's modern industrial system.