Xinhua News Agency reporter Peng Qian Zhang Manzi
After the big model, "embodied intelligence" has become a new hot spot in the science and technology industry this year and is considered to be the key direction in the new wave of artificial intelligence (AI). Many start-ups have emerged, financing has repeatedly hit new highs, and technology has continued to break through... As the most representative entity in this field, humanoid robots are accelerating industrialization under the catalysis of big models.
If the big model is regarded as an "interesting soul", the humanoid robot enabled by "embodied intelligence" has a "good-looking skin" and has become a powerful assistant to humans in many fields.
Redefining robots
As a bridge connecting virtual space and real space, "embodied intelligence" refers to the integration of AI into physical entities such as robots, giving them the ability to perceive, learn and interact dynamically with the environment like humans.
The term "embodied intelligence" itself has a strong philosophical color of technology. In 1945, French philosopher Maurice Merleau-Ponty proposed the concept of "embodiment", believing that humans need to interact and perceive the surrounding environment through their bodies to understand the world. In 1950, British computer scientist Turing, known as the "father of AI", first proposed the concept of "embodied intelligence" in his paper "Computing Machinery and Intelligence".
In fact, industrial robots (robotic arms) with relatively low intelligence levels have long been widely used in the manufacturing industry, bringing about quality and efficiency improvements. However, traditional industrial robots are a combination of "fixed programs + robotic arms", while robots empowered by "embodied intelligence" are an iteration of "multimodal perception + brain decision-making".
Xu Huazhe, assistant professor at the Institute of Cross-Disciplinary Information Sciences of Tsinghua University, believes that robots in the future will present a variety of forms: bipedal, quadrupedal, wheeled, mechanical dogs, intelligent drones and even mechanical bees, but humanoid robots are the most adaptable to human society and will become the robots that can best help humans.
Humanoid robots can solve the problem of the "last mile" of the production line. Many personalized and customized products cannot be assembled uniformly on the assembly line, which requires humanoid robots with generalization capabilities to "help" assemble mass-produced parts into products according to customer customization needs. In more complex and changeable scenarios such as home services and public services, humanoid robots also have more advantages and can adapt to different environments and needs to complete a variety of tasks.
Three major difficulties to be overcome
The development of humanoid robots begins with learning and imitating humans. The difficulties in its development can also be understood by comparing it with the human brain, cerebellum and body. The "brain" is mainly responsible for the robot's autonomous learning, planning and decision-making center; the "cerebellum" is responsible for motion control, including walking to running and jumping, and from simple grasping to complex hand movements; and the "body" includes the trunk and limbs structure and dexterous hand design.
Jiao Jichao, vice president of UBTECH Robotics and director of the research institute, told reporters that there are many technical difficulties to be overcome in these three major areas: in terms of the "brain", cloud-edge-end integrated computing architecture, multimodal perception and environmental modeling are the focus of technology in recent years. "The biggest difficulty in human imitation lies in the imitation of the human brain, and the existing scientific theories are far from enough to study the human brain"; in terms of the "cerebellum", human-computer interaction capabilities, complex terrain passage, whole-body coordinated fine operations, etc. are important directions; in terms of the "body", key technologies such as rigid-flexible coupling bionic transmission mechanism, highly compact robot limb structure and dexterous hand design are important hardware foundations required for the flexible movement of humanoid robots.
The emergence of large models has significantly "evolved" the robot's "brain", greatly improved the versatility and generalization of robots, and is expected to reduce the development cost of humanoid robots and accelerate their entry into thousands of households.
According to Yang Fengyu, founder and CEO of Uniqlo Robotics, the industry now mainly uses pre-trained large models to pre-train robots, giving them stronger learning capabilities; large models can transfer learning of specific tasks to robot tasks to improve their adaptability; in addition, the multimodal processing capabilities of large models can be used to combine various inputs such as vision, hearing, and touch to improve the robot's understanding of complex scenes.
China is not behind in the start
Looking around the world, humanoid robots have entered the initial stage of industrialization and have begun to "test the waters" in the fields of industrial manufacturing, commercial services, and family companionship. Whether it is technological breakthroughs, progress in implementation, or financing scale, the humanoid robot R&D competition is basically dominated by China and the United States.
UBTECH's humanoid robot Walker focuses on key manufacturing fields such as automobiles and consumer electronics this year, and has entered many car factories for practical training; Yushu Technology's robots have achieved completely human-like natural walking; Uliqi Robot is brewing a "home" plan; Boston Dynamics' new version of Atlas robot has achieved flexible movement of parts between different lockers in the factory; Tesla's humanoid robot "Optimus Prime" plans to start mass production in 2025...
Jiao Jichao said: "If the humanoid robot industry is likened to a marathon, China and European and American countries are currently almost at the starting stage of the first 1,000 meters."
Yang Fengyu holds the same view. The leading large-model R&D capabilities and perception technology make American companies more competitive in robot decision-making systems and complex task processing. China's advantages are more reflected in the field of industrial robots, especially in the relatively mature application of manufacturing. In terms of humanoid robot technology patents, China has also taken the lead.
Although the core algorithms and high-end chips of the robot's "brain" still face challenges, the Chinese robot industry has a wealth of application scenarios and a large potential user population, and data is one of the biggest competitive advantages.
The key is how to collect a large amount of data in real-world scenarios and coordinate the circulation and supply of data to the humanoid robot industry. Xu Huazhe said that real-world data is generally collected by various robot companies themselves. Beijing Humanoid Robot Innovation Center Co., Ltd. is planning to create an open source data set for use by academia and industry. In the future, high-quality shared data sets will greatly help the development of the industry.