From transporting heavy objects to serving tea and water, from working in factories to cooking and folding clothes... When large models inject "soul" into robots, the evolution of humanoid robots accelerates. Industry insiders point out that the landing of humanoid robots in industrial scenarios can be achieved in some pilot projects as early as this year or next year, while "entering households" will take about 10 years. Changes - Entering the "Humanoid Era" Recently, the domestic humanoid robot field has taken frequent actions: the world's first pure electric drive full-size humanoid robot "Tiangong" has achieved "humanoid running" for the first time at the Beijing Humanoid Robot Innovation Center, the humanoid robot Unitree G1 targeting the elderly care market has been launched, and the industrial version of the humanoid robot Walker S has entered the NIO assembly workshop for "practical training" and become a "trainee factory worker"... Not only in China, humanoid robots that have taken the lead are sweeping the global primary and secondary markets. From startups to tech giants, from software platforms to hardware development, everyone is competing to enter. After NVIDIA, Tesla, and Microsoft all expressed their optimism with practical actions, Goldman Sachs adjusted its expectations for the global market size of humanoid robots by 2035 from $6 billion to $38 billion. "A track with a scale of billions of yuan requires three indispensable factors: market demand driven, complete industrial chain, and breakthroughs in core technologies." Jiao Jichao, Vice President of Ubisoft and Executive Director of the Research Institute, said that today, with the acceleration of the implementation of "robots+big models", the enthusiasm for technology, capital, and market to achieve embodied intelligence has stirred up a vast space for imagination. The combination of an aging society and a society with fewer children, as well as the relative shortage of labor, has stimulated a practical demand for humanoid robots in society. "In the manufacturing industry (such as automobile assembly, component sorting) and special operations (such as high-altitude operations, dangerous tasks), humanoid robots can complete high-risk, repetitive dirty, hard, and tiring work, reduce the injury and death rates of related positions, and improve occupational satisfaction. Therefore, the market is full of expectations for the arrival of humanoid robots." said Wang He, doctoral supervisor at the School of Computer Science at Peking University and director of the Peking University Galaxy General embodied intelligence joint laboratory. The reporter arrived at the Beijing Humanoid Robot Innovation Center and met a tall robot climbing stairs. This humanoid robot named "Tiangong" not only has a smart brain and dexterous hands, but also has a slight "softness" in its smooth movements. Even its appearance is close to that of humans: it is 1.63 meters tall and weighs 43 kilograms. When the engineer gave instructions to "her", he saw "Tiangong" leaning his upper body slightly forward, his arms cooperating with his legs to increase the swing, and he took small steps to run. Xiong Youjun, the general manager of the Beijing Humanoid Robot Innovation Center, who has been deeply involved in robotics for more than 20 years, believes that the future form of universal robots will definitely be humanoid. In the past, traditional specialized robots that only engaged in a single operation, such as welding robots and polishing robots, were essentially automated devices that only needed to execute fixed programs written in advance by humans. Therefore, in form, they only needed one leg or one arm. After injecting 'soul' into the robot through the large model, the robot has gained unprecedented perception, understanding, reasoning, and decision-making abilities, and can complete diverse tasks in complex environments. On the one hand, the infrastructure, scene construction, and tool design of human society are all developed based on human physical characteristics. Humans can reduce adaptation costs and quickly adapt to the environment; On the other hand, the ultimate goal of future universal robots is to enter human life and serve humanity, with 'humanoid' being the most natural and experiential form of interaction with humans Xiong Youjun said. Evolution - Achieving the Unity of Knowledge and Action: What is the Wonder of this generation of humanoid robots? Compared to the previous robotic arm, the domestically produced humanoid robot Walker S, which has been "trained" in the parking lot, provides the answer with a silky motion: her agile hands align with the front of the car to accurately affix the logo, her agile feet drive the torso to coordinate and automatically move to each seat for seat belt tension testing, and her "smart eyes" can also align with the car door and body surface for defect detection. The breakthrough of this generation of humanoid robots is the integration of knowledge and action. Manifested as a 'smart brain' commanding 'agile hands' and' agile feet ', making robots more autonomous and flexible in action, and able to continuously evolve in learning Ji Chao, Chief Scientist of iFlytek Robotics, said. Why are "smart minds", "agile hands", and "agile feet"? On the one hand, the iteration of large models makes the 'brain' smarter; On the other hand, the improvement of the software and hardware capabilities of humanoid robots makes their movements more flexible. Two forces are driving the evolution of humanoid robots Ji Chao said. It seems that the big model has a body, and the robot has a brain. Looking inside, it is actually the perceptual and reasoning abilities of the large model, integrated with the embodied abilities of the robot. How are these human like abilities achieved? "Derived from three parts: perception, planning, and execution." Wang He said that from a technological implementation perspective, if we compare humans, it not only requires the core brain of human intelligence, but also relates to the cerebellum responsible for motion control. The "brain" of humanoid robots is responsible for perception, understanding, and planning. "When we say 'I'm thirsty' to a robot, the robot first understands and breaks down the intention, and then plans and designs in steps: for example, the first step is to find where the water is, the second step is to think about how to open the refrigerator door, take out the water from the refrigerator, close the refrigerator door, and the third step is delivered to us. Each step, including understanding semantics and task planning, is carried out in the 'brain'." Wang He gave an example. The mission of the humanoid robot "cerebellum" is motion control and task execution. Wang He further explained, "For example, how to find the location of the refrigerator, avoid roadblocks, use a certain amount of force to reach out and pull the door, open the refrigerator door to a certain extent, and how to mobilize the limbs and torso to achieve coordinated movements of hands, eyes, and feet during each step of walking are the current challenges in research and development." "On the technological route, today's humanoid robots are becoming increasingly close to the process of human thinking, decision-making, and execution." Xiong Youjun said that in the past, cognition was achieved through voice interaction, visual recognition, natural language understanding, or traditional artificial intelligence methods. Now, multimodal large models such as VLA (Visual Language Actions) are used to complete "brain" functions. ". In the past, motion control was achieved through model optimization, but now the running and jumping of humanoid robots are more based on reinforcement learning and imitation learning to achieve the functions of the "cerebellum" and "trunk". The revolutionary change brought to robots by the iteration of large models is that they are not only applied to the understanding, reasoning, and planning of the brain, but also integrated into the perception and execution processes involving the cerebellum and torso. After possessing stronger generalization ability, humanoid robots will better interact with the physical world, further reducing deployment costs and application barriers, which is also the direction for the advancement of embodied intelligence. "Embodied intelligence emphasizes the interaction and feedback between intelligent agents and the physical world, serving real-life scenarios and addressing human practical needs," said Jiao Jichao. Evolution - "Entering the Home" for ten years, humanoid robots can imitate and replicate human intelligence with the power of the wilderness, but they are ultimately not humans. The Moravik paradox points out that the unique high-order intelligence of humans only requires very little computation for machines, such as reasoning, while the unconscious skills and intuition of humans require machines to expend a great deal of computational power, such as catching a parabolic object. "At present, the bottleneck of humanoid robots' abilities is not the perception, understanding, and judgment of the 'brain', but rather the lack of generalized execution ability similar to the human 'cerebellum'." Wang He said that the problem caused by this bottleneck is that humanoid robots are not fast enough and not agile enough to work. The root cause of this bottleneck is the lack of high-quality data. "The data required by the robot 'brain' can be obtained from the texts and images on the Internet, while the data required by the 'cerebellum' is very little. One of the limited sources is the manual remote control machine training method adopted by some foreign enterprises, and the data collected is difficult, expensive, and slow." Wang He said. High precision simulation of real-world physical scenes and synthesis of data in simulation environments can be an effective solution. For example, synthesizing dexterous hands to simulate real-world object grasping for training and collecting data, and then testing and generalizing it in a real environment, can solve the data feeding problem required by the robot's cerebellum. Experts also point out that real-time performance and inference speed are another bottleneck currently faced by embodied intelligence. At present, online decision-making for robot control based on large models requires at least 50 milliseconds, with a general delay of 1 to 5 seconds, which cannot meet the needs of 3 to 10 milliseconds in industrial scenarios. The humanoid robot is a hard nut to crack, and its breakthrough requires time accumulation and technological accumulation. Mass production and commercialization are also challenges Multiple respondents stated that although the growth rate of humanoid robots is visible to the naked eye, there is still a long way to go before the "iPhone moment" of the industry explosion. It can be confirmed that humanoid robots were the first to be implemented in industrial settings, followed by their application in commercial settings, and finally entered thousands of households. Among them, the landing of industrial scenes can be achieved in some pilot projects as early as this year or next year, and the "entry" of humanoid robots into households can be achieved in about 10 years Xiong Youjun said. Jiao Jichao provides a set of data: Currently, China has the highest number of patent applications for humanoid robot technology in the world, with a total of 6618 cases. "This will lay a solid foundation for the steady development of China's humanoid robot industry." "When can we achieve a 'brain like model', put all modal data into a dataset, respond quickly enough, understand semantics, and interpret intentions, and truly achieve embodied intelligence? This is also the problem we will focus on solving." Wang He said. (Lai Xin She)