Want to use humanoid robots? First, pass four hurdles.
At the just - concluded 2025 World Artificial Intelligence Conference (hereinafter referred to as "WAIC"), the Embodied Intelligence Robot exhibition hall was the most popular. Robot dogs running around and robots of various shapes attracted audiences to stop and watch.
It is reported that there were more than 150 humanoid robots competing on the same stage at this year's WAIC, which was the largest - scale collective exhibition of domestic humanoid robots ever. Different from last year when most robots were presented in a static display, the robots exhibited this year mainly showcase their ability to work in real - world scenarios.
The exhibition booth of Zhipu Robotics covers an area of 300 square meters, which is also the largest booth among the Embodied Intelligence enterprises participating in this conference. The company set up five major commercial scenarios at the booth, including industrial logistics, commercial services, household use, scientific research and education, and cultural and entertainment performances. Different robots were "assigned" to show their capabilities: The general humanoid robot Yuanzheng A2, which focuses on interactive services, was responsible for booth explanations and provided self - guided tours without human intervention throughout the process; The Elf G1 served as the "all - around assistant" at the booth, delivering items to audiences and stamping commemorative seals; Another robot, Yuanzheng A2 - W, was diligently performing de - palletizing and handling operations, accurately grasping turnover boxes of different specifications.
Yinhe General built a 1:1 restored real - life supermarket scenario, and its robot Galbot was on duty around the clock.
Xingdong Jiyuan simulated a logistics and warehousing scenario to demonstrate the robots' ability to perform "real work" such as intelligent sorting, scanning codes, and providing services.
Goldman Sachs Research predicts that by 2035, the global humanoid robot market size is expected to reach $154 billion. In a more optimistic scenario, the market size could reach $205 billion. Morgan Stanley said in a research report that as the world's largest robot market and manufacturing center, China's robot market size reached $47 billion in 2024, accounting for 40% of the global total. It is expected to increase to $108 billion by 2028, with a compound annual growth rate of 23%.
On July 27, CITIC Think Tank released the in - depth research report "The New Era of AI: Forging Ahead and Spreading the Flame of Intelligence" on AI + Industry Development at WAIC. The report points out that currently, the training method of robot large - scale models has begun to develop towards end - to - end large - scale models + world models (physical rule modeling). At the same time, the modal capabilities are becoming more and more diverse. However, due to pain points such as the limited generalization ability of the robot "brain", the high difficulty of obtaining datasets, the need to optimize the structural design, and the high cost of the supply chain, it is still difficult for humanoid robots to achieve large - scale commercialization in the short term.
Why Humanoid Robots?
Humanoid robots are generally considered one of the most promising application directions of AI. In the future, they will have broad application scenarios in industries such as industry, commerce, household use, and outer - space exploration, liberating humans from low - level and high - risk industries and improving human productivity and work efficiency.
It should be noted that humanoid robots are just a typical carrier of embodied intelligence, but the scope of embodied intelligence is far beyond humanoid robots. The core of embodied intelligence lies in "generating intelligence through the dynamic interaction between the body and the environment", and the form of the "body" can be diverse. As long as it has the ability to perceive, act, and interact with the environment, it can become a carrier of embodied intelligence.
CITIC Think Tank pointed out in the report that the humanoid design is considered for various reasons.
Firstly, they can better adapt to the environment. In reality, many scenarios, such as stairs and tools, are designed according to human ergonomics. If a robot has a humanoid shape, it can be directly applied in these scenarios.
Secondly, they have better task versatility. The human body has a high degree of flexibility. By imitating human joints and sensory systems, humanoid robots can be more flexible and perform diverse tasks.
Thirdly, the human - like appearance reduces the psychological gap in human - robot interaction. Compared with other types of robots, humanoid robots will face lower barriers in emotional communication in the future.
Fourthly, the data availability is higher. Deep bionics means that a large amount of motion data can be obtained from human movements, which can significantly reduce the difficulty and training time required for training methods such as deep reinforcement learning.
Tesla's humanoid robot Optimus adheres to a human - like structure. The Optimus robot is 1.73 meters tall. It uses a spring design imitating tendons in the actuator connection, and its hand shape imitates the human palm, thus achieving high sensitivity to grasp heavy and thin objects.
According to the plan announced by Tesla in January this year, Optimus plans to mass - produce thousands of units in 2025 and conduct preliminary tests in Tesla's factories; Mass - produce 50,000 - 100,000 units in 2026 and increase the production by 10 times in 2027.
Since the beginning of this year, it has become a trend for automobile manufacturers to enter the humanoid robot field. According to an incomplete statistics by Caijing based on official announcements, currently, 15 domestic automobile manufacturers, including GAC, SAIC, BYD, XPeng, Chery, and Xiaomi, as well as supply - chain enterprises such as Huawei, Horizon Robotics, and Hesai Technology, have successively entered the humanoid robot track. Overseas, 6 automobile manufacturers, including Tesla, BMW, Mercedes - Benz, Honda, Toyota, and Hyundai, have also proposed their own ideas for humanoid robots. Most of the world's mainstream automobile manufacturers have basically entered the humanoid robot track (for related reports, see "21 Automobile Manufacturers Increase Investment in Humanoid Robots: The Next Ticket?").
In China, humanoid robots have also become one of the hottest areas in the capital market this year. On July 18, information on the official website of the China Securities Regulatory Commission showed that Unitree Technology started the listing guidance process, with a valuation of over 12 billion yuan. Unitree Technology was founded in 2016 and started with the research of robot dogs. It began to release humanoid robot products in 2023 and is one of the most - watched humanoid robot startups in China. The humanoid robot company Tashizhihang has started its third - round financing. In March this year, the robot company Vita Power announced the completion of a 200 - million - yuan financing. This company has completed two rounds of financing within 100 days of its establishment.
According to the valuation, currently, Chinese humanoid robot startups have formed distinct echelons. There are three companies in the first echelon: Unitree Technology, Zhipu Robotics, and Yinhe General, all with valuations exceeding 7 billion yuan. The second echelon includes more than a dozen companies such as Tashizhihang, Zibianliang Robotics, Zhujidongli, and Qianxun Intelligence, with valuations ranging from 2 billion to 5 billion yuan.
Commercialization Difficulties
Although the humanoid robot industry is highly popular, it still faces the industry problem of being difficult to achieve large - scale commercialization in the short term. CITIC Think Tank pointed out in the report that the resistance mainly comes from the following four aspects.
Firstly, the "brain" of the robot, that is, the generalization ability of the large - scale model of the embodied intelligence robot is limited, resulting in restricted current application scenarios.
The large - scale model of embodied intelligence refers to a general intelligent agent model that has the ability of environmental perception → task understanding → action generation through pre - training with a large amount of multi - modal data (vision, language, and action). It gradually realizes the leap from "single - instruction execution" to "autonomous decision - making evolution". Its core is to enable the robot to "understand the world, understand instructions, and take actions" like humans, promoting intelligent upgrades in scenarios such as industry, medical care, and household services.
CITIC Think Tank believes that the large - scale model of robots can be roughly divided into L1 - L5 stages, referring to the technological development of autonomous driving. Currently, it is in the L4 stage, which is characterized by the realization of real - world interactive feedback. The representative model is Gemini Robotics launched by Google DeepMind in March this year, which achieves a deep integration of the three modalities of vision - language - action (VLA) and has strong spatial reasoning ability.
Currently, the large - scale model of embodied intelligence is iterating rapidly, but it still faces major pain points such as insufficient datasets, slow thinking compared to movement, and lack of an ecosystem. In terms of effectiveness, humanoid robots have made good progress in the generalization ability of single tasks, but their generalization ability in multi - tasks and cross - tasks is still relatively weak, which affects the performance of humanoid robots in real - world application scenarios.
Data is the "cognitive cornerstone" of embodied intelligence. Currently, the amount of data required for the general robot large - scale model is extremely large, and it is difficult to obtain data. As a result, it is difficult for general robots to collect a sufficient amount of real - machine data for large - scale model training in the short term.
Different from large - scale language models that can directly learn a large amount of knowledge from the Internet and books, it is difficult to quickly improve the movement and planning ability of robots just by watching Internet videos or human movement videos.
Nie Kaixuan, the CEO of Songying Technology, a developer of physical AI simulation systems, told Caijing that humanoid robots need a large amount of data from interactive operations in the real physical environment to train their decision - making and action abilities. The shortage of operation data in the physical world limits the development of humanoid robots.
The structural design of robots still needs to be optimized, and problems such as the reliability of components and battery life need to be solved urgently.
In April 2025, the 2025 Beijing Yizhuang Human - Robot Half - Marathon was held at Beijing Nanhaizi Park. The half - marathon challenges the comprehensive performance of robots and is a comprehensive test of their joint structure design, thermal management, battery life, control algorithms, communication interference, and software - hardware coupling.
During the race, most robot participants experienced movement imbalance, falling, etc., showing poor stability. At the same time, many robots exposed heat - dissipation problems and needed continuous spraying of coolant to cool down. This is because the high - power motors and dense electronic components of robots generate a large amount of heat during continuous operation, especially under high - speed continuous movement or heavy - load working conditions, the heat accumulates quickly. Due to the human - like design of humanoid robots, their internal space is compact, and it is difficult to dissipate heat quickly.
Currently, the battery life of humanoid robots is limited. Most humanoid robots have a battery life of 2 - 5 hours, which is difficult to meet the continuous operation requirements of scenarios such as household services and industrial collaboration and still needs further improvement.
Currently, the supply - chain cost of humanoid robots is relatively high, and the hardware price still needs to be further reduced. Due to the limited mass - production scale of the humanoid robot industry, the demand for relevant components has not been fully released. Some components, such as precision sensors and planetary roller screws, are relatively expensive due to limited application scenarios. The further expansion of the industry scale in the future also depends on further cost reduction in the supply - chain link.
(This article is also contributed by Liu Yiqin, a reporter of the magazine, Huang Siyun, a researcher, and Gu Lingyu, a contributing writer.)
This article is from the WeChat official account "Caijing AI Wave", author: Zhou Yuan. Republished by 36Kr with permission.