Humanoid robots are waiting for their ChatGPT moment.

The wave of AI continues to gain momentum.

The robotics field is a typical example. Recently, a press conference was held for the 2025 China Robot Industry Development Conference, hosted by units such as the China Federation of Machinery Industry. Data from the press conference shows that the domestic robot industry has witnessed rapid growth in scale. The revenue has increased from 106.1 billion yuan in 2020 to 237.89 billion yuan in 2024. In the first three quarters of 2025, the domestic robot industry's revenue increased by 29.5% year - on - year. The production of industrial robots reached 595,000 units, and the production of service robots reached 13.5 million sets, both exceeding the annual production in 2024.

As a key carrier and core implementation direction of AI, robots are injecting new impetus into industrial transformation. A more popular term for this trend is "embodied intelligence", which refers to intelligent agents with physical bodies that can perform tasks in the real world through perception, decision - making, and interaction capabilities and continuously evolve through interaction with the environment. Embodied intelligence is driving technology from algorithmic models to the real world, expanding the boundaries of AI applications and exploring more possibilities for achieving general AI.

By definition, embodied intelligence includes not only humanoid or other forms of robots but also drones and intelligent vehicles equipped with AI models. Among them, the humanoid robot track is particularly eye - catching. From overseas companies like Figure AI and Tesla to domestic ones like Unitree Technology and Zhipu Robotics, global innovation forces are driving the industry forward at a rapid pace and constantly setting new records.

On October 29, 2025, Norwegian technology company 1X launched the household humanoid robot NEO, opening pre - orders at a price of about $20,000 (approximately 142,000 yuan) or a monthly subscription fee of $499 (approximately 3,500 yuan), with delivery scheduled for 2026. Unitree Technology brought its humanoid robots into the "Double 11" promotion, selling them on the JD platform at a price of 29,900 yuan.

Since ChatGPT ignited the global AI boom in November 2022, AI has quickly entered the public eye, transforming from an out - of - reach high - end technology into a tool accessible to everyone. The "Report on the Application and Development of Generative Artificial Intelligence (2025)" released by CNNIC (China Internet Network Information Center) shows that as of June 2025, the number of domestic generative AI users has reached 515 million, with a penetration rate of 36.5%.

The development of generative artificial intelligence, also known as generative AI or AIGC, has promoted the development of related fields, especially the embodied intelligence industry, bringing the scenarios of humans and robots coexisting in science - fiction works such as "I, Robot" and "WALL - E" closer to reality. As a result, tech giants are making heavy investments, and startups are rushing to enter the market. In this competition for the future of technology, players are striving to build moats, attempting to create the "ChatGPT moment" for humanoid robots first.

01

Solving the Movement Problem

Nowadays, the iteration of humanoid robots is advancing by leaps and bounds.

Robot performances are quite prominent. At the CCTV Spring Festival Gala in January 2025, the performance of "Yang Bot" by Unitree Technology's robot H1 only involved simple movements like swaying the body and rotating a handkerchief according to the rhythm. By October 2025, in the curtain - call segment of the dance drama "The Exploitation of the Works of Nature", Unitree Technology's robots could accurately reproduce the dancers' postures and complete "human - robot co - dancing" with smooth flips and backflips.

The relevant performance videos of Unitree Technology's robots have spread rapidly on platforms like Douyin and Kuaishou, and the total number of likes has exceeded 1.3 million. Some users commented in the comment section, saying that the robots' movements seemed uncoordinated at the beginning of the year, but within just a few months, they had become so coordinated, as if they had "received a martial - arts secret manual".

The breakthrough progress of robots is the result of decades of continuous accumulation in technological development.

It should be noted that AI pioneer Alan Turing proposed in a 1950 paper that intelligence must rely on physical entities to interact dynamically with the outside world to form. However, limited by the technological level, for more than half a century, robots have been far from true embodied intelligence.

In the Fukushima nuclear power plant accident in 2011, there were no mature robots with practical operation capabilities at the rescue site. The limited equipment was frequently trapped in the complex radiation environment, and some were even tripped by scattered cables, making it difficult to perform key tasks. After that, DARPA (Defense Advanced Research Projects Agency) announced the launch of a robot challenge to promote the R & D of disaster - relief robot technology.

The first DARPA Robot Challenge started in October 2012 and the winner was not determined until June 2015. In the final, robots were required to perform tasks such as reaching the mission area, getting out of the vehicle autonomously, opening doors, closing gates, and making holes with tools. Most of the participating robots walked clumsily, fell frequently, and most were unable to complete all the operations. The champion was the robot HuBo developed by KAIST in South Korea, which used omni - directional wheels instead of its feet to ensure speed and balance during movement. The runner - up was the robot Atlas developed by Boston Dynamics.

At that time, the final video sparked public discussion - the robots were slow and made many mistakes, which was completely different from the public's expectation of a sensitive and intelligent assistant.

As a global leading humanoid robot company, Boston Dynamics, founded in 1992, was once a pioneer in the industry. As early as 2017, Boston Dynamics' Atlas demonstrated a backflip. However, Atlas originally used a hydraulic drive system, which had problems of high energy consumption, high noise, and high cost while having high strength and precision, making it difficult to commercialize. Boston Dynamics was acquired by Google in 2013, transferred to SoftBank in 2017, and then acquired by South Korean Hyundai in 2021. During the SoftBank period, Boston Dynamics launched the robot dog Spot into the market at a price of about $75,000 (approximately 530,000 yuan), but only about 400 units were sold.

The reason why the flip action is regarded as a key milestone in robot technology development is that it systematically integrates and promotes the progress of multiple core fields such as robot hardware design, dynamic control, and real - time decision - making.

According to Haike Finance, to complete a flip, the robot's drive system needs to release sufficient power density instantaneously, and the high - load duration is extremely short; the system needs to solve the six - degree - of - freedom motion equation including translation in the front - back, left - right, and up - down directions and rotation around three axes in real - time, and an angle deviation of more than 0.5 degrees may lead to imbalance when landing; the robot's ankles, knees, and hips need shock absorption and buffering, requiring the foot force sensor to sense the ground reaction force and respond within 0.01 seconds.

New players represented by Unitree Technology have abandoned the hydraulic drive solution and chosen the pure - electric drive technology route, overcoming the problem that the power of the electric drive was inferior to that of the hydraulic drive in the past. They ensure the balance between action completion and cost through self - developed high - torque motors and lightweight structural design. For example, Unitree G1 uses 23 - 43 joint motors, with a maximum joint torque of 120N·m, enabling it to maintain overall stability even in actions such as side flips that require extremely high lateral inertia control.

Boston Dynamics also launched an electric - drive version of Atlas in April 2024, marking wider recognition of the electric - drive technology path. In February 2025, domestic company Zhongqing Robotics successfully completed the world's first front flip by a robot, achieving an important technological breakthrough. Compared with the common backflips in the previous stage, the front flip places higher requirements on the robot's dynamic balance ability, instantaneous explosive power, and precise landing control.

02

Where Does Intelligence Come From?

The breakthroughs in high - difficulty actions such as flips are far more than just technological demonstrations.

These actions can systematically verify the maturity of the overall machine control system and key components, laying the foundation for the application of robots in complex real - world environments. In a public demonstration in September 2025, Unitree Robot G1 could quickly react and return to a standing state when faced with continuous pushing, shoving, and kicking, demonstrating considerable motor intelligence.

This marks an acceleration of the process of robots moving from the laboratory to the complex real world.

Since the breakthrough of AIGC in 2022 and Tesla's launch of the prototype of the robot Optimus, the global humanoid robot industry has entered a period of rapid development. A research report released by Guotai Junan Securities in November 2025, integrating data from multiple sources, shows that in 2024, the number of domestic humanoid robot enterprises registered was 104, a year - on - year increase of 104%. The humanoid robot is also a hot spot for investment and financing. From January to July 2025, there were 101 financing events in the domestic humanoid robot industry, with a financing amount exceeding 26 billion yuan, exceeding the total financing amount in 2024. Before 2024, the humanoid robot industry was in the experimental testing stage, and the products were prototypes, mostly with a scale of less than 10 units. From 2024 to 2025, the industry entered the trial - production stage, and some leading enterprises started pilot deliveries of dozens to hundreds of units. After 2025, the industry will enter the stage of large - scale mass production.

It should be noted that players in the robot industry can be divided into two development paths: hardware - oriented and software - oriented according to their business focus. Hardware - oriented enterprises take the robot body as the core entry point, focus on the independent R & D of key components such as joint modules, motors, reducers, and controllers, and focus on breaking through motion control algorithms. This is similar to the "cerebellum" of humans, and their products are usually measured by load - carrying capacity, speed, and motion performance, such as Boston Dynamics and Unitree Technology.

Software - oriented enterprises start more from embodied intelligence technology, taking cutting - edge visual - language large models, world models, and simulated synthetic data as the starting point for R & D. They usually integrate the robot body by purchasing components externally and emphasize the robot's cognitive and decision - making intelligence, such as Galaxy Universal. Automobile companies like Tesla, which have a foundation in large - scale manufacturing, can demonstrate full - stack capabilities of both hardware and software in the robot field with their deep hardware manufacturing genes and software accumulation from intelligent driving.

In the stage when artificial intelligence technology was not yet mature, robots completely relied on precise trajectory codes written by engineers to perform tasks, which was essentially no different from traditional production equipment. Just like the backflip demonstrated by Boston Dynamics' Atlas in 2017, it was essentially the precise execution of a preset program.

Later, robot learning entered the data - driven stage, where robots learned skills autonomously through observation, imitation, and repeated trial - and - error. Further on, the intelligent system was deeply integrated with autonomous learning, and robots began to understand abstract instructions, actively try solutions in unfamiliar environments, and gradually evolve into autonomous intelligent agents capable of dealing with complex realities. Thus, global players are showing their strengths in algorithms.

After announcing the termination of its cooperation with OpenAI in February 2025, overseas leading player Figure AI turned to self - developed end - to - end AI models. It is reported that its large AI model Helix has achieved a major technological breakthrough. Helix introduced a dual - system thinking into the VLA model (Visual - Language - Action model) for the first time. System 1 focuses on real - time action control and can process visual information at a very high response speed; System 2 has strong scene understanding and language parsing capabilities and is responsible for interpreting complex instructions, identifying environmental elements, and formulating action plans. The dual - system architecture also has advantages in modular iteration ability; the two systems can be optimized independently without readjusting the overall model.

Domestic player Zhipu Robotics announced in September 2025 that it would fully open - source its general embodied base large model GO - 1, which uses the innovative ViLLA architecture and is the world's first open - source general embodied intelligence model using this architecture. The full name of the ViLLA architecture is Vision - Language - Latent - Action, which can effectively bridge the semantic gap between image and text input and the final action execution of the robot by introducing implicit action markers, enabling the robot to understand human instructions more accurately and translate them into precise actions.

In addition, new players such as Physical Intelligence and Skild AI in the United States are exploring the cutting - edge field of world models, aiming to enable robots to build an internal physical world model so that they can predict the results of their actions.

03

Many Challenges Remain

The robot industry has begun to build a systematic technological development framework.

Analogous to the L1 - L5 autonomous driving grading system, Zhipu Robotics proposed a G1 - G5 embodied intelligence technology roadmap. According to Haike Finance, G1 - level is customized for specific scenarios and lacks the ability to migrate across scenarios; G2 - level can understand multi - scenario tasks and achieve limited generalization by combining large language models; G3 - level turns to end - to - end data - driven, achieving a paradigm shift at the architectural level; G4 - level introduces general operation large models and simulation data, significantly improving performance in complex tasks; and G5 - level, as a long - term goal, will achieve complete end - to - end autonomous operation from perception to execution.

Generalization is an extremely important challenge in the process of robot technology moving towards a higher level.

Generalization means that robots can perform various tasks flexibly in different scenarios without being retrained or adjusted for each new scenario. At present, the skills that robots master proficiently in specific environments are difficult to transfer effectively to new scenarios, tasks, or objects. It may be able to perform a grasping action accurately in a laboratory environment, but as long as the shape of the cup is changed or the lighting conditions are altered, its performance will decline significantly or even fail completely.

The root cause of this challenge lies in the infinite complexity of the real world. The real environment is open and dynamically changing