Unitree, China's Top Robot Company, Says "Humanoid Robot Commercialization Will Take Over 10 Years"

Wang Xing, founder of Unitree, a leading company in the Chinese robot market, expressed a very conservative stance on the commercialization of humanoid robots. He asserted that it would be difficult for robots to be widely used within the next 10 years.

Wang Xing predicted that robots would reach a level where they could perform about 80% of tasks even in unfamiliar environments within 2-3 years at the earliest, or 5-10 years at the latest. However, he added that the point at which they could be used as naturally as ChatGPT in daily life is much further away.

Last week, the World Robot Conference, which drew global attention, was grandly held in Beijing. This event was an important platform to showcase the latest trends and potential for advancement in humanoid robot technology.

China is considered a global leader in the humanoid robot industry, and this conference revealed its intention to demonstrate its technological capabilities. Various robot demonstrations and technical presentations were made, raising expectations for a future robotic era.

In artificial intelligence robot technology, conversational AI and physical object control are distinct areas with entirely different levels of difficulty. Conversational robots, equipped with large language models (LLMs), can already achieve a considerable level of implementation, allowing for natural conversations.

However, controlling physical objects requires much more advanced technology. For example, the ability to feel the thickness of an eggshell and lift it with precise force, or to recognize an object's friction, weight, and tactile feel for delicate manipulation, is currently very difficult to implement.

Robots often fail to function properly when they leave a specific learning environment, which indicates that their ability to cope with complex variables in real-life settings is still insufficient.

Most robot demonstrations are not about flexibly moving by perceiving the environment in real-time, but rather are the result of thousands of training sessions in a predetermined environment. This can be described as a result of programming to move precisely according to specific coordinates.

Such robots can malfunction due to perception failure if the demonstration environment changes even slightly. For example, if the position of an object in a kitchen environment changes slightly, the robot may not recognize it and perform an incorrect action.

A realistic limitation has been raised that it will take at least 2 years, and possibly up to 10 years, for humanoid robots to reach a level where they can perform about 80% of tasks. At this level, only simple repetitive tasks would still be possible.

While simple repetitive tasks are possible, the ability to cope with unexpected situations that may arise when performing complex tasks, such as packaging items in a factory or home, is still insufficient. This is considered an inadequate level for commercial use of robots.

Some robot company founders make overly optimistic claims, such as 'commercialization next year,' which are interpreted primarily as attempts to attract investment. The analysis suggests a strong intention to garner investor interest and secure funding.

In contrast, the outlook presented by the head of a leading company like Unitree's founder, Wang Xing, is considered much more realistic. Unlike startups eager to persuade investors, they directly experience the difficulties of actual technology development and the limitations of commercialization.

The elaborate group dances, martial arts, and somersaults shown in robot demonstration videos are mostly the result of thousands of repetitive training sessions. These are pre-programmed for specific environments and movements.

These robots show limitations in responding appropriately to new songs or unexpected environmental changes. Some critical views even describe these demonstrations as mere 'fraud,' indicating a significant discrepancy between actual ability and demonstrations.

Current robot technology has made significant advancements in balance control and joint actuation capabilities. However, its ability to cope with minor unexpected situations that may occur in factories or homes is still insufficient.

For example, robots often fail to respond appropriately to unforeseen variables such as parts shifting out of place or packaging material sticking to their hands due to static electricity. A key challenge in robot development is acquiring real-world, manually generated data to train them to handle such situations.

The widespread adoption of household robots requires far more stringent technical perfection and reliability than industrial robots. Household robots must simultaneously meet several conditions, including low price, low noise, and privacy protection.

In particular, if a robot malfunctions at home, it can cause significant damage, so the demand for technical reliability is very high. This means that complex technical challenges beyond mere functional implementation must be addressed.

The data required for robot learning is fundamentally different from that for language AI. While language AI can learn from vast amounts of document data accumulated on the internet, robots face the limitation of having to generate physical action data from scratch.

Even for a minor action like peeling an onion, extensive motion data is needed: how to move which finger muscles, at what angle to apply force, and so on. Such data does not exist in traditional libraries or online, so all data must be built directly from reality.

Even for a simple action like a robot grasping a cup, camera footage alone is insufficient. Precise data for every action sequence must be recorded, including the pressure applied when holding the cup, the friction of the object, the exact angle of the robot's joints, and the degree of slippage of the object.

While video learning can help imitate movements to some extent, physical sensory data—such as touch, pressure, and slippage—is essential for safely and effectively manipulating objects in a real environment. Only by directly sensing and learning this data can robots adapt to real-world conditions.

The advanced movements shown in robot demonstrations are merely the result of meticulously planned and repeatedly trained actions. This starkly reveals the problems facing the humanoid industry. Demonstrations are nothing more than programmed actions performed according to a set schedule.

There is a significant gap between these demonstrations and actual reality. They are far from the ability to cope with numerous unexpected situations that may arise in homes or industrial sites, and some critics even describe them as close to 'fraud.'

Part of the current revenue structure in the robot industry involves artificial revenue generated by government funds circulating between manufacturers and training centers, without real external market demand. This is akin to money simply moving around internally.

If successfully established, such a structure could be considered a 'new business model.' However, if it fails to generate genuine external market demand and the fund circulation ceases, it risks collapsing into a massive 'fraudulent scheme.'

China's robot industry has experienced explosive growth in the past two years, with approximately 370 new startups established. Over 50 of these companies have either already gone public or are preparing to do so, indicating vibrant investment activity.

Currently, about 60 robot training centers, like those shown in the photo, are operating across China. These centers serve as key hubs for robot technology development and talent cultivation.

This infrastructure and capital investment demonstrate China's strong will to secure global leadership in the robot industry. However, concerns about excessive competition and the formation of a bubble are also being raised.

Following the remarks by Unitree founder Wang Xing, skepticism about the pace of humanoid robot development has emerged among experts. Some are even raising concerns about whether the fundamental pace of humanoid development should be slowed down.

This reflects the reality that current technological levels still face limitations for immediate use of robots in homes or industrial sites. There is a growing call for a realistic approach rather than excessive expectations.

The everyday assistance capabilities of humanoid robots that we've seen in videos and hoped for differ from reality. While robots can walk around a house, their ability to perform delicate tasks like picking up, holding, or moving objects is still insufficient.

Beyond simple movements, dealing with unexpected situations and performing complex daily tasks is still considered premature. Compared to expectations, the actual functionality of robots is limited, and more time is needed for them to completely replace human labor.

Within the robot industry, there is a clear difference in outlook on humanoid robot commercialization between hardware companies and the AI software camp. Robot hardware companies tend to be relatively conservative and skeptical, as they directly experience technical limitations.

Conversely, experts in the AI software camp are more optimistic, expecting that the rapid advancement of AI technology could accelerate robot commercialization. This difference in stance is attributed to varying predictions of the future based on their respective areas of expertise and perspectives.

Currently, there are almost no consumer reviews or successful corporate adoption cases in the humanoid robot market that indicate efficient use after purchasing a robot. It is difficult to find interviews with people saying, 'I bought a humanoid and I'm using it so well.'

Companies sending prototype robots to factories is essentially akin to the role of robot training centers. This suggests that commercialization verification has not yet been sufficient, and it is not yet at a stage leading to actual labor cost savings or productivity improvements.