After "cheaper than a smartphone", embodied intelligence enters its second half

Recently, JD.com launched a "Hot Sale List of Humanoid Robots".

The top-ranked robot on the list is a humanoid robot called "Xiaobumi". After applying government subsidies and platform promotions, the final price is 9,118.1 yuan - below 10,000 yuan, even cheaper than a high-end mobile phone. The platform shows that more than 400 units have been sold so far.

Humanoid robots have truly entered the consumer price range for the first time.

Looking back further, the trend becomes clearer. The unit price of Unitree's humanoid robots was 593,400 yuan in 2023, dropped to 260,000 yuan in 2024, and fell to 167,600 yuan in 2025. By April 2026, the starting price of the R1 dual-arm version was 26,900 yuan. In just three years, the price has dropped by nearly 95%.

This is not just a product price cut; it is also a signal that an industry has undergone a certain transformation.

01

First Half: Make It Move

During the Spring Festival in 2025, Unitree's robots performed a yangge dance on the CCTV Spring Festival Gala stage, completing a series of highly difficult coordinated movements. That moment was remembered by many - smooth, stable, and without falling. This would have been almost an impossible task five years ago. The next year, the number of robot manufacturers on the Spring Festival Gala stage expanded to four, and they could do everything from dancing and performing skits to shooting short films.

Being able to move is the core proposition solved in the first half of the embodied intelligence era.

What made "being able to move" possible was a deep restructuring at the supply chain end. The core components of humanoid robots - reducers, servo systems, and controllers - almost entirely relied on imports in the past, with high prices and long lead times. By 2026, the localization rate of these three types of components had risen to 75% - 90%. More importantly, the joint motors and sensing components of robots began to reuse the mature supply chains of new energy vehicles and consumer electronics on a large scale. With a large enough demand, the cost of the same parts dropped.

The material cost of the basic version of Unitree G1 is 41,600 yuan, with the joint cost accounting for the majority, about 27,500 yuan. Even if the overall machine selling price is pressed down to 85,000 yuan, the gross profit margin still remains at around 40%. In other words, this price is already based on normal business logic, not a loss - making operation.

The increasing shipment volume amplifies this logic. In 2025, the annual shipment volume of Chinese humanoid robots reached 14,400 units. The High - tech Robotics Industry Research Institute predicts that the annual shipment volume is expected to exceed 62,500 units in 2026. As the scale increases, the R & D cost and production cost per robot are further reduced, and there is still room for price reduction.

When many people see that robots can move and the price is right, their first reaction is: Will robots soon enter every household?

The answer is not necessarily. This is the challenge in the second half of the embodied intelligence era that started in 2026: being usable.

02

Second Half: The Dilemma of "Usability" in the Industrial Sector

A detail can illustrate the problem well.

The "dexterous hand", the core component for robots to interact with the physical world, is scientifically called the "end - effector". To some extent, it ultimately determines what a robot can do and how precisely it can do it. However, the prices of dexterous hands on the market currently range from a few thousand yuan to 200,000 yuan, and no product has truly dominated the market. What's more troublesome is that in real - world usage environments, the lifespan of many dexterous hands is only one to three months - the shortest ones start to malfunction within a week.

This is just a problem at the hardware level. A deeper challenge lies in the matter of "judgment". Traditional industrial robots face a single assembly line and repetitive tasks such as welding, screwing, and spraying. Their working logic is "hit where it's pointed" - with clear instructions and fixed actions, and they stop and wait for human intervention when there is a change. What an embodied intelligence robot needs to do is to autonomously perceive, judge, adjust, and solve temporary problems in a complex real - world environment, such as factory patrols, circuit maintenance, and cargo handling. These are two fundamentally different things.

The current state of the industry is that simulation training and multi - modal perception have been implemented on a large scale in leading enterprises, but the autonomous decision - making ability of robots is still in the production line verification stage. More directly, the industry has not yet witnessed the "ChatGPT moment" of embodied intelligence - that is, a breakthrough in autonomous decision - making ability in an open environment.

This gap will ultimately be reflected in the balance sheets. When manufacturing enterprises purchase robots, they mainly look at two indicators: an investment payback period of less than two years and an hourly comprehensive cost of less than $15. According to this standard, even the most advanced enterprises can only barely meet the threshold under ideal working conditions. However, a turning point has emerged - small and medium - sized factories in the Pearl River Delta and the Yangtze River Delta have started to introduce medium and small - sized industrial models such as Unitree R1 dual - arm version (starting from 26,900 yuan) and Stardust Intelligence T1 (starting from 89,900 yuan). The payback period for a single unit has been shortened to 14.7 months, and the proportion of manufacturing - scenario orders has risen to about 37.5%. The accounts are starting to make sense, but only in some scenarios.

Currently, in the domestic market, there are roughly three approaches to overcome the "usability" hurdle.

Unitree represents the hardware school. Starting from robotic dogs, it has extremely strong engineering capabilities, and cost control is its core competitiveness - it was able to cut the price from 590,000 yuan to 26,900 yuan precisely because of this ability. However, there are limits to the "hardware - centric" approach: while the price can continue to be lowered, the ability of a robot to make autonomous judgments in complex scenarios cannot be solved by price cuts. Unitree's IPO registration on the Science and Technology Innovation Board was approved on July 2nd, and 85% of the raised funds will be invested in R & D, with nearly half going into the large - scale model of embodied intelligence - it clearly knows that the battle in the second half is not about hardware.

ZhiYuan takes a different approach, focusing on the "brain" of the robot. It has launched the "Lingchuang" platform, which allows users to choreograph robot actions and define task processes without writing code, significantly lowering the development threshold. The idea is correct, but the platform is just starting, and the developer ecosystem is still in the process of formation. There is currently no public data to determine whether it can achieve scale.

Leju's logic is more down - to - earth: first, establish mass - production capabilities. It has jointly built the first automated production line for humanoid robots with an annual output of over 10,000 units in China with Dongfang Seiko. One robot can be produced every 30 minutes, and the annual production capacity exceeds 10,000 units. More importantly, it has released Taskor, the first industrial embodied intelligence software platform in China - production line engineers can choreograph robot actions by dragging graphic modules without writing code, deploy them to an entire robot team with one click, and directly connect to the factory's MES system. Leju is betting on first expanding the volume, occupying actual usage scenarios, and then gradually iterating in real - world scenarios.

These three approaches are actually competing for the same thing: whoever can first run through the scenarios can build a data moat.

Consumer Sector: A More Difficult Testing Ground Than the Industrial Sector

In addition to the industrial sector, the home is also a scenario that many embodied intelligence enterprises are betting on.

At the end of June, Ubtech released the consumer - grade brand "YouWorld" U1 series, full - size hyper - realistic humanoid robots, mainly for emotional companionship, with a price range of 119,800 yuan to 990,000 yuan. The pre - orders across all channels have exceeded 13,000 units. Meanwhile, 1X Technologies in Silicon Valley launched the humanoid robot NEO, priced at $20,000, which can do laundry, water flowers, and sweep the floor, and is planned to be delivered in 2026. There is also a company called Weave Robotics with a completely different approach - its Isaac 1 doesn't take on a human form. It has a wheeled base and a two - finger gripper and only does two things: collect sundries and fold clothes, priced at $7,999.

"Can it handle different situations every second?" - The same question becomes much more difficult when the testing ground changes from the factory to the living room. Clothes have different materials, toys are randomly scattered, and each grab is different from the previous one. Isaac 1's solution is to compromise: when the robot gets stuck, a remote human expert can access through VR to make a 5 - to 10 - second fine - adjustment and then hand back the control. The AI vice - president of 1X admitted that the product "may make mistakes". Ubtech U1 is even more straightforward - it simply avoids housework and only focuses on companionship. These three approaches all confirm the same fact: household robots still can't pass the seemingly simplest test of "folding clothes".

However, Silicon Valley is still enthusiastically betting on it. The reason is simple: if a robot really learns to handle soft, constantly deforming objects without a standard answer, the capabilities it acquires won't be limited to folding clothes. Fabrics, wiring harnesses, and packaging bags in factories, soft packages in warehouses, and gauze and consumables in hospitals all rely on the same ability - understanding complex environments, stable grasping, and precise operation.

Folding a piece of clothing is a test of a robot's ability to face the real world.

Conclusion

Buying a humanoid robot for less than 10,000 yuan was unimaginable three years ago. However, if we take "less than 10,000 yuan" as the end goal, we misinterpret the meaning of this signal.

The price drop means that the threshold is lowered, allowing more people to enter the market, make mistakes, and accumulate data. This is exactly what is needed in the second half - friction, failure, and iteration in the real world, with no shortcuts.

In the 2026 government work report, "embodied intelligence" was clearly written as a future industry, and the term "embodied intelligence" was even selected as one of the annual buzzwords in "Yao Wen Jiao Zi" last year - the fact that an industry term has entered the public language shows that external expectations have outpaced technological realization.

In the first half, the industry solved the problem of "whether it can be manufactured". This hurdle has been overcome. However, there is still a considerable distance to reach the level of "being easy to use".

In the second half, the problem to be solved is "whether it can really be used after being manufactured". To judge the development stage of the robot industry, we should look at the scenarios rather than the price - the price only tells you "whether you can buy", while the scenarios tell you "whether you can use". The reliability of dexterous hands, the boundaries of autonomous decision - making, and whether the data flywheel can really start spinning - these issues cannot be solved by subsidies or press conferences.

Reaching a mass - production volume of over 10,000 units and a price of less than 10,000 yuan is the answer sheet for the first half. The answer sheet for the second half will be written in factories, warehouses, and living rooms.

This article is from the WeChat official account "Manfon Slow - Motion", written by Manfon, and is published by 36Kr with permission.