Wang Xingxing's latest prediction: The inflection point of embodied intelligence will come within one to two years.

On October 12, at the opening ceremony of the China International Industry Fair, Wang Xingxing, founder of Unitree Robotics, put forward a more radical judgment than ever when talking about the technical inflection point of embodied intelligence: "If things go well, it may come true in even just one or two years, because when GPT-7 comes out next year, it may directly break through all the bottlenecks of embodied intelligence."

This is not the first time he has given such a timeline. In August 2025, at the World Robot Conference, he said "if everything goes smoothly, it will take one or two years, or two or three years". In August this year, he revised his statement to "two to three years at the fastest, five to ten years at the slowest". By October, the timeline has been shortened to "one or two years".

Every time a new model is released, his timeline is moved forward accordingly.

What Insights Did Unitree Gain From GPT-6

Wang Xingxing's optimism is not groundless. He specifically mentioned GPT-6 at the conference: "We have deployed it on various robots to conduct tests of different tasks, and the performance is actually very good."

To what extent is the performance "very good", reference can be found in tests of other teams. Xu Huazhe, founder of Peking Robotics, revealed at the Bund Conference that the success rate of GPT-6 Astra in robot control tasks reaches 95%, while that of its predecessor Fable 5.1 is only 40%. GPT-6 can complete operations such as grabbing, rotating and delivering. Even for objects that require precise force control like chopsticks, "it can recognize objects on the desktop and move to the correct position".

However, Xu Huazhe also pointed out that the physical generalization capability of GPT-6 is still limited. It can solve problems of "seeing what an object is, where it is, and how to approach and grab it", but cannot complete tasks involving complex physical contact such as using chopsticks to move objects, stacking boxes, and folding clothes. "Complex physical interaction is still the last fortress for embodied intelligence companies."

This judgment exactly explains why Wang Xingxing bet on GPT-7. General large models have proven their capabilities in semantic understanding and spatial generalization. The next hurdle to overcome is physical interaction. Wang Xingxing's judgment is that GPT-7 is likely to "directly break through all relevant problems" — not an incremental improvement, but a leapfrog breakthrough.

The Real Bottleneck in Factories Is Not That AI Is Not Smart Enough

Expectations for technological breakthroughs are accelerating, but the actual progress of robots entering factories is not as fast as imagined.

Wang Xingxing himself admitted: "At present, our company and other robotics companies are all carrying out pilot deployments in factories. It is true that for large-scale application, further improvement is still needed in both generalization and efficiency."

Unitree has carried out pilot deployments with several automotive enterprises as early as 2024. But two years later, it is still in the "pilot" stage. IDC data shows that in the first half of 2026, the global shipment of humanoid robots is close to 25,000 units, of which scenarios such as scientific research and education, performance demonstration and government data collection centers still account for 69% of the global shipment, while scenarios including industrial manufacturing, retail and logistics have not yet become the main shipment directions.

What exactly is the bottleneck? Wang Xingxing once said a very straightforward sentence at the World Robot Conference: "Every time a new task comes, retraining is required, and the overall efficiency is lower than that of human beings." After a robot learns to tighten screws at one station, it may need to be retrained when moving to another station. Insufficient generalization means that every new scenario brings cost that starts from scratch.

The efficiency problem is more practical. Industrial scenarios have rigid requirements for takt time, and there are clear standards for the hourly output of workers on a production line. If the robot is not fast enough, or has a relatively high error rate, the entire production line has to slow down for it. According to a survey by Xinhua Finance, some enterprises reported that robots have a "relatively high error rate", and insufficient perception accuracy and disconnection between decision-making systems and industrial scenarios are common problems.

Wang Xingxing said these problems "are very likely to be solved in the next few years", but he did not give details of the solution path. Judging from Unitree's own actions, the idea is to apply AI to the model development chain — to form a "physical AI self-evolution" system, so that code writing, model training, deployment, testing and feedback form a self-loop. But he also emphasized that "human participation in evaluation is still required".

Who Is Making Money and Who Is Betting on the Future

Wang Xingxing's optimistic statement has another meaning when viewed against the current capital environment of the humanoid robot industry.

According to data from Frost & Sullivan, the top five companies in the global humanoid robot revenue ranking in the first half of 2026 are all Chinese companies: Ubtech ranks first with 620 million yuan, Agibot ranks second with 610 million yuan, Unitree Robotics ranks third with 580 million yuan, Galaxy Universal ranks fourth with 210 million yuan, and Leju Robotics ranks fifth with 150 million yuan.

However, Ubtech, which ranks first, posted a loss of 339 million yuan in the first half of the year. Full-size embodied intelligence humanoid robots — the type that can actually "do work" — only sold 921 units in the first half of the year. In the past few years, Ubtech has continued to lose money, and it still failed to make a profit in the first half of this year.

The situation of Unitree is slightly better. In 2025, its shipment exceeded 5,500 units, with a revenue of about 1.7 billion yuan and a non-net profit margin of about 34.8%. But its main products are half-size models such as the G1 that cost tens of thousands of yuan, and its customers are mainly from the scientific research and education sector. The shipment of full-size models for industrial scenarios is far less than that of half-size products.

This is not just a problem for Unitree. The structure of the entire industry is: low-cost models contribute most of the sales volume, while high-priced robots support most of the revenue, but no company has yet achieved positive profit from full-size humanoid robots in industrial scenarios. Of the 25,000 units shipped in the first half of the year, most are models with a unit price of less than 30,000 US dollars. Full-size models only account for about 20% of sales volume, but take more than 75% of the industry's revenue.

Capital is pouring in. As of October 11, there have been 565 primary market investment and financing events in the embodied intelligence sector this year, with a total amount of 1.12622 trillion yuan. The figure for the whole year of 2025 was 395.11 billion yuan, and in less than ten months this year, it is nearly three times that of last year. However, the direction of capital investment is changing. In the first half of the year, the proportion of capital absorbed by complete machine enterprises declined, while the investment and financing activity in core components and data infrastructure rose. The investment logic is shifting from "paying for dreams" to "paying for scenarios".

If the Inflection Point Really Comes, Factories Will Not Be the First to Benefit

Wang Xingxing also mentioned an easily overlooked judgment at the Industry Fair: the landing of commercial services may be faster than that of the industrial sector.

This judgment is noteworthy. If the inflection point of embodied intelligence arrives within one to two years, the first scenarios to achieve large-scale deployment may not be factories, but commercial services — scenarios such as guided tours, inspection, and retail assistance that have less strict requirements for efficiency and larger fault tolerance space. Unitree's current main customers in scientific research and education essentially fall into this category.

In contrast, Wang Xingxing ranks the landing of home scenarios last. The home environment puts the highest requirements on the adaptability of robots. The placement of tables and chairs, the position of objects, and the ground conditions are changing every day, and the threshold of generalization capability is far higher than that of fixed stations in factories. Wang Xingxing said that the industrial sector and To B commercial services "will be faster and more mature than home scenarios", and this ranking is pragmatic.

For investors, this means that the commercialization timeline of humanoid robots needs to be understood in layers. The large-scale deployment of factory scenarios depends on whether generalization and efficiency can be substantially improved in a few years; the volume growth of commercial service scenarios may come earlier; home scenarios are a longer-term matter.

The mass production pace of Tesla Optimus also points to a similar structure. The production line at the Fremont Factory that originally produced Model S and Model X has been transformed into a robot manufacturing facility, with a target annual production capacity of 1 million units. As of March 2026, more than 1,000 Optimus units have been sorting batteries and transporting parts in Tesla's own factories. But external sales will not start until after 2027, and the long-term target price is as low as 20,000 to 30,000 US dollars, on the premise that the annual output can reach 1 million units.

Tesla's strategy is to use the robots itself first, iterate with real production line data, and then consider selling to external customers. This flywheel — the spatial understanding cultivated by millions of miles of FSD training, the amplified training of Cortex 2.0 supercomputing, and the real operation data fed back by robots in factories — cannot be replicated by pure robotics startups.

Belief Cannot Replace Verification

The most widely spread sentence in Wang Xingxing's speech this time is his reflection on his past misjudgments: "The biggest mistake I have made in the past ten years is that I did not believe enough in AI."

Coming from the mouth of a robotics company founder, this sentence carries a lot of weight. It means that he believes that his previous estimate of AI's capabilities was too conservative, and he is now correcting this deviation. The way to make the correction is to keep moving forward the timeline of the technological inflection point.

But belief cannot replace verification. The fact that GPT-6 enables robots to pick up chopsticks does not mean that robots can work stably in factories; if GPT-7 can really "break through" the problems of embodied intelligence, there is still a long chain of data collection, hardware adaptation, scenario debugging and cost control between model breakthrough and engineering implementation.

Wang Xingxing said that "within one or two years" is the technological inflection point, not the commercialization node. This distinction is very important. Technology is accelerating, but the robots in factories still need to answer a more plain question: how many screws can it tighten in a day, and when can its cost be lower than that of a human worker. Before this question is answered, all predictions of inflection points are just predictions.

This article is from the WeChat official account "Tech Business", written by Tech Business, and authorized for release by 36Kr.