Is the Leader of Yushu Pouring Cold Water? How Long Until the 'ChatGPT Moment' of Embodied Intelligence?

At a time when humanoid robots are gaining significant attention, the leader of the "first humanoid robot stock" in A-shares seems to have poured cold water on the industry.

On August 20, during the 2026 World Robot Conference, Wang Xingxing, founder of Yushu Technology, candidly stated in his first public speech after the company's listing that the "ChatGPT moment" for embodied intelligence has not yet arrived, predicting it could be two to three years at best, or as long as five to ten years at worst.

In the past two years, humanoid robots have become the hottest "fashion item" in technology. They are increasingly visible on stages of major events, at gaming exhibitions, and even at the entrances of shopping malls and retail stores. These futuristic robots can be seen dancing, walking, and waving, interacting with audiences, and are rapidly moving from laboratories into public view.

However, behind the lively displays, the actual capabilities of these robots are clearly far from the public's imagination of "future robots." At least in more general expectations, robots should not only perform on stage or undertake marketing tasks but should genuinely enter factories and homes to transport, organize, operate tools, or take on some household and caregiving tasks. The public's question has always been specific and direct: When will robots truly be able to work?

In the field of large language models, ChatGPT once provided a highly symbolic answer. It was not the first language model, but it was the first to make a large number of ordinary users intuitively aware that AI had crossed a certain capability threshold. When a similar moment will occur in the field of embodied intelligence has gradually become one of the most closely watched questions in the robotics industry.

The Key to Robots Entering Homes Lies in Generalization Ability

At the World Robot Conference on August 20, Wang Xingxing pointed out that the biggest bottleneck for embodied intelligence currently is a problem that has been repeatedly discussed in the industry—generalization ability.

In his view, many robot models can achieve nearly 100% success rates in fixed scenarios as long as sufficient data collection and training are completed. However, if the objects being manipulated change or the environment shifts slightly, the success rate of the robots may drop significantly.

This is also the most important threshold for robots to truly enter daily life and homes.

Wang Xingxing provided a rather specific criterion: If one day a robot can be taken to a completely unfamiliar home or environment and can complete about 80% of tasks solely through voice or language commands, then embodied intelligence will have approximately reached its "ChatGPT moment." He believes that at best, this could happen in two to three years, while at worst, it could take five or even ten years.

Wang emphasized that the real difficulty is not getting the robot to "generally know what to do," but rather the last few centimeters or even millimeters.

For example, if a robot is to pick up an object, the model may have correctly planned the entire action, and the robotic arm may have moved close to the object, but if the final positional error, tactile feedback, or grasping force is not correctly adjusted, the entire task could fail.

This is also where robots differ significantly from large language models.

The input and output of language models always occur in digital space, while every perception and action of a robot must interact with the real physical world. Sensors have noise, actuators have errors, and the position, material, and weight of objects are constantly changing; these errors can accumulate as tasks are executed. Therefore, for embodied intelligence, moving from "basic capability" to "stable correctness" may be more challenging than initially learning a task.

Almost simultaneously with Wang Xingxing's judgment, Wang He, founder and CTO of Galaxy General, provided a more specific timeline. Wang He stated that with continuous data accumulation and further technological breakthroughs, embodied intelligence is expected to reach its "ChatGPT moment" in 2028. He defined this milestone as the point at which robots can complete about 70% to 80% of daily tasks without specialized training for specific tasks.

The judgments of the two are actually very close. Wang Xingxing focuses on the task completion rate of robots in unfamiliar environments, while Wang He is concerned with the direct generalization ability of the base model without specialized training, but both set the critical point at around 70% to 80% success rate for unknown tasks.

This also means that the "ChatGPT moment" they refer to is not the movement capabilities demonstrated by humanoid robots today, nor is it a fixed demo that can achieve over 99% success rate after extensive training. The real change should occur when robots begin to break free from dependence on fixed scenarios, fixed objects, and specialized training.

When Will Robots Truly "Get It"? A Collective Focus on 2 to 5 Years

Just a month ago, at the 2026 World Artificial Intelligence Conference (WAIC), a roundtable posed the same question to six leading entrepreneurs and researchers in the field of embodied intelligence.

On July 19, during the "Intelligent Initiation of Embodied Forum" hosted by Zhiyuan Robotics and Mifeng Technology, the moderator asked six guests to answer a very direct question: How many years until robots reach their ChatGPT moment?

The results were surprisingly concentrated. Yao Maoqing, partner at Zhiyuan Robotics and chairman and CEO of Mifeng Technology, answered two years; Tony Zhao, co-founder and CEO of Sunday Robotics, believed it would be within three years; Zhang Zhengyou, chief scientist at Tencent and director of Robotics X Lab, gave a timeframe of three to five years; Ma Yecheng, co-founder and chief scientist of Dyna Robotics, estimated about four years; Ren Zhiyi, research scientist at Physical Intelligence, predicted four to five years; and Xu Danfei, professor at Georgia Tech, provided the most conservative answer of five years.

The six individuals come from completely different companies and research institutions, with varying technical routes, but ultimately their answers all converged within the next two to five years.

Among them, Yao Maoqing was the most optimistic. His judgment is primarily based on the growth of data scale. In Yao's view, the current constraints on the further development of physical AI can be summarized as three walls: "data, representation, and closed loop." Compared to the vast amounts of text and images that can be cheaply obtained in the internet world, the real interaction data for robots is not only expensive but also limited by the differences in robot bodies, tasks, and scenarios.

Robots that can be used out of the box need to understand open natural language commands and achieve about 70% to 80% basic success rates on common tasks. To achieve this, embodied intelligence needs data scales far exceeding current levels. He even suggested that in the future, data scale might need to reach hundreds of millions of hours.

In other words, in Yao Maoqing's judgment framework, the "ChatGPT moment" is largely a scaling problem: as real-world data, simulation data, internet videos, first-person perspective data, and data flowing back from robot deployments gradually form a scale, the model's capabilities may also cross the critical point.

Tony Zhao of Sunday Robotics, on the other hand, compressed the timeline to within three years. Rather than pursuing increasingly complex robot bodies, Zhao is more focused on when the robot's "brain" can truly mature. Sunday Robotics has long focused on household robots and foundational models, believing that if a sufficiently strong foundational model can understand complex environments and tasks, even if the robot uses relatively simple and cost-effective grippers, it may still solve many practical problems first.

Zhang Zhengyou's attitude is much more cautious. He provided a timeframe of three to five years but emphasized that this window does not mean the industry only needs to wait for a larger model to suddenly appear. One important premise for large language models to quickly form scalable capabilities is that the Transformer has gradually become a relatively unified architecture. However, robotic intelligence has not yet formed such a clear technical paradigm.

From high-level cognition, visual perception, and task planning to real-time motion control, bodily feedback, and safety responses, a robot needs to handle problems across different time scales simultaneously. Therefore, real breakthroughs may require data collection, simulation, real deployment, failure recovery, hardware, and models to progress together, rather than simply replicating the scaling laws of large language models. For this reason, Zhang Zhengyou also emphasized that the industry still needs to "work steadily."

Ma Yecheng of Dyna Robotics estimated about four years. His judgment is more derived from commercial deployment. For a laboratory demo, an 80% or 90% success rate may be sufficient to prove the effectiveness of the technical route; however, for a factory or service enterprise that actually purchases robots to work, such reliability is often far from enough. Therefore, the "ChatGPT moment" for embodied intelligence not only means that robots can understand more tasks but also that they must further improve to a level acceptable in real commercial environments.

Ren Zhiyi of Physical Intelligence predicted four to five years. He mentioned that before joining Physical Intelligence, he once thought this process might take ten years, but with the development of foundational models for robots over the past year, his expectations have significantly advanced.

Physical Intelligence itself is also one of the current representative companies in the "robot foundational model" route, aiming to gradually acquire broader operational capabilities through cross-body and cross-task data training. In this logic, what is truly important is whether the robot's capabilities can continuously emerge as the model and data scale expands, ultimately producing effects similar to zero-shot transfer in large language models.

Xu Danfei, a professor at Georgia Tech, provided the most conservative answer of five years. Rather than betting on a specific model route, he focuses on three more fundamental indicators: whether the model can truly learn from data, whether the learned capabilities can generalize to new environments and tasks, and whether the reasoning speed can meet the real-time operational requirements of robots in the real world.

When we put together Wang Xingxing's "fastest in two to three years," Wang He's "2028," and the answers from the six guests at WAIC, an increasingly evident phenomenon is that the embodied intelligence industry is compressing the previously distant timeline for "general robots" into the next technological cycle.

A Clear Historical Node or a Gradual Process?

However, although these entrepreneurs and researchers are increasingly converging on time predictions, they are not actually discussing the "ChatGPT moment" using the same set of standards.

Wang He focuses on the capability leap of the foundational model itself. In his definition, when robots no longer need to undergo specialized training for each new task and can complete 70% to 80% of daily tasks solely based on the foundational model, embodied intelligence will have crossed a critical point similar to ChatGPT. This standard primarily measures whether the model can transition from "specialized intelligence" for specific tasks to general intelligence with strong transfer capabilities.

Wang Xingxing's standard is closer to the actual performance of robots in the real world. He also views an approximately 80% task completion rate as an important threshold but emphasizes whether robots can complete most tasks relying solely on language commands after entering unfamiliar environments. In his view, many robots today can achieve high success rates in well-trained fixed environments, but the real challenge is whether the model can continue to work when the environment, objects, or tasks change.

These two definitions may seem similar, but the levels of focus are not entirely the same. A model may possess good zero-shot or cross-task capabilities, but that does not mean a robot equipped with it has become a reliable production tool. Wang Xingxing mentioned at the WRC that Yushu has already deployed robots in automotive factories and its own factory, which can complete some simple assembly tasks, but there is still no large-scale promotion, one important reason being that the efficiency of robots is still lower than that of humans, and they often require retraining when faced with new tasks.

Gao Jiyang, CEO of Xinghaitu, provided an interesting dual judgment on this issue. In the industrial roadmap presented at this WRC, Xinghaitu clearly marked 2027 as the "GPT moment" in terms of technology, believing that once the scaling law drives the foundational model's capabilities past a turning point, vertical scene applications will begin to accelerate; however, the true "commercialization turning point" is set for 2028, followed by deep penetration in 2029 and widespread deployment in 2030.

However, when discussing whether embodied intelligence will replicate a historical node like ChatGPT, Gao Jiyang clearly stated that he believes "it is highly unlikely that such a moment will occur." The reason is that embodied intelligence will not spread through a software product aimed at everyone instantaneously but is more likely to unlock gradually from To B scenarios, one industry and one task at a time.

The simultaneous occurrence of technological breakthroughs and public recognition is difficult. The explosion of GPT relied on everyone being able to personally experience it through a smartphone or computer, while the first scenarios where embodied intelligence lands are factories, warehouses, and other production ends, which ordinary people find difficult to perceive intuitively. Therefore, the industry turning point will gradually emerge in a "subtle manner," and by the time people suddenly realize it, it will already be ubiquitous.

These two statements are not entirely contradictory. Gao Jiyang actually distinguishes between the technological turning point of model capabilities and the social turning point of industrial diffusion: the former may appear relatively clearly in a certain year, while the latter requires scene validation, large-scale production, cost reduction, and mature business models to gradually transmit to the real world.

This is also one of the most significant differences between robots and large language models. After the release of ChatGPT in November 2022, changes in model capabilities could almost be communicated to global users on the same day. For a software product, as long as the server and computing power can handle it, the cost of adding a new user is relatively limited; users can open a webpage, input questions, and immediately perceive the leap in model capabilities. Thus, technological breakthroughs, product releases, and large-scale user adoption were almost compressed into the same time window in the case of ChatGPT.

Robots, however, must enter the physical world through a real body. Even if a foundational model suddenly gains significantly stronger generalization capabilities tomorrow, this capability still needs to be transformed into real actions through chips, sensors, joints, dexterous hands, and actuators, while also facing a series of engineering issues such as precision, stability, safety, lifespan, maintenance, and cost.

Therefore, embodied intelligence may ultimately find it challenging to replicate a clear historical node like November 30, 2022. It is more likely to manifest as a series of gradually emerging critical points. Perhaps, even in the future, when people look back to confirm the "ChatGPT moment" of embodied intelligence, they may not be able to find a universally recognized node.