Full Speech by Wang Xingxing at the 2026 World Robot Conference: Embodied Intelligence of ChatGPT Will Arrive in Two to Three Years at the Fastest

After the listing of Yushu Technology, its founder Wang Xingxing delivered a keynote speech titled "From Exhibits to Products: The Next Decade of the Humanoid Robot Industry" at the 2026 World Robot Conference. He believes that the current robotics industry is at a new starting point for the accelerated development of artificial intelligence, and the biggest bottleneck in the field at this stage is still the insufficient generalization ability of embodied intelligence.

"From the perspective of the development cycle, it could be as fast as two to three years, or as slow as five to ten years, and the industry is expected to usher in a key breakthrough." Regarding the timeline for embodied intelligence to reach a critical point similar to ChatGPT, Wang Xingxing gave a relatively optimistic judgment.

When discussing when general-purpose robots will truly enter households, Wang Xingxing believes that the biggest bottleneck worldwide is still the lack of generalization ability. "If one day everyone can see a robot brought into any unfamiliar environment, and it can basically accomplish about 80% of tasks, I think we would have nearly achieved the embodied intelligence 'ChatGPT moment'."

The following is the speech content organized by PANews:

Reflecting on Yushu's Decade: From Quadruped Robots to Humanoid Robots

Dear leaders and guests, hello everyone. It is a great honor to share here. First, let me briefly review the company's situation. Yushu Technology was founded in August 2016, and this August marks almost a full 10 years. We initially focused on quadruped robots, and in recent years, we have started working on humanoid robots.

One of our more successful products in recent years is the G1, launched in 2024. After its launch, it became a representative product in the global humanoid robot field. Many humanoid robot products you see now, whether in overall shape or design concept, have some similarities with the G1.

Earlier this year, a representative project was our performance "WuBOT" during the Spring Festival Gala. This program integrates very classic traditional Chinese kung fu and martial arts culture. We collected dozens of representative Chinese kung fu movements in advance, trained them with AI, and finally showcased the most exciting movements on stage. The highlight of this program is the combination of cutting-edge humanoid robot technology with traditional Chinese kung fu and martial arts culture. This video has also spread very well overseas, possibly with hundreds of billions of views. I think this is a good example of the fusion of technology and culture.

In February, we also did a demonstration at the Temple of Heaven. At that time, about 49 robots performed fully automatically. The Temple of Heaven itself has hundreds of years of culture and history, and robots performing in such a setting creates a very strong sense of time travel—melding China's past culture with the latest technology, both visually and emotionally stunning.

In recent years, we have been promoting robots to truly enter daily life and factories. In 2024, we began collaborating with automotive factories to implement some practical applications, and we are also deploying robots in our own factory for simple testing and applications. Currently, the vast majority of our AI team is focused on how robots can truly enter households or factories to work.

Why is there still no large-scale promotion? The main reason is that the efficiency and generalization ability of robots are still insufficient. Robots can now perform some simple assembly tasks, but their efficiency is still lower compared to humans; additionally, every time they encounter a new task, they often need retraining, which further reduces overall efficiency. Therefore, we hope to improve the generalization ability of the technology before proceeding with large-scale promotion. This is also a core bottleneck faced by the global robotics industry.

In April this year, we also set a new record for the running speed of humanoid robots. This robot was modified based on our first-generation humanoid robot H1, launched in 2023. H1 is Yushu's first humanoid robot and is a classic product. Subsequently, we have also tried different forms based on it, such as adding wheeled structures to make the robot more flexible in certain scenarios.

Driving Robots: AI, Data, and Real-World Scenarios

As you know, current robots are basically AI-driven, and AI is essentially data-driven. The amount of data, especially high-quality data, largely determines the capabilities of AI. Therefore, on one hand, we collect data ourselves, and on the other hand, we collaborate with third-party companies to obtain more data for both industry use and our own use.

I personally believe that the data used for training AI robots should primarily consist of a large amount of real human data or internet data as the main pre-training data, supplemented by some real machine data. Both types of data are essential. In terms of data scale, real human data will be more abundant and important; however, real machine data is also indispensable because ultimately, we need to match robots with the real physical world.

In May this year, we launched the world's first mass-produced manned transforming mech GD01. This robot stands over 3 meters tall and weighs about 500 kilograms when carrying a person. You might wonder what applications such machines have. To put it simply, it is somewhat like an "off-road vehicle"; it is not primarily aimed at daily use in households or cities but is more suitable for outdoor, complex environments, and some transportation scenarios. For example, it can be used for outdoor hiking or transporting goods in complex terrains.

GD01 also has an interesting feature: it can transform into a quadruped mode. Switching to quadruped mode improves stability and enhances obstacle-crossing ability. This is also why we are willing to truly offer it for pre-sale: in quadruped mode, it already possesses relatively good stability and the ability to navigate complex terrains. In simple terms, it can be understood as a robot-shaped "off-road vehicle".

A few months ago, we also demonstrated an end-to-end AI action generation system based on multimodal models. In the past, many robot action demonstrations were pre-collected and pre-trained; however, this system can generate actions based on real-time voice commands. The voice needs to be recognized first, then uploaded to the cloud for AI action generation, and finally, the actions need to be verified to ensure there are no issues before being sent to the robot, so the entire process currently has a few seconds of delay.

We hope that in the future, when robots are truly applied, actions can be generated completely in real-time, rather than relying on pre-programming. This automatic generation method also has a characteristic: even if the same command is input, each generated action may vary slightly. In other words, if you tell the robot the same thing today and tomorrow, the actions it performs may not be exactly the same.

We are also promoting robots to truly enter daily scenarios like meeting rooms. For example, if a meeting room in a company is sometimes messy, then organizing the meeting room is a clear and practical task. Our training goal is: after randomly messing up a meeting room, the robot should be able to restore it to a clean and tidy state.

This task is currently fully end-to-end implemented, so the robot's movement efficiency is still relatively slow. However, it is not a single-task system. In this scenario, we have set about 7 to 8 different tasks, allowing the robot to automatically switch between different tasks and execute them directly, while also possessing a certain degree of interference resistance. Because AI needs to recognize in real-time, generate in real-time, and continuously reason, the actions are not yet very smooth; every step and every action requires reasoning, which affects the overall "smoothness". However, the demonstration video itself is a continuous shot without editing, and environmental interference has been included.

On this basis, we have combined voice-driven and multimodal models, allowing robots not only to generate actions based on voice but also to truly complete interactive and operational tasks. In the live demonstration, the robot can recognize the color of objects in front of it, determine the positions of different medicine boxes, and then retrieve the specified object based on verbal instructions. In this case, the robot is not executing a pre-set fixed action but needs to understand language, recognize the environment, and then complete the operation.

Recently, we also released the latest generation of wheeled-quadruped robot As2-W. This robot is relatively lightweight, weighing about 25 kilograms with a battery, and has strong load and endurance capabilities, while also possessing IP54 dust and water resistance, making it suitable for both indoor and outdoor scenarios.

We have reapplied some technologies accumulated in humanoid robots to quadruped robots, hoping to further enhance the flexibility of quadruped robots while improving load, endurance, and adaptability to complex environments. We also hope that this robot can truly be used for outdoor hiking, transportation, and some industrial scenarios in the future.

Meanwhile, we continue to upgrade lightweight humanoid robots like the R1 to further improve their movement and flexibility. The R1 itself has a relatively high cost-performance ratio and is already available for purchase through platforms like JD.com and Taobao.

Recently, we also showcased a new robot. This robot was developed in a short time, taking about three months so far, and is not yet a formal release version. Its top speed has reached 12.65 meters per second, exceeding the peak running speed of the fastest human athlete in history; according to current test results, its jumping height has also surpassed the highest jump level in human history. In 2025, I, the company's products, and Yushu Technology were also featured on relevant lists in Time magazine.

Core Bottleneck of Embodied Intelligence: Generalization Ability and Alignment with the Real World

Looking back at the development of embodied intelligence AI models in recent years, the main technical routes include VLA models and world models. This year, the world model direction has probably been the most discussed.

Our company has invested heavily in AI models, which is one of the areas with the largest financial and human resource investments. In early 2020, we began exploring video-based world models. By the end of 2020, the results were not particularly ideal, so we set it aside for a while; however, starting last year, we increased our investment in the direction of video-based world models again.

Many people ask: When will truly general-purpose robots, whether humanoid or semi-humanoid, really enter daily life and households?

I believe the biggest bottleneck currently is still the insufficient generalization ability of embodied intelligence. Many AI models can achieve nearly 100% success rates in fixed scenarios as long as sufficient data collection and training are done. However, as soon as the objects being manipulated change slightly or the environment undergoes even minor changes, the success rate drops significantly.

I hope that in the future, it could be as fast as two to three years, or as slow as five years or even longer. If one day we can bring a robot into a completely unfamiliar environment, such as a household it has never seen before, and it can accomplish about 80% of tasks just through voice or language commands, then I think we would have nearly reached the embodied intelligence "ChatGPT moment". This will also become a key critical point for the true explosion of the future industry.

This evaluation metric is actually quite simple: in about 80% of unfamiliar scenarios, the robot should be able to complete about 80% of tasks solely through voice or language commands. I think this is a very important critical point.

Why are we not yet at this level? One of the biggest bottlenecks is that the alignment between the input and output of AI models and the real robots, and the real physical world is still insufficient.

For example, when you ask a robot to move or to assemble two things together, or to move an object from one place to another, its general direction is often correct; it can basically understand your task and has a rough idea of how to execute it. However, what truly affects the final success rate is often the last few centimeters or even the last few millimeters.

When a robot is about to grab something, it may seem like its hand is almost there, but the final tactile feedback and the last bit of positional error cannot be corrected well. If this error is not corrected properly, it may lead to the failure of the entire task, and the success rate will drop significantly.

This is a very core issue faced by embodied intelligence worldwide: the discrepancy between the input and output of AI models and the real world.

Why does this problem exist in robots, while general language models or multimodal models do not exhibit such obvious issues? The content processed by language models is essentially numerical encoding. The input and output are strictly confined within a numerical space or vector space, so errors do not accumulate in the same way.

But robots are different. Every input and output of a robot must enter the real physical world, and every action executed generates certain deviations and losses. These deviations affect the next perception and the next action, ultimately leading to error accumulation. Because of this, the generalization ability and actual task success rate of current robot models have not yet reached a sufficiently high level.

I believe this issue can be resolved in the coming years, but it does require some time.

Enabling Robots to Self-Evolve: Using AI to Restructure the Robot Development Process

Here, I would like to introduce something we are currently promoting: directly enabling physical AI robot models to achieve self-evolution.

In recent years, AI has been widely used for programming and various AI development, but the use of AI itself in the robot development process is still relatively lacking. What we are currently promoting is to try to use the most advanced and top-tier large models to drive robot development.

Specifically, we can first set some rules, experiences, constraints, and tools for the AI, allowing it to search for the latest papers, the best research results, and various open-source solutions online. Then, let it program itself and generate the control code for the robot.

Once the control code is generated, it can be run and trained in a simulation environment, or it can further call models and control systems to drive real physical robots.

If deployed to physical robots, we can conduct real tests and evaluate the results. Part of the evaluation can be completed by the robot AI model itself, while another part can involve human participation. Human experience is still very important because, many times, humans can easily judge whether an action is performed well or poorly.

If this logic can truly form a system and operate, the overall development efficiency of robots will significantly increase, and the iteration speed will be very fast. After this process runs for a while, there will be a chance to truly enhance the self-evolution capability of robots.

Here, I can give a very simple example. We can use a coding intelligent agent to write the control code for the robot itself. Currently, many relatively simple deep reinforcement learning algorithms have already shown good results when using coding agents for automatic programming.

After the code is generated, we can place it in a simulation environment for validation and training; once the results reach a certain level, we can deploy it to physical robots for testing. After testing, one part of the evaluation can be done by the AI model, while another part can involve human scoring to assess the effectiveness. Then, we can feed these results back to the programming intelligent agent, entering the next round of code generation and optimization, ultimately forming a continuous iterative positive cycle.

What I just mentioned is just a very simple example; the actual use will be much more complex. However, I believe this is something very worthwhile for the global robotics industry to pursue in the present and the coming years.

Why do this? There are several direct reasons.

First, the capabilities of foundational models are improving every month and every year, so the self-evolution cycle's capabilities will also improve accordingly. In other words, this system can naturally strengthen itself alongside the progress of global AI foundational capabilities.

Second, it can make fuller use of diverse data. In the past, relying on humans to process and utilize data was relatively inefficient. If a self-loop can be formed, we can utilize various training data, real-world data, and human data, and the efficiency of data utilization will be higher.

Third, it can form a closed loop with more real machine deployments. As real robots are deployed more and more, we will continuously obtain new testing data and more evaluation metrics. This way, the data utilization rate and growth speed of the entire system will be guaranteed.

Fourth, it can continuously accumulate the skills of robots. We can keep adding new skills every day or every month, rather than developing a skill today and wasting it tomorrow or starting over from scratch. Skills, data, control strategies, and evaluation results can all be continuously accumulated in this system.

Therefore, I think this is a very important matter: truly using AI to significantly enhance the development and evolution efficiency of robots. In the future, this approach may become an important path for the continuous evolution of physical AI robots and may open a new phase.

I believe that more and more people will promote this direction in the future.

Looking back, Yushu has been through almost 10 years. We first participated in the World Robot Conference around 2017 or 2018, and over the years, the experience has been profound.

Standing at this new stage today, especially against the backdrop of rapid AI development and the global focus on robots, the evolution speed of robots has already significantly accelerated. The future development speed of robots may be even faster than I currently estimate.

I believe this is a brand new starting point and a brand new beginning.

Due to time constraints, today's report is relatively simple. Thank you all, and I ask for your understanding.