At the Bund Summit, Wang Xingxing Shares New Perspectives on Robots

On September 11, at the 2025 Bund Summit, Wang Xingxing, founder and CEO of Unitree Robotics, shared his latest perspectives. He offered his views on issues such as robot models, data, and hardware.

Wang Xingxing believes that the entire field of getting AI to do work is currently a desert, with only a few small grasses growing; the eve of explosive growth has not yet arrived. Moreover, current hardware is more than sufficient—even hardware from one or two years ago is enough.

But Wang Xingxing pointed out that current AI models are not yet capable of effectively utilizing hardware, and the robotics field faces significant challenges. In response, he shared some of his perspectives. 42 Wave edited this content without altering its original meaning; the following is for reference only.

Data and models are both important, but the model still struggles to leverage hardware effectively.

On the data and model issues that are of great concern in the industry, Wang Xingxing believes that both data and models are currently very important. From a data perspective, there are significant issues with robot data, whether in terms of collection or data quality. Furthermore, there are no established standards regarding how to collect high-quality data, what level of data quality should be achieved, what types of data should be collected, or the scale of collection; the field is still in a relatively ambiguous stage.

Moreover, it is essential to maximize data utilization efficiency. If the model itself possesses a stronger capacity for understanding data, less data will be required. From the perspective of the model, one can identify which parts of the data are more valuable, thereby enabling targeted data collection. In the context of language models, many scenarios require distinctive or characteristic data rather than merely relying on volume, as certain characteristic data points are critically important.

At the Bund Summit, Wang Xingxing shared new perspectives on robots. He noted that current models struggle with multimodal fusion; while standalone language and video models perform well, effectively integrating language and image is a significant challenge today. Controlling details in image and video generation via text alone is difficult, but using sketches or line drawings as references yields better results.

In the robotics sector, aligning robot control modalities through language or other methods remains a major hurdle. For instance, using video generation to instruct robots to perform household chores is challenging: although video generation quality has improved, achieving precise alignment between generated content and robotic control inputs is still highly demanding.

Wang Xingxing believes that current hardware—and even hardware from one or two years ago—is more than sufficient. The primary bottleneck lies in AI models themselves, which lack the capability to fully leverage existing hardware. For example, effectively utilizing dexterous hands is extremely difficult for current AI. Whether it involves data collection or enabling dexterous manipulation beyond simple grasping, these tasks remain highly challenging for the field of AI.

The domain where AI performs actual work is still in its desert phase

Neither AI nor robotics has seen development progress as optimistically as previously anticipated. Wang Xingxing, who has been working in robotics for over a decade, recalled that his first project in 2009 was a small bipedal robot. Later, Unitree Robotics shifted its main business toward humanoid and quadruped robots.

Although Unitree Robotics is now a star company in the robotics industry, Wang Xingxing expressed some regret when discussing AI. He shared an anecdote: his biggest regret was initially being passionate about AI around 2011, when it was still niche. After reading available books, he felt the practical applications were limited, so he did not devote much time to AI subsequently, focusing instead on robotics for several years.

At the Bund Summit, Wang Xingxing shared new perspectives on robots. He believes that recent developments in AI, including robotic AI models, have provided him with another opportunity to seize the AI era and enable AI to truly take action and perform work. Current language models are already outstanding in the information domain, including text and images, surpassing 99.99% of people. However, the field of making AI perform tasks is still in a desert stage; only a few small grasses have grown on this desert, and the eve of true large-scale explosive growth has not yet arrived. But this era is very exciting. The AI era is very fair: as long as you are smart, want to get things done, and aim to achieve your goals, everyone has the opportunity to grow towering trees from the desert.

AI is an Opportunity for Young People

In the AI-native era, Wang Xingxing believes that recent years, especially for the younger generation, including students who are still studying, present a very good opportunity. Previously, programming mainly involved writing basic code from scratch, but now it involves using tools, including AI. Current AI models can be regarded as a very good set of pre-programming tools. In the past, people might have been adjusting some basic code, but now they can use more advanced model capabilities to create works, whether generating content or coding AI agents, which is much more convenient than before.

Wang Xingxing stated that people's perception of AI models themselves can be more radical. Instead of treating them merely as models, we should view them as all-around toolsets. We need to forget some past concepts, relearn, and accept them to generate more new inspiration. Relying on past experience for future decision-making is not beneficial. Instead, grasping what is currently happening and making new decisions is more likely to lead to new creations.