In the embodied AI race, China can opt to look beyond bigger models

The global fascination with robots has surged in recent years, with advancements in their abilities to perform a variety of tasks. However, these captivating feats do not necessarily equate to commercial viability. One significant challenge faced by all players in the robot industry is the availability of data. Achieving widespread adoption of robots necessitates them to be versatile, dependable, and economically feasible simultaneously.

Currently, no entity has managed to conquer all three aspects of this puzzle. The primary obstacle lies in the scarcity of high-quality physical interaction data, which is essential to fuel the development of advanced robotic intelligence.

Building sophisticated AI models for general-purpose robots may require hundreds of millions of hours of real-world interaction data. However, as of early 2026, the global repository of compliant and high-quality physical interaction data had only reached approximately 500,000 hours. Researchers worldwide are in pursuit of innovative solutions, yet different regions are adopting distinct strategies.

In the United States, tech behemoths are favoring a scale-driven approach, focusing on larger models, increased computational power, and extensive AI infrastructure. This strategy is particularly advantageous for companies with substantial financial resources and computational capabilities.

Europe and Japan prioritize regulation, safety, and hardware advantages but have yet to solidify their position in AI algorithms and large-scale data ecosystems. Geopolitical tensions have also escalated, with the United States imposing restrictions on Chinese-made humanoid robots, highlighting the intertwining of hardware accessibility and geopolitical competition.

While hardware limitations can be erected, data exchange is more challenging to control, thereby creating an opportunity for China to adopt a different approach in the realm of embodied AI.

China's strategy in embodied AI is to move beyond constructing larger models and instead emphasize data efficiency. Its greatest asset lies not only in its abundant engineering talent, extensive supply chain, and abundant capital but also in its highly developed physical world, which serves as an extensive source of real-world tasks for robots to learn, adapt, and be tested in.

China's manufacturing ecosystem possesses an enormous, continuously evolving pool of real-world scenarios where robots can acquire experience. Consequently, China stands to benefit significantly by converting its manufacturing depth into a competitive advantage in the robotics sector through the accumulation of task experience.

As the AI industry has traditionally relied on scaling models through increased parameters and computational power, China's manufacturing environment suggests an alternative approach: scaling through a greater number of tasks, experiences, and skills learned from the real world. Presently, robotic skill acquisition typically begins in large teleoperation centers, where humans manually control robots to collect training data.

However, much of this data is tied to specific hardware, compelling companies to restart from scratch when switching robots. A burgeoning trend in China is the utilization of technologies like the Universal Manipulation Interface, which captures human skills in a reusable format that can be applied across various platforms. Lumos Robotics, the company I lead, has observed that such methods can reduce data collection costs to approximately one-fifth of those incurred through conventional approaches.

Beyond the debates surrounding task experience and data efficiency, China's AI sector also presents a valuable lesson in engineering wisdom. Companies such as DeepSeek and Moonshot AI have demonstrated that breakthroughs do not always necessitate augmenting computational power. By implementing system-level optimization, algorithm innovation, and hardware-software co-design, engineering efficiency can rival brute-force scaling.

The same principle applies to embodied robotics, where competition transcends the mere construction of larger models. Instead, it involves extracting maximum intelligence from every data point, interaction, and real-world experience.

In the broader context of the global competition to develop ever-larger AI models and expand computing infrastructure, China is pioneering a more cost-effective route that could potentially inspire technological communities in developing nations by lowering the barriers to innovation. China's electric vehicle industry serves as an illustrative example.

A decade ago, advanced features like head-up displays and adaptive air suspension were primarily available in luxury vehicles. Today, China's supply chains and engineering capabilities have enabled the integration of these technologies into mainstream vehicles priced below US$40,000. Embodied intelligence may follow a similar trajectory.

If high-quality training data and robot skills become affordable, reusable, and scalable, the future of robotics will not be monopolized by a handful of tech giants with seemingly limitless financial resources and computational capabilities. Instead, a larger number of startups, academic institutions, and research teams worldwide will gain access to affordable, reusable, and scalable robotics technologies, thereby democratizing innovation.

China's approach to the embodied intelligence race indicates that the future will not be determined by those with the largest models or the most expensive robots. Rather, it will belong to innovators who can establish a complete closed loop: collecting real-world experience, transforming data into intelligence, converting intelligence into reusable skills, and delivering measurable value across various industries.

While this opportunity is currently exclusive to China, it is not exclusively theirs to claim. It is an opportunity that can be shared by humanity as a whole, as we strive to dismantle the walls of technological monopolies that impede others from participating in and reaping the benefits of the next wave of AI. The ultimate goal is to foster a more open ecosystem that enables broader participation and collective advancement in the realm of artificial intelligence.