The Global AI Race: The Case for Embodied AI (Part 2)

(Continued from The Case for Embodied AI – Part 1)

For most of its history, artificial intelligence has existed behind a screen. It has analyzed, predicted, generated, and optimized, all without ever touching the world we wanted it to understand. That’s beginning to change. A new generation of systems are emerging that can move, manipulate, and learn through interaction, much as we ourselves do. No longer theoretical, embodied AI is quickly becoming a reality.

And like every transformative technology before it, it is being driven by many different mindsets and approaches.

In the United States, the center of gravity lies in what might be called the “learning layer.” Companies like NVIDIA are building the infrastructure that allows machines to simulate reality, generate synthetic experience, and train at scale. Around that foundation, a new class of companies is emerging with an ambitious goal. Figure AI, Skild AI, and Physical Intelligence are not just building robots. They are building systems designed to improve through interaction. Each action becomes data. Each mistake becomes a lesson. Intelligence, in this model, improves, compounds, grows.

More established robotics companies are also beginning to shift in this direction. Boston Dynamics, long known for its precision-engineered machines, is now pairing its physical platforms with learning systems developed alongside Toyota Research Institute. Meanwhile, Google DeepMind continues to develop world models that allow machines to anticipate how environments evolve over time. Another important example is Covariant, which is deploying AI-driven robotic systems in warehouses that learn from millions of real-world picking interactions, steadily improving performance through experience.

But the U.S. is hardly the only center for developing embodied AI. In fact, China is rapidly becoming the leader in building the robotics that will enable this new form of AI learning.

Companies like AgiBot, Unitree Robotics, and UBTECH Robotics are moving at extraordinary speed. Their focus is less on perfecting a single system and more on deploying many. That matters because in embodied AI, learning is driven by interaction. The more machines in the field, the more data they generate, and the faster they can potentially improve.

This is a familiar pattern. China has followed it in electric vehicles, drones, and solar manufacturing. Now it is applying the same logic to robotics. Unitree is pushing aggressively on cost and iteration. AgiBot is building large-scale data environments to accelerate learning. XPENG is attempting to bridge autonomy in vehicles with humanoid systems, potentially compressing years of development into a much shorter window. Even more focused players like Galbot are quietly advancing manipulation, which may prove more commercially valuable than full humanoid form factors.

China’s advantage is not just technical. It is systemic. Tight supply chains, rapid manufacturing cycles, and national-level prioritization are creating a dense ecosystem where iteration happens quickly and at scale.

Europe, by contrast, is taking a different path.

Rather than chasing scale or spectacle, European leaders are focusing on reliability, safety, and integration into real-world environments. Companies like ANYbotics are already operating in industrial settings where failure is not an option. 1X Technologies is pushing toward general-purpose robots that can function in human spaces, particularly the home. Meanwhile, firms such as NEURA Robotics are working toward what they describe as cognitive robotics, systems that combine perception, reasoning, and action in a more unified way.

Supporting all of this is a deep institutional foundation. Organizations like ETH Zurich and Max Planck Institute for Intelligent Systems continue to produce some of the most important work in robotics and intelligent systems. Europe may not dominate headlines, but it is quietly shaping the standards and expectations that will determine how these systems are trusted and deployed.

Across all three regions, a pattern is emerging. The field is dividing into two distinct camps. The first builds machines that perform tasks. The second builds systems that improve by performing them.

That distinction is more than technical. It can be evolutionary.

A robot that can sort packages or inspect pipelines is valuable. A robot that becomes better at those tasks simply by doing them is something else entirely. It is a system that accumulates experience. Over time, that experience becomes a form of intelligence. Not intelligence like our own, but intelligence that can supplement it in evermore powerful ways.

This is why the most important players in embodied AI are not necessarily the ones with the most impressive demonstrations. They are the ones building closed learning loops, systems where action leads to feedback, feedback leads to adaptation, and adaptation leads to improved performance across contexts.

In the United States, that loop is being built into the software layer. In China, it is being accelerated through deployment at scale. In Europe, it is being refined through real-world constraints and expectations of safety and reliability.

The future will not belong to any one of these approaches in isolation.

It will belong to whoever combines them in the most useful, reliable ways.

Because in the end, embodied AI is not about giving machines a body. It’s about giving intelligence a way to learn from the world itself. Once that happens, intelligence stops being something we program. It becomes something that can grow on its own.