Chinese robot makers pivot from physical stunts to intelligent AI brains

The shift to embodied AI

Chinese humanoid robots have recently captured public attention by performing impressive physical feats like dancing, running, and doing backflips. While these demonstrations showcase high-level engineering, the industry is now moving toward a more complex goal. The next major hurdle is providing these machines with the cognitive ability to carry out actual work without needing specific instructions for every individual action.

This transition is defined by the rise of embodied AI. This technology allows machines to understand and interact with the physical world rather than just processing text or images. Unlike a chatbot, an embodied system must learn how to grip specific objects, navigate around unexpected obstacles, and handle unfamiliar situations in real time.

The focus is moving away from simple hardware improvements toward building the AI brain. The goal is to create systems that can understand their surroundings and learn new skills on the fly. This shift is turning the robotics landscape into a competition over who can build the most adaptable and general-purpose intelligence.

A massive investment boom

The push for smarter robots is driving a significant surge in capital. According to research firm ITjuzi, firms in China building robot intelligence gathered 93.5 billion yuan ($13 billion) through 322 separate deals during the first half of 2026. This total represents a fivefold increase compared to the amount of funding recorded during the same period in the previous year.

This influx of money is being directed toward companies that specialize in robot intelligence. Some specialist startups in this space have already reached valuations of around $3 billion, according to company disclosures. This highlights how much the market values the software and data required to make robots functional in the real world.

Established manufacturers are also joining the race to secure talent and technology. For example, UBTech recently advertised a chief scientist position with a compensation package of up to 124 million yuan. This includes equity and benefits, showing that the competition for the best minds in robotics AI is becoming increasingly intense.

How robots learn to work

Training a robot is fundamentally different from training a standard chatbot. While chatbots learn from internet text, robots require examples of physical actions to understand how the world works. Companies are currently using several different methods to collect this necessary data to train their models.

Some firms are using massive human-led data collection efforts. Spirit AI, for instance, has hired approximately 1,000 people to record everyday and industrial activities. By capturing these movements, the AI can learn the nuances of human behavior and task completion in various environments.

Other developers are using a mix of human-guided movements and computer simulations. Robbyant reported training its model using 160,000 hours of material, which included 50,000 hours of human-guided robot movements. While simulations help teach machines industrial tasks, developers acknowledge that these digital environments cannot perfectly reproduce all real-world conditions.

The global robotics race

The competition for robotics dominance is global, with the United States also developing systems to help robots learn tasks. Companies like Physical Intelligence and Skild AI are working on similar AI systems. This creates a high-stakes environment where both regions are trying to perfect the underlying intelligence of humanoid machines.

China currently holds a significant advantage in manufacturing and access to industrial environments for testing. This allows Chinese developers to put their robots into practical use more quickly than many competitors. However, they still face a major hurdle in the form of restrictions on advanced Nvidia chips, which are essential for high-level AI processing.

The ultimate milestone for the industry will be creating robots that can complete unfamiliar tasks reliably. Unitree founder Wang Xingxing suggested that future AI models could eventually lead to a "ChatGPT moment" for robotics. However, experts warn that a successful demonstration does not mean a product is ready for commercial use, as making robots dependable across different workplaces may take years.