Chongqing - Since the beginning of 2025, China’s humanoid-robot sector has experienced a surge in financing. From late June to early July, various Chinese humanoid-robot startups secured substantial funding.

On July 10, at the 2025 Embodied Intelligence Ecosystem Conference, humanoid robots from Unitree performed a dance show. (Photo/MSXF)

Galaxea announced the completion of its series A4 and A5 financing rounds, raising over 100 million USD combined; Unitree confirmed completion of its series C round; Deep Robotics, a humanoid company on a par with Unitree, announced the completion of a new financing round totaling nearly 500 million yuan; Robot Era closed a nearly 500 million yuan series A; and Galbot raised 1.1 billion yuan in a new round, reaching a valuation exceeding 1 billion USD.

Data from IT Juzi, a Chinese investment data service provider, shows that financing in the field exceeded 23 billion yuan (3.2 billion USD) in the first five months of 2025, surpassing the total amount raised in 2024.

China’s humanoid-robot industry also holds vast future potential. According to Morgan Stanley’s report, Humanoid 100, China’s humanoid-robot market will reach 12 billion yuan with shipments of 1.5 million units by 2030. Globally, humanoid robots are becoming a hot investment focus.

The financing boom is driven by the broader “embodied AI” technology sector. Mo Lei, Vice President of AI² Robotics, another startup, said that humanoid robots represent the next revolutionary general-purpose platform after smartphones and new energy vehicles. “Every investor hopes to back the next Apple or Tesla,” he said, emphasizing the vast potential of AI-robotics integration—i.e., embodied AI’s wide application scenarios.

Li Chunzhi, COO of the National and Local Co-built Embodied Artificial Intelligence Robotics Innovation Center, discussed the challenges of humanoid robot deployment on July 10 at the 2025 Embodied Intelligence Ecosystem Conference. (Photo/MSXF)

According to NVIDIA’s official glossary, embodied AI refers to the integration of AI into physical systems, enabling interaction with the real world. These systems include general-purpose robots, humanoid robots, and autonomous vehicles. Li Chunzhi, COO of the National and Local Co-built Embodied Artificial Intelligence Robotics Innovation Center, stated at the 2025 Embodied Intelligence Ecosystem Conference that humanoid robots are the most typical application within embodied AI systems.

Mo noted that the embodied AI is currently in a rapidly developing stage. Since major AI breakthroughs in 2022, the industry has gradually formed a clear technical roadmap, with consensus emerging on key areas such as large-scale models, robot hardware design, and multimodal perception. The current focus is on driving technology to practical application and developing market-ready products; the following three years will see large-scale commercialization, advancing widespread industry adoption.

However, technical challenges remain in the practical application. Li identified three major issues. First, many robots lack sufficient mobility—walking remains unstable, and fine movements are limited, making it hard to navigate complex environments. Second, real-world training data is scarce, with simulated data often failing to reflect reality, leading to poor model performance. Third, robots struggle with generalization, typically confined to narrow tasks and unable to adapt to diverse scenarios— a long way from working with human-like flexibility.

Regarding mobility, Li said that in 2023, industry benchmarks were simply “able to walk a few steps. " In 2024, they evolved to compete on running capability, and in 2025, the focus shifted to upper limb manipulation skills.

On data scarcity, Li explained that robot training data falls into several levels. The publicly available internet data—such as videos of robots making coffee—is easy to access but provides limited value. Enterprises generate the synthetic data through computer simulations. The multimodal data mimics human senses, like vision and touch. The most valuable one is real-world data collected by humans wearing motion capture suits, though it comes at the highest cost. Li noted that Tesla, for instance, pays $30 per hour to hire people to collect such real-world action data.

Zhang Ningning, Deputy General Manager of Baidu AI Cloud’s Smart Industry Solutions, introduced key points for training large AI models on July 10 at the 2025 Embodied Intelligence Ecosystem Conference. (Photo/MSXF)

Zhang Ningning, Deputy General Manager of Baidu AI Cloud’s Smart Industry Solutions, emphasized that improving humanoid robots’ generalization relies on continuous training of large models, which depends critically on computing power. Training often requires mixed use of chips from NVIDIA, Baidu, and Huawei.

Due to architectural differences between chips from different vendors, mixing them results in poor ecosystem compatibility and low cross-platform communication efficiency,reducings model training performance and operational costs. To tackle these challenges, Zhang said Baidu is developing a computing power scheduling platform to integrate and manage diverse chips in a unified manner.

Globally, embodied AI is advancing rapidly. Morgan Stanley also forecasts that it could become a key investment trend in technology over the next decade, with a total addressable market (TAM) estimated at 60 trillion USD worldwide.

Elon Musk’s AI startup xAI recently launched Grok 4, claiming it as “the world’s most powerful AI model.” According to Fortune, Musk plans to integrate Grok into Tesla’s Optimus robots.

Meanwhile, NVIDIA became the first company to surpass a market value of 4 trillion USD, overtaking Apple and Microsoft. Since the generative AI boom in 2022, demand for computing power has surged, making NVIDIA GPUs the preferred infrastructure for large model training by tech giants such as OpenAI, Google, and Meta.