Professor Yang Yaodong from Peking University's Institute for Artificial Intelligence, who also serves as the Chief Scientist at Lingchu Intelligent, recently shared his insights on the development of the embodied artificial intelligence sector.

Addressing Industry Competition

When discussing the intense competition within the embodied AI field, Professor Yang stated that such a situation is entirely expected. The progression of large language models and the entire embodied AI domain operates on a logarithmic scale encompassing four orders of magnitude, where each step represents a tenfold difference. Therefore, the pressures felt from capital, technology, product development, and commercialization during advancement are very typical.

He pointed out that every dimension will undergo a reshuffling process. This includes not just data, models, and computing power, but also the number of companies, funding shares, product development cycles, model iteration speeds, and data volume growth. The significant gaps across these orders of magnitude translate into the individual experience of intense market competition.

The Current State and Expectations

Professor Yang noted that following this year's Spring Festival Gala, public discourse has been questioning whether robots can truly perform practical work. He expressed a shared anticipation for the "ChatGPT moment" in embodied AI, but acknowledged that, regrettably, there is still some distance to cover before that milestone is reached.

Achieving this pivotal moment is a crucial prerequisite for robots to genuinely begin undertaking useful tasks. Currently, both industry and academia are iterating rapidly, and the substantial funding raised is aimed at accelerating this very process.

Challenges in Scaling and Data

He highlighted that for embodied intelligent robots, there is currently no mature, large-scale application model at what could be considered an "L2" level. The data production process itself is a capital-intensive manufacturing环节. Early this year, estimates suggested that revenue from data manufacturing and services could account for as much as 90% of the total, whereas selling the robots themselves is not highly profitable. The price of a single robot is not very high, almost sold by the weight of its metal, due to the fiercely competitive nature of the industry.

Professor Yang emphasized that the greatest bottleneck for embodied AI models to achieve true generalization and accomplish various universal tasks remains data. An academic consensus suggests that approximately 100 million hours of human operational data are needed to support a qualitative leap in model capabilities. Just considering hand operations, the current cost of collecting one hour of such human operational data domestically is around 300 yuan. For embodied AI companies, despite significant investment, the total collected data remains far from the 100 million-hour target.

Furthermore, this cost is specific to China. Collecting similar data in Silicon Valley or other regions of the United States would involve significantly higher labor and time expenses. Therefore, from a global perspective, the data bottleneck represents a critically scarce resource for training viable embodied AI models.