Reference News Network, March 17 report: On March 12, the UK's Financial Times website published a report written by journalist Ryan McMorrow, titled "Entering the Humanoid Robot Training Lab in China." The following is the translated content:

In a new facility covering 12,000 square meters in Wuhan, young Chinese university graduates spend their days operating humanoid robots, making them deliver buns, wipe tables, and fold clothes.

In this 200 million yuan experimental base, every movement of the robots in the simulated kitchen and simulated bedroom is tracked and recorded in real time by cameras and sensors.

The Hubei Humanoid Robot Innovation Center is one of the government-funded robot training bases that have emerged across China, aiming to create large-scale robot-specific training databases.

Zhang Jia (音), the 21-year-old project manager at the base, said: "We are like teachers, and the robots are students. Humans can master something after repeating it a few times, but teaching robots is different. The same action must be repeated hundreds, thousands, or even tens of thousands of times."

Officially, they hope that accumulating such data will help China's still nascent humanoid robot industry, promoting solutions to some major challenges in the process of artificial intelligence moving from software to the real world.

Beijing has included embodied intelligence as one of the key industries to cultivate in its five-year plan from 2026 to 2030.

Experts say that the lack of specialized training data for humanoid robots remains a major obstacle to transforming AI technological breakthroughs into practical robot applications. Large language models such as ChatGPT and DeepSeek rely on massive text data scraped from the internet. Data collection for humanoid robots is still in its early stages.

Beijing has formulated plans to promote the development of the humanoid robot industry, including expanding the scale of robot training and data collection. From the wealthy coastal city of Hangzhou to inland cities like Mianyang, local governments are investing in building new training centers.

Jay Huang (音) analyst at Bernstein Research said: "China's approach to supporting emerging industries facing developmental bottlenecks is becoming increasingly mature, and establishing data collection centers is an example of this." He said: "Government support means that these data can be shared, benefiting the entire industry and driving all participants to work in the same direction."

In Wuhan, Zhang Jia assists in managing 70 young operators who take turns training 46 robots. Operators control the robots via remote controls or sensor-equipped handheld devices, repeatedly performing the same action.

The idea behind this work is that continuous data collected by sensors, along with video footage recording the position, speed, and torque of the robot's components, can eventually be input into a robot-specific artificial intelligence model—commonly referred to in the industry as a 'vision-language-action model.'

This process aims to replicate the breakthroughs achieved by large language models, enabling robots to acquire more general skills, such as independently picking up a bottle of water without specific programming.

Experts say that China's ambitions in the humanoid robot industry also face a more fundamental challenge: currently, data collected from one robot is difficult to directly apply to another robot with different hardware configurations. Moreover, due to the rapid pace of hardware technology updates, the data collected now may not be as applicable to next year's new models.

Jay Huang said that cross-device data transfer capabilities are a current research hotspot, and progress is expected. For example, Google DeepMind's robotic AI model has already shown preliminary potential in skill transfer across hardware platforms. (Translated by Guo Jun)

Original: toutiao.com/article/7618036066330722831/

Statement: This article represents the views of the author themselves.