Training Ground Turns AI into Action

At the DeTao Technology Park in Tong'an, a district in Xiamen city, Fujian province in southeast China, an area over 2,000 square meters has been transformed into 10 real-world training zones. There, a humanoid robot bends down to pick up a product and moves to place it on a shelf. Just a few steps away, a wheeled-arm robot is practicing stacking cartons and tying shoelaces. These machines are not putting on a show. They are "learning in school."

The facility, which began trial operations in July, is Xiamen's first heterogeneous robot training facility led by X-Era AI (Xiamen) Technology Co., Ltd. and jointly developed by an AI innovation consortium. The consortium comprises a national-level laboratory, leading AI research institutions and major robotics companies, including X-Era AI. The facility houses various types of robots. Its primary focus is to train robots to accurately perform complex tasks in real-world environments.

"Once AI enters the physical world, simply being able to 'talk' is far from enough," said Dong Haoye, vice president of R&D at X-Era AI.

An action as simple as picking up a cup requires a robot to overcome a series of challenges that humans handle almost instinctively: where is the cup? From which direction should it be reached? How much force should be applied? Will the water spill? The ability to independently perceive the surroundings, make decisions, and execute tasks is a crucial step toward bringing AI into the physical world.

As AI accelerates its penetration of production and daily life scenarios, the training ground is working to accelerate the transition from "algorithms on paper" to "real-world scenarios" by providing enterprises with pilot-scale technical validation based on X-Era Lab's self-developed robot brain foundation.

According to Dong, the technology enabling AI to move from "talking" to "doing" is the team's proprietary large model for physical-world actions. Unlike the mainstream approach of training vision, language and action modules separately and then integrating them, the model incorporates three-dimensional space, physical laws and action trajectories into a unified learning framework from the outset.

Rather than simply learning a fixed sequence of movements, the robots trained here learn the underlying rules of the physical world. This allows them to anticipate and plan the consequences of their actions and generate precise, reliable movements.

To collect high-quality training data from robots, data collection personnel first break down tasks such as folding clothes, tying shoelaces, flexible packaging and product sorting into a series of individual steps. Depending on the task and scenario, they then guide the robots through the exercises using master-slave arm control, virtual reality control and first-person video recording, among other methods, refining the smoothness and reliability of their execution.

Source: Science and Technology Daily