Shijingshan district in Beijing, on Friday, opened the third phase of its humanoid robot training center, combining advanced simulation and tactile technologies to create China’s largest robotics database.
The project, led by Beijing Shijingshan Science and Technology Innovation Co and Beijing Tashan Technology Co, provides simulated environments such as homes, hotels, supermarkets and factories, where more than 100 robots can train and collect essential data.
Dong Jianqiang, head of Tashan’s training platform, said robots first learn tasks in a controlled simulation scenario, then conduct real machine training for about a week before being deployed in real-world applications. The data collected serves three main purposes: supplying datasets to companies developing advanced AI models, training robots for specific applications, and supporting larger projects that include building scenarios, generating data and scaling up deployments.
What sets the project apart is its combination of vision, movement and tactile sensing. While visual systems alone can struggle with blind spots or capturing fine details, the addition of touch sensors allows robots to handle delicate tasks, such as gripping eggs without cracking them or manipulating fragile items like potato chips and tofu. A robot has to sense how an object changes shape when touched and how it returns to normal when released. “Physical contact data is essential for large-scale deployment in the future,” Dong said.
The third phase has introduced dozens of robot models from multiple manufacturers, allowing cross-platform training that was not possible in earlier phases, which relied on a few models from a single brand. Dong said this flexibility is key to developing more versatile robotic systems.
According to Fu Hang, director of the Zhongguancun Shijingshan Park integrated service center, the first two phases laid the groundwork for large-scale data collection. Phase one, completed in March 2025, deployed 100 wheeled robotic arms, creating Beijing’s first specialized data training facility. Phase two added 100 humanoid robots across multiple scenarios, generating millions of high-quality data points per year, Beijing News reported.
The project plans a fourth phase to introduce “data collection gloves”, offering a simpler and more agile way to gather training data, it added.
The push for such projects is fueled by strong market demand. A report released by Zheshang Securities on March 22 estimated that by 2030, China and the United States will need about 2.1 million humanoid robots for manufacturing and household services, representing a potential market of around 314.6 billion yuan ($45.6 billion).
The report added that humanoid robots are expected to see large-scale adoption first in industrial settings before gradually moving into household applications over the medium to long term.
The pace of adoption differs mainly because of environmental complexity. Dong explained that industrial sites are relatively controlled, with standardized components and predictable conditions, making it easier to deploy robots efficiently.
“Household environments, on the other hand, compress the full complexity of the real world into a small space,” he said. “This makes operations less predictable and more challenging, so widespread use of robots in homes is likely to take longer — perhaps five to 10 years — before we see meaningful progress.”
Yang Cheng contributed to this story.