As the government has set in motion the development of a domestic "robot brain," solving the data shortage—the biggest bottleneck to implementing physical AI (AI that moves in the real world) that drives robots—has emerged as the top task. That is because collecting action data needed to make "smart robots" that move and decide on their own takes substantial expense and time. Overcoming pushback from highly skilled workers in manufacturing sites such as autos, shipbuilding, and steel—who are essential for securing high-quality industrial site data—is also cited as a challenge.
Bae Kyung-hoon, Deputy Prime Minister and Minister of the Ministry of Science and ICT, on the 29th designated physical AI as a national strategic industry and declared that within three years Korea will build a general-purpose physical AI foundation model that understands the real world and predicts actions. Bae said, "The golden time to become No. 1 in physical AI is the next three years," adding, "With physical AI, we will raise the productivity of leading industries by 20% to create an ultra-gap, and achieve the spread of home robots and zero deaths from industrial accidents."
Robots based on physical AI are distinguished from existing robots that operate passively under fixed rules in that they must move directly and carry out tasks in line with constantly changing surroundings. There is not enough category-specific action data to train such active robots, so the government and corporations must start building it now. According to Ken Goldberg, a professor of industrial engineering at UC Berkeley and an expert in physical AI, while Generative AI such as ChatGPT, Claude, and Gemini has been able to learn from 100,000 years' worth of text data—books, articles, and websites published to date—the data needed to train robots amounts to about 10,000 hours.
The problem is that if various motion data—such as a robot reaching out and picking up an object—are collected one by one, it will be impossible to secure the necessary data within the three years the government set as the golden time. It is not only time that is lacking; the expense is also significant. According to the Silicon Valley Robotics Center (SVRC) in the United States, producing 500 demonstration data samples through "teleoperation," a method in which a person remotely controls a robot to collect on-site data one by one, costs at least $50,000 (about 77 million won) to as much as $200,000 (about 300 million won). Here, 500 refers to high-quality data at a level usable for actual services, and the total expense includes only the cost of robots and various equipment (hardware) used for data collection, excluding labor costs.
On top of that, when adding labor costs for researchers and developers overseeing teleoperation, infrastructure expenses for building dedicated workspaces to conduct it, and post-processing expenses to convert collected data into training data, it is estimated that at least tens of millions of won more will be needed.
In response, the government decided to supplement teleoperation (on-site data collection) by generating synthetic data through virtual simulation. It will use digital twins that replicate real environments in virtual space to produce synthetic data. Through this, the plan is to secure large amounts of robot training data within the short span of three years and cut expense.
Major U.S. corporations such as Nvidia and Google are actively using large-scale synthetic data production. Nvidia takes a small amount of real demonstration data obtained by remotely controlling robots and has robots learn repeatedly in diverse scenarios at virtual factories. For example, when simulating on-site data of a robot picking up a cup and placing it on a shelf, changing the cup's size and weight or moving its position, and adjusting the shelf height—altering the work environment—can mass-produce from thousands to millions of virtual training data samples.
In the past, robots had to go through trial and error in real environments over long periods to collect data, but by using simulation, companies say they can produce large-scale, high-quality synthetic data with only a small amount of on-site data, let robots learn from diverse situations that are hard to collect in reality, and drastically reduce expense and time.
Using this approach, Nvidia early this year unveiled the world model "Cosmos 3," trained on about 400 million real and synthetic videos, 1 billion images, and human and robot action data. Rev Lebaredian, Nvidia's vice president in charge of simulation technologies, said, "The success of physical AI depends on how quickly you can secure vast amounts of data."
China, by contrast, has chosen a strategy of using manufacturing sites as robot training grounds under full government support. Viewing physical AI as part of a strategy to advance manufacturing, China is focusing on combining AI and robots on industrial sites through its "AI + Manufacturing" policy. Notably, Chinese robot corporation Ubtech this year deployed about 1,000 units of its Humanoid Robot "Walker S" to local auto plants and other sites to begin on-site training and application. China is also operating 20 regional data factories that collect robot data. There, people wear Virtual Reality (VR) devices and the like and repeatedly perform actions such as folding clothes, wiping tables, and stacking items to create motion data for robot training.
Korea plans to pursue both the U.S. strategy of producing synthetic data and China's approach of integrating with manufacturing sites. As part of that, it will work with major domestic manufacturing corporations and startups to gather on-site data and build data factories at regional hubs that can produce and process real action data and synthetic data.
However, strong resistance from manufacturing-site workers and skilled technicians worried about job losses is a challenge facing the government and the related physical AI industry. An official at a domestic physical AI corporation said, "To secure high-quality data, we need data that contains the know-how of skilled workers in manufacturing sites, but there is strong resistance from them over data collection."