One of the biggest obstacles to the spread of physical artificial intelligence is a shortage of data. Big Tech firms have been able to rapidly advance generative AI models because they had ample language data to train on, alongside computing infrastructure. Robots are a different story. There is far too little action data for robots to understand and learn from. Config Intelligence is a startup that has thrown down the gauntlet, saying it will secure that data itself and go on to develop a robot foundation model.
Seo Min-joon, chief executive of Config Intelligence, and Lee Ki-min, chief science officer, met with The Seoul Economic Daily at the company's office in Seoul's Gangnam district on the 14th. "There are many areas we can commercialize with robot data technology," they said. "Selling data is one of them, and we also want to focus on a business built on models trained with that data."
Config is a robot software startup co-founded in 2024 by Seo, Lee and others. Seo previously worked at NAVER and Meta and is currently a professor at the Kim Jaechul Graduate School of AI at the Korea Advanced Institute of Science and Technology. The founding team also includes Lee, a KAIST professor who formerly worked at Google; Son Hyung-mok, chief technology officer and a former Google Waymo engineer; and Bang Keuk-jae, chief operating officer and a finance and business specialist.
Although the company is not yet three years old, it has drawn concentrated interest from large Korean conglomerates. In April this year it secured Samsung Venture Investment, Hyundai Motor Group's ZER01NE, LG Group's LG Technology Ventures and SK Telecom as seed investors. It has also been named a key partner in the robot transformation, or RX, businesses of LG CNS and Samsung SDS.
What draws the conglomerates to Config is its sizable trove of robot data and its ability to convert that data and train models on it. "Even when robot hardware is advanced enough, not every task gets automated, and that is because of the limits of software," Seo said. "All of the co-founders agreed that 99% of what hardware can do depends on data." Lee added: "Data that people generate while working in physical spaces cannot be used for robot training as it is. We convert it into a form robots can learn from, deploy it on actual robots, and feed the performance results back into improving the data."
To that end, Config runs what it calls a "human factory" in Hanoi, Vietnam. With about 500 local staff, it collects roughly 20,000 hours of action data a month. The robot data gathered so far spans more than 400,000 tasks and more than 10 million episodes. Converted into hours, that comes to more than 150,000 hours.
"Early on we gathered data by repeating a small number of task types, but now we have greatly expanded the range of tasks," Lee said. "We perform things like lifting a tray, putting shoes into a box and picking bread off a moving conveyor belt, and then convert that so robots can learn from it."
The converted data is not tied to any single robot manufacturer. "If a client prefers robots from a particular maker, we have to secure data suited to that robot and train on it," Seo said. "We need to become familiar with a wide range of robots." That is also why robot hardware companies, not just general manufacturers, have proposed partnerships. Earlier this year, the company ran a proof of concept in the United States with American robot maker Dexmate.
Config is not stopping at securing data; it is also developing the foundation model that serves as the robot's brain. The flagship example is CFG-1, trained without outside data on material the company has secured and converted itself. CFG-1 was trained mainly on more than 100,000 hours of human task data. "Most robot foundation models operate by focusing only on the current task and situation, whereas our model has memory capability," Lee said. That means the robot can grasp what task preceded the one it is performing.
Config sees the robot foundation model business as still at an early stage but expects strong market potential ahead. "Most robot foundation models train on everything posted on the internet, but our model is trained primarily on data we secured ourselves," Seo said. "It is a mechanism that shows the better the model runs, the better the data is."
The company counts as a particular advantage its ability to respond quickly to problems, since it handles data training, robot foundation model development, field deployment, performance evaluation and improvement as a whole. "With robot data, you only know whether it is any good after you train a robot on it and check that the robot ultimately works well," Seo said. "When something goes wrong with a robot, it is hard to tell whether it is a data problem or a model problem, but at Config we can."
Behind the pair's focus on robot software is strong demand for automation at manufacturing sites. "Take beauty companies alone: their worry is that they cannot produce as much as they have been ordered because factory labor is short," Seo said. "For these companies, bringing in robots can translate directly into revenue."
Market potential also looks high given the state of robot hardware technology. "If the work a person does is 100, the current level of robot hardware is close to 1," Lee said. "Even doing just 1 is not small, and if hardware advances to 10, the variety of tasks that can be automated will widen." That, however, requires the development of robot hands and grippers capable of picking up smaller and more delicate objects.
Config's current clients include domestic chipmakers, automakers and systems integration firms, and inquiries from overseas Big Tech companies continue to come in. Cumulative investment stands at 50 billion won ($34 million). The company plans to spend the money on buying graphics processing units, hiring and research and development. "As data grows larger, training times also lengthen, and GPUs are in short supply relative to how fast our data is growing," Lee said. Seo stressed: "We are preparing to show new capabilities in models trained on our data. The goal is to demonstrate the world's best technology."