Asia’s Manufacturing Powerhouse Drives Physical AI Advancement
Asia’s push toward physical AI is fueled by the same manufacturing prowess that has made the region a global industrial powerhouse. In South Korea, Japan, China, and Taiwan, the manufacturing sector remains a central pillar of economic growth. Unlike economies more heavily oriented toward services or software, these countries have long relied on large-scale production, export-oriented industries, and highly optimized supply chains. This structural foundation now determines how artificial intelligence is adopted and where investments flow.
Venture Capital Investment in Config
Config, a startup with dual bases in Seoul and San Jose, is making waves by building the data layer for robotic foundation models (RFM). Its approach has garnered attention from the venture capital arms of South Korea’s largest manufacturers. In a significant milestone, Samsung Venture Investment led an oversubscribed $27 million seed round, valuing Config at over $200 million. This brings Config’s total raised to $35 million.
Other strategic investors include ZER01NE Ventures, the venture capital arm of Hyundai Motor, LG Tech Ventures, and SKT America, the venture capital unit of a South Korean telecommunications giant. Notable angel investor Pieter Abbeel, co-founder of Covariant AI and a professor at UC Berkeley, also joined the round, alongside backers such as Mirae Asset Ventures, Korea Development Bank, GS Futures, Kakao Ventures, and Z Ventures.
Config’s Innovative Approach
Founded in January 2025 by CEO Minjoon Seo, previously a researcher at Meta and chief scientist at Twelve Labs, Config is steering away from building robots. Instead, the company focuses on providing high-quality data that robots need to learn and operate effectively. Seo emphasizes the importance of better data, stating that it will be crucial in making robots more useful.
Training large language models is often costly due to the computing power needed. However, the raw material—vast amounts of text from the Internet—is readily available. Conversely, teaching robots to move is more challenging. Each piece of training data must be physically collected, requiring robots, facilities, and personnel, making robotic AI development more expensive than software-only chatbots. As companies pursue more advanced bots, data collection and labeling costs can escalate rapidly.
Config’s Role as a Data Provider
Config envisions itself as the TSMC of robotics, providing data without competing directly with manufacturers. This approach is gaining traction as large manufacturers seek to create proprietary AI robots. The startup is already generating revenue, with customers including large manufacturers, systems integrators, and companies in the agriculture and defense sectors.
Peers in this space include Physical Intelligence, General AI, and Skild AI. Config records human tasks in controlled environments, operating in Seoul and Hanoi, where nearly 300 people are involved in data production. To date, it has amassed over 100,000 hours of human movement data, far surpassing the 3,000 hours of AgiBot World, the largest comparable open-source dataset.
Data Transformation and Future Plans
Unlike typical robotics teams that train AI models on human movement data, Config focuses on transforming data before training to better suit robotic movement. Seo likens this to language translation, where training a model on one type of data and expecting it to work seamlessly in another context is akin to teaching Korean using only English material. “The data needs to be converted, not the model. This conversion technology is Config’s key technical differentiator,” Seo explains.
The recent funding will prioritize scaling data operations in Vietnam and Seoul to one million hours of collected data, expanding its enterprise platform business to $10 million in ARR by the end of 2027, and introducing a cloud-based Robot-as-a-Service product. This service will allow companies to run Config’s base model without needing embedded hardware.
For further details, you can read the original article here.
“`
