Asia’s Manufacturing Powerhouse Accelerates Advances in Physical AI

Asia’s leading role in global manufacturing continues to fuel groundbreaking progress in physical artificial intelligence. Nations like china, Japan, South Korea, and Taiwan maintain thier economic strength through expansive production networks and export-driven industries supported by highly optimized supply chains. This solid industrial foundation is instrumental in shaping how the region integrates AI technologies and channels investments into innovative sectors.

Key Investments from Asia’s Industrial Giants

A prime illustration of this momentum is Config, a startup with bases in Seoul and San Jose that focuses on building data infrastructure tailored for robotic foundation models (rfms). Recently, Config closed an oversubscribed $27 million seed round led by Samsung venture Investment, pushing it’s valuation beyond $200 million. Other strategic backers include Hyundai Motor’s ZER01NE Ventures, LG Tech Ventures, SKT America-a major South Korean telecom venture capital firm-and notable angel investor Pieter Abbeel alongside financial institutions such as Mirae Asset Ventures and Korea Growth Bank.

Expert Founders Driving Robotics Data Innovation

Founded at the beginning of 2025 by CEO Minjoon seo-formerly a Meta researcher and chief scientist at Twelve Labs-config was co-established with specialists from Waymo, Google, and Naver. Rather of manufacturing robots themselves, their core mission revolves around delivering premium-quality data essential for training robots to execute complex tasks efficiently. they emphasize that elevating data quality will be critical to expanding robot functionality across diverse industries.

The Complexities Behind training Robotics AI Models

Unlike large language models trained on vast amounts of freely available internet text-which are relatively inexpensive to compile-robotics AI requires physically gathered training datasets involving real robots operating within controlled labs or actual environments.Seo highlights that this makes robotics AI development considerably more resource-heavy compared to software-only chatbots due to expenses tied to hardware usage and human oversight during data collection phases.

“Every piece of training details must be physically captured using robots inside specialized facilities managed by skilled operators,” Seo explains. “This complexity rapidly escalates costs as companies pursue increasingly sophisticated robotic capabilities.”

A Semiconductor-Inspired Data-Centric Strategy

Config envisions itself playing a role similar to TSMC-the Taiwanese semiconductor powerhouse producing chips for Apple and Nvidia without competing against them-in the robotics ecosystem. By concentrating solely on providing foundational datasets rather than developing proprietary robot hardware or software that might compete with manufacturers’ own products, Config aims to become an indispensable partner enabling various companies’ ambitions in robot AI development.

Diverse Industry Applications Supported by Robust Datasets

The company already generates revenue through contracts spanning major manufacturers, system integrators as well as agriculture technology firms and defense contractors according to COO jack Bang. Competitors within this emerging field include Physical Intelligence, Generalist AI, and Skild AI-all striving toward advancing physical intelligence via enhanced datasets.

An extensive Human Motion Dataset Powers Robotic Learning

Operating out of Seoul and Hanoi with nearly 300 employees dedicated exclusively to data production efforts, Config has compiled over 100,000 hours of human motion recordings captured both inside studios designed specifically for precise movement tracking and also field environments simulating real operational conditions.This dataset size exceeds AgiBot World-the largest open-source equivalent dataset containing roughly 3,000 hours-by more than thirtyfold.

A Transformative Approach: Refining data Before Robot Training

The company distinguishes itself further by not only collecting raw motion capture but also converting these datasets into formats optimized specifically for robotic movement patterns before any model training begins.Seo compares this process to language translation: expecting an unadapted model trained solely on one type of input (human motions) to perform flawlessly when applied directly onto another domain (robotic actions) would be akin to trying learn Mandarin using only Spanish textbooks without translation aids.

“Our key innovation lies in transforming the data rather than altering the model itself,” Says Seo.This conversion technology forms Config’s primary competitive edge.”

Aiming high: Growth Plans Backed by Recent Capital Injection

• Expanding operations: Increasing workforce presence across Vietnam and Seoul targeting one million hours of motion data collection within two years;
• SaaS platform scaling: Pursuing $10 million annual recurring revenue (ARR) from enterprise clients by late 2027;
• Pioneering Robot-as-a-Service: Introducing cloud-based solutions enabling customers remote access to Config’s foundational models without requiring onboard computing resources embedded directly into robots.