Config (company)
Last reviewed
Jun 4, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 1,394 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 4, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 1,394 words
Add missing citations, update stale details, or suggest a clearer explanation.
Config (legally Config Intelligence) is a robotics data startup that builds the data layer for robot foundation models, positioning itself as a neutral data supplier rather than a robot maker. Headquartered in both Seoul and San Jose, the company was founded in January 2025 by chief executive Minjoon Seo and three co-founders, and it collects and processes large volumes of human motion data used to train general-purpose, two-handed (bimanual) robots. Config draws an explicit analogy to the contract chipmaker TSMC, which manufactures chips for Apple, Nvidia, and AMD without competing with any of them. The company says it wants to play the same role in robotics, supplying the underlying training data to every robot builder instead of fielding its own hardware. In May 2026 it raised an oversubscribed $27 million seed round led by Samsung Venture Investment, with strategic backing from the venture arms of Hyundai, LG, and SK, drawing widespread coverage as "the TSMC of robot data."
The central problem Config addresses is the scarcity of training data for embodied AI. Whereas large language models can be trained on text scraped from the open internet, robots that learn to manipulate the physical world need demonstrations of physical tasks, and that data largely does not exist at scale. Teaching a robot to fold a shirt or pack a box typically means physically recording a person performing the action, hour after hour, in controlled studios and in real environments. Config's thesis is that this data, not the robot chassis or the model architecture, is the binding constraint on progress in robot learning, and that a specialized supplier can produce it more cheaply and at higher quality than each robot company building the capability in-house.
As of its May 2026 funding announcement, Config said it had accumulated more than 100,000 hours of human motion data. The company described this as roughly 30 times the size of AgiBot World, which it characterized as the largest comparable open-source robot dataset at about 3,000 hours.
Config was founded in January 2025. Its chief executive, Minjoon Seo, is a machine learning researcher who was previously an associate professor at the Korea Advanced Institute of Science and Technology (KAIST), a research scientist at Meta, and chief scientist at the video-understanding startup Twelve Labs. Earlier in his career he was a research manager at Naver and earned his PhD at the University of Washington. The three co-founders came from Waymo, Google, and Naver, according to TechCrunch. Reported coverage also names Jack Bang as the company's chief operating officer.
The official company site describes Config as "a dedicated team of engineers and scientists" whose objective is "developing the data infrastructure necessary to realize general-purpose, bimanual robotics."
In May 2026, Config announced a $27 million seed round that it described as oversubscribed, raised at a valuation of more than $200 million. The round brought the company's total funding to $35 million. Samsung Venture Investment led the round. What drew the most attention was the breadth of strategic backing from South Korea's largest industrial conglomerates, several of which are themselves building robots and could become customers.
| Detail | Value |
|---|---|
| Seed round | $27 million (oversubscribed) |
| Valuation | More than $200 million |
| Total raised | $35 million |
| Lead investor | Samsung Venture Investment |
| Strategic investors | ZER01NE Ventures (Hyundai Motor), LG Technology Ventures, SKT America |
| Financial investors | Mirae Asset Ventures, Korea Development Bank, GS Futures, Kakao Ventures, Z Ventures |
| Angel investor | Pieter Abbeel (UC Berkeley professor; co-founder of Covariant) |
The presence of Pieter Abbeel, a prominent robotics and reinforcement-learning researcher, as an angel investor was widely noted. The participation of the venture units of Hyundai (through ZER01NE Ventures), LG (through LG Technology Ventures), and SK (through SKT America), alongside Samsung, meant that four of South Korea's biggest manufacturing groups backed the company at once.
Config's core activity is producing and refining human motion data for robot training. Its operations are concentrated in Seoul and in Hanoi, Vietnam, where a workforce of roughly 300 people records humans carrying out physical tasks, both in controlled studio settings and in field environments. The recorded demonstrations are intended to capture the fine-grained, two-handed manipulation that general-purpose robots struggle with, such as packing, folding, and rearranging objects.
A key part of the company's pitch is what happens after collection. Config says it transforms or retargets the raw human data before it is used for training so that it better matches the way a robot's body actually moves and interacts with the world, an approach the company has likened to translation between two languages. The idea is that raw human demonstrations are not directly usable by a robot with different kinematics, so the data must be converted into a form suited to the target robot's embodiment. Config frames data quality as something that can only be validated by "closing the loop," meaning training models on the data and then testing those models against the friction of the physical world rather than treating the dataset as a finished product on its own.
The company has said it plans to scale collection toward one million hours of data, a roughly tenfold increase from the 100,000 hours reported in May 2026.
Config sells access to its data and related technology to companies building robots and robotic systems. As of its 2026 seed announcement the company was already generating revenue, with reported customers including large manufacturers, system integrators, and companies in the agriculture and defense sectors. The official site emphasizes general-purpose bimanual manipulation and describes work on long-horizon robot memory for tasks such as playing piano, playing shell games, and rearranging objects, as well as on robots coexisting with humans in shared spaces.
Looking ahead, Config said it intends to launch a cloud-based Robot-as-a-Service product, described in coverage as a hosted inference layer that would let customers deploy robot capabilities through the cloud rather than only consuming raw datasets. The company set a target of reaching $10 million in annual recurring revenue by the end of 2027.
Config is part of a wave of companies betting that the bottleneck for embodied AI is data rather than hardware or model design. It operates adjacent to robot-brain and foundation-model companies such as Physical Intelligence, Generalist AI, and Skild AI, but its stated strategy is deliberately different: rather than building the model or the robot, it aims to be the upstream supplier of the data those companies and manufacturers depend on. That neutral, infrastructure-style positioning is the basis for the recurring "TSMC of robot data" framing in press coverage, which contrasts Config's supplier role with the vertically integrated approach of robot makers that build hardware, models, and data pipelines together.
The launch also fits a broader 2026 push in South Korea to build domestic "physical AI" data infrastructure, as the country's manufacturing giants invest in robotics and data facilities to keep pace with rivals in the United States and China.
robot foundation model - embodied AI - robot learning - physical intelligence - humanoid robot - robot manipulation - robot teleoperation