I think competition in the physical AI market will shift from who built the better model to who built more mass-production cases on actual manufacturing floors.
Met at the company's office in Seocho-gu, Seoul, on the 16th, Moon Tae-yeon, CEO of CarbonSix, said, "Until last year the focus was proof of concept (PoC), but this year is the first year that revenue is generated on real mass-production lines," adding, "We will make this year the first year of mass production for physical AI in manufacturing."
Unlike generative artificial intelligence (AI) corporations competing on model performance, in manufacturing the corporations that apply AI to real production lines, prove results, and accumulate mass-production experience will gain a competitive edge.
Moon is a co-founder and former vice president and chief strategy officer (CSO) of SUALAB, an AI machine vision corporation, and founded CarbonSix last year based on experience commercializing AI in manufacturing. CarbonSix is a physical AI startup developing SigmaKit, a software platform that equips industrial and collaborative robots with AI. Rather than developing new robots, it focuses on using AI to make the industrial and collaborative robots already deployed on manufacturing floors intelligent so they can perform "unstructured tasks" that were hard to automate because part positions or shapes were not uniform.
Conventional industrial robots are efficient for tasks that repeat fixed positions and motions, but even slight changes in the work environment required a person to manually modify the program. CarbonSix combined AI vision with robot control technology so existing robots can respond to diverse work environments, and because it does not develop new robots and can use the industrial and collaborative robots already installed in factories as they are, it lowered the adoption burden for manufacturing corporations.
A representative case is Taelim Industrial. After applying SigmaKit to Taelim Industrial's production line, CarbonSix is also discussing building a demo factory where local manufacturing corporations can directly review the technology and consider adoption.
There are a little over 20 contracted clients, most of them domestic manufacturing corporations. Until last year the focus was on verifying whether AI operates on actual production lines through proof of concept (PoC), but starting this year it has entered a phase where it is applied to real mass-production lines and generates revenue. The company is targeting 10 billion won in revenue this year and is expanding applications to automobiles as well as semiconductors, displays, electronics, bio, food, and chemicals.
Moon said, "Global physical AI corporations are fiercely competing on model performance, but there still aren't many cases that have led to large-scale mass production on real manufacturing floors," adding, "In the end, as mass-production experience accumulates, field data will build up, leading to a virtuous cycle that further improves AI performance." He added, "We aim to secure mass-production references as fast as possible and, based on that, to advance a manufacturing-specialized robot foundation model." The following is a Q&A with Moon.
-How many clients do you have? How far along are PoC and mass production?
"We currently have a little over 20 contracted clients, most of them domestic manufacturing corporations. Until last year the focus was on technology verification and PoC, but starting this year we are in a phase where real mass-production revenue is generated. Making this year the first year of mass production is the most important goal."
-Are there additional mass-production cases after Taelim Industrial?
"We are running projects across several manufacturing sectors, including automobiles. With Taelim Industrial, we are not stopping at building in a single factory but are also discussing establishing a demo factory where local manufacturing corporations can directly review the technology and consider adoption."
-Which industry is adopting the fastest right now?
"Automobiles are the fastest. In addition to semiconductors, displays, and electronics, any manufacturing with unstructured tasks performed by people—such as bio, food, and chemicals—is a target for application."
-What most differentiates you from existing physical AI corporations?
"We are not a company that makes new robots. We make the industrial and collaborative robots already installed on manufacturing floors intelligent with AI. We are compatible with a variety of robots, including FANUC, KUKA, Doosan Robotics, and Universal Robots. From a manufacturing corporation's perspective, adoption burden is much lower because they can use already proven hardware as is."
-What new features are slated for release in the second half?
"We will launch a dual-arm solution based on direct teaching. Previously, teleoperation—remotely controlling and collecting data—was common, but on manufacturing floors it is far more efficient for operators to move the robot directly and teach it. We expect usability to increase significantly in mass-production environments."
-Where will you focus the 62.3 billion won Series A investment?
"There are three main areas. First is hiring. We plan to expand from about 40 people now to around 100 by the end of next year. Second is scaling mass production. It is important to apply AI robots to as many manufacturing floors as possible to create actual mass-production cases and secure data. We also plan to support the necessary infrastructure so manufacturers can adopt new technology without burden. Third is building global partners."
-Why the emphasis on securing data?
"Physical AI is not a field where you can use open internet data. Ultimately, work data and diverse exception situations (edge cases) must accumulate in real mass-production environments. The data secured this way then improves model performance and creates a virtuous cycle that leads to more mass production."
-How far along is the global business?
"We are identifying partners in the United States, Japan, Taiwan, Southeast Asia, and Europe. The channel strategy is to have systems integrators and manufacturing automation corporations in each region use SigmaKit. This year we will secure partners and from next year we will expand the global business in earnest."
-What kind of company do you want to build long term?
"In the short term, the goal is to become the company that mass-produces the most intelligent robots in manufacturing. In the long term, we want to build a manufacturing-specialized robot foundation model based on data accumulated on manufacturing floors. Ultimately, the goal is to implement an intelligent manufacturing cell that performs assembly on its own when given only drawings and parts."
-What are your next funding and IPO plans?
"We are targeting a Series B round in the second half of next year. But securing sufficient mass-production revenue and customer references comes first. In the long term, we aim for an initial public offering (IPO). Our goal is to grow into a proven company that can raise capital continuously and build out the physical AI ecosystem in manufacturing."