CarbonSix, a startup developing artificial intelligence (AI) for robots—so-called "physical AI"—unveiled its robot intelligence strategy on the 19th aimed at manufacturing sites. CarbonSix is distancing itself from a universal humanoid-centered approach and focusing on developing AI and hardware for robots that can be deployed immediately in manufacturing processes.
CarbonSix was founded in July last year by Chief Executive Moon Tae-yeon, who founded the vision AI startup SuaLab in 2019 and sold it to Cognex in the United States. Based on 10 years of experience applying AI to manufacturing sites, Moon chose "manufacturing sites" as the starting point for robot AI. Moon said, "Images and work data generated in manufacturing processes do not exist on the internet and are buried on the floor," and noted, "Because this is volatile information that disappears once a person completes a task, the approach itself has to be different."
Moon also pointed out the fundamental differences between large language models (LLMs) and robot AI. Moon said, "Language models like GPT grew on the vast data base of the internet, but the data needed for robot foundation models is not on the internet; it is hidden in human movements and on-site tasks," and explained, "It is structurally difficult to bring over a 'general play' like LLMs as-is in the robot domain."
This line of thinking leads to skepticism about the omnipotence of humanoids. Seo Hyeong-ju, CarbonSix chief technology officer (CTO), said, "The approach of solving every manufacturing problem with a single humanoid is far from reality," and added, "In manufacturing, each process has different environmental conditions, tools, precision, and speed requirements, making it difficult to secure both versatility and efficiency with a single platform."
Seo particularly pointed to the limits of a humanoid-centered strategy in terms of return on investment. Seo said, "In manufacturing, cycle time, expense, and reliability must all align for a technology to be adopted," and explained, "A general-purpose humanoid is more likely to increase cost and complexity than to meet on-site requirements."
The core of CarbonSix's physical AI strategy is its method of data collection. Moon defines the data accumulated but unused in manufacturing sites as a "digital history," saying, "The core of the business is how to secure the data buried in human motion and work processes, and what value to provide in return." Moon added that Korea, where manufacturing clusters such as semiconductors, batteries, and shipbuilding are concentrated, offers a favorable environment for commercializing physical AI.
As an outcome of this strategy, CarbonSix recently launched SigmaKit, a standard manufacturing robot product based on AI imitation learning. SigmaKit is a toolkit-style solution designed to be applied to manufacturing processes without AI expertise or complex system setup, implementing a structure in which the robot learns when a person demonstrates the task.
SigmaKit consists of ▲ AI specialized for manufacturing ▲ a robot gripper tailored for delicate tasks ▲ an intuitively operable teaching tool ▲ sensor modules. It is designed so that robots can perceive and make judgments even in manufacturing environments with frequent model changes and high irregularity, aiming to cover processes that were difficult to automate with conventional methods.
Starting with SigmaKit, CarbonSix plans to expand the scope of physical AI to manufacturing processes that have been difficult to automate. Moon said, "We have received sales inquiries and preorders since launch, and we are currently conducting proofs of concept (PoCs) with major domestic manufacturing corporations," and added, "Rather than dishwashing or household robots, we aim to focus on areas that can be commercialized fastest on actual manufacturing sites and build the most advanced robot AI model within manufacturing."