By Hamza Faisal
South Korea's race to build a homegrown "physical AI" stack has a new centerpiece: Mobilint, the AI semiconductor firm, announced on October 22 that it will lead a state-funded project to develop an on-device AI computing platform built specifically for humanoid robots.
The project, titled "Development of On-Device AI Accelerators for Humanoids," sits under the Ministry of Trade, Industry and Energy's K-On-Device AI Semiconductor Technology Development Program, with the Korea Evaluation Institute of Industrial Technology (KEIT) serving as managing agency.
One brain for the whole robot
The platform bundles a humanoid-specific neural processing unit (NPU), AI acceleration semiconductors, a compact accelerator module that mounts inside the robot, and the software needed to convert and optimize AI models for the hardware — all delivered as a single computing platform.
That integration matters because a humanoid does not run one AI model — it runs several at once. A vision-language model parses what the robot sees and hears; a vision-language-action model links that judgment to actual walking and joint movements. Mobilint's approach co-optimizes the silicon and the software so compute resources are allocated efficiently and every required task finishes on time.
"In physical AI, what matters is not only the computing performance of an AI model but whether the AI runs reliably inside an actual product and whether its judgments can feed quickly into the movement of the machine," said Park Joo-hyun, head of Mobilint's research planning division.
Tested inside LG's humanoids
The developed platform will be validated in stages against LG Electronics' humanoid systems built for everyday spaces. Testing will cover control latency with multiple AIs running simultaneously, power consumption and heat under battery power, stability during long operation, and module reliability against the vibrations and shocks of walking and joint movement.
Unlike data centers, humanoid robots run on batteries inside cramped bodies, so managing power and heat is as important as raw compute. Running everything on-device also cuts dependence on external servers or network links — a requirement for robots that work in real homes.
Mobilint says the bundle spares robot makers from separately connecting and optimizing chips, models, hardware and software every time they add a new AI capability.
From humanoids to smart homes
The technology is meant to travel well beyond humanoids. Mobilint plans to extend it across the physical AI sector — service robots, autonomous mobile robots, smart home appliances and AI cameras.
The move also fits South Korea's broader push for a domestic AI stack, with Korean firms racing to pair locally designed NPUs with robots, appliances and data centers — reducing reliance on foreign compute at every layer.
By Hamza Faisal
