Mobilint 將開發實體 AI 運算平台

Mobilint is developing a computing platform for physical AI. (Photo: Mobilint)

Fabless semiconductor company Mobilint said Oct. 22 it will develop a computing platform for physical artificial intelligence (AI) applications, including humanoid robots.

Mobilint has been selected as the lead research organization for a project to develop on-device AI accelerators for humanoids. The project is part of the Ministry of Trade, Industry and Resources' K-On-Device AI Semiconductor Technology Development Program, overseen by the Korea Planning & Evaluation Institute of Industrial Technology (KEIT).

Mobilint will design a neural processing unit (NPU) to accelerate AI computing for humanoids. The company ultimately aims to develop an accelerator module that can be installed directly in robots. It will also develop a software environment for optimizing AI models.

The completed platform will undergo phased validation and improvement through integration with LG Electronics' humanoid system designed for everyday environments. Mobilint plans to expand the technology beyond humanoids into other physical AI applications, including service robots, autonomous mobile robots (AMRs), smart home appliances and AI cameras.

Humanoid robots use cameras and other sensors to perceive their surroundings, process information through computing devices and control walking and joint movements. Multiple AI models, including vision-language models (VLMs) and vision-language-action models (VLAs), are processed internally through on-device computing.

Mobilint aims to implement an on-device AI computing platform capable of supporting stable humanoid operation. Even minor control delays in robots and other physical AI systems can cause serious accidents, requiring ultra-low-latency control and high safety standards. On-device AI with low latency is therefore essential to avoid reliance on external communications or server connections.

Mobilint will jointly optimize its NPU and software to ensure that AI decisions are rapidly translated into physical robot movements. The platform will also allocate computing resources efficiently to support timely execution of critical tasks.