Mobilint Leads Humanoid Robot AI Chip Project in Korea

South Korean chip designer Mobilint says it is leading a government-backed project to develop AI chips that let humanoid robots run their AI models on board, with testing planned in LG Electronics’ humanoid robot set-up. The company announced the role on 22 October as part of South Korea’s K-On-Device AI Semiconductor programme, which the industry ministry has budgeted at ₩800.2bn for 2026 to 2030.

The aim is a robot that can see, decide and move without sending data to a remote server. Mobilint says it will design a neural processing unit (NPU) for humanoids, a small accelerator module to fit inside the robot, and the software needed to run AI models on it. The company has not said how much of the national budget its project will receive.

In brief

Mobilint is the lead research organisation for a sub-project to develop on-device AI accelerators for humanoid robots, inside a ₩800.2bn state programme.
It plans to build a chip, a robot-mounted module and software as one platform, and test it in stages with LG Electronics’ humanoid system for homes and other everyday spaces.
Performance figures for Mobilint’s chips come from the company and have not been independently tested.

Mobilint will design the chip, the module and the software

According to the company’s statement, reported by Seoul Economic Daily and The Elec, Mobilint is the lead research and development organisation for a sub-project titled “Development of on-device AI accelerators for humanoids”. The programme is run by the industry ministry and managed by KEIT, the state agency that manages industrial R&D funding.

Mobilint says a single AI model is not enough for a robot that has to act on its own in spaces shared with people. A humanoid needs vision-language models to understand scenes and instructions, vision-language-action models to turn decisions into movement, and other systems for object recognition, navigation, safety checks, voice and sensors, all running at once.

The company plans to optimise its NPU and software together so that, when several models run at the same time, the most urgent tasks are handled first. It is also developing a high-performance, low-power NPU and a compact accelerator module to keep power use and heat under control inside the robot. Mobilint argues that delivering all of this as one platform would spare robot makers the work of connecting and tuning chips, models, hardware and software themselves.

“In physical AI, it is not only computational performance that matters, but also whether the AI’s decisions are rapidly translated into the machine’s movements,” Park Ju-hyun, Mobilint’s head of research planning, said, according to Edaily.

LG Electronics’ humanoid set-up will be the first test bed

The platform will be tested in stages with LG Electronics’ humanoid operating scenarios and system environment for everyday spaces, the reports say. Mobilint says the tests will check how quickly and reliably the system moves from sensor input to decision and control while several AI models run together. They will also cover power use and heat on battery power, stability over long periods, and how well the module copes with vibration and shocks from walking and joint movement.

Beyond humanoids, Mobilint says it plans to use the technology in service robots, autonomous mobile robots, smart home appliances and AI cameras.

The company already offers an edge AI chip, REGULUS. Mobilint describes it as a system-on-chip with 10 TOPS (trillion operations per second) of computing power at under 3 watts, aimed at uses such as drones and security systems. Those are company figures, and vendors measure TOPS in different ways, so they are not directly comparable with rivals’ numbers. None of the sources says whether the humanoid chip will build on REGULUS.

The ₩800bn programme spans cars, appliances, robots and defence

The ministry confirmed the programme’s budget on 2 June at ₩800.23bn in total, of which ₩511.11bn is state funding, running from 2026 to 2030. It supports custom AI chips, the modules they sit in and the software that runs on them across four sectors: cars, connected devices and home appliances, machinery and robots, and defence.

The robotics strand includes chips for collaborative robots in factories and food service, for farm robots, and for humanoids that help with household tasks. The ministry argued that on-device AI chips, unlike server chips such as GPUs, have no dominant supplier yet, and that South Korea’s mix of chip design start-ups and large manufacturers that would use the chips gives it an advantage.

On-board AI is a safety requirement for robots in homes

The case for this kind of chip is less about speed records and more about where AI runs. A robot that relies on a distant server for its decisions is exposed to network delays and dropouts, and in a home that can mean a robot arm that reacts late or not at all. Running models on the robot itself removes that dependency, but it creates a different problem: doing heavy AI work inside a battery-powered body without overheating or draining the battery.

That is why the testing plan matters more than the headline budget. Mobilint’s checks on control delay, heat, battery use and resistance to vibration are the practical questions that decide whether a humanoid can work safely around people for hours at a time. It is also why the claimed chip figures should be treated with caution until independent tests are published.

The project also shows how countries are approaching humanoids differently. China’s industry ministry is focused on volume, and expects humanoid output to pass 100,000 this year. Japan’s Mitsubishi Electric is pairing humanoids with machine tools on the factory floor. South Korea is betting on the silicon inside, in a market where chipmakers such as Synaptics are also building small NPUs for on-device AI.

What happens next

The first results to watch for are data from the LG Electronics tests and any independent benchmarks of Mobilint’s humanoid NPU. The value of Mobilint’s share of the programme, and a timetable for a chip that robot makers can buy, have not been disclosed.