Mobilint builds an NPU, module and software stack for humanoids, to be tested in LG robots

Korean AI chip start-up Mobilint said in a release carried by HelloT and other Korean outlets on Oct 22 that it is developing an on-device computing platform for humanoid robots that bundles a low-power NPU, a compact accelerator module and software for converting, optimising and running models; the release has not been posted on Mobilint's own site.

The aim is to run several models at once inside the robot, such as a vision-language model, a vision-language-action policy, object detection, navigation, safety checks and speech, and pass their output to motion control without a server.

Validation will use LG Electronics' humanoid system environment and will measure the whole path from sensor input to control under multi-model load, plus power draw, heat, long-run stability on battery and module reliability under the vibration and shock of walking.

The work is a funded task of the Ministry of Trade, Industry and Resources' K-On-Device AI Semiconductor programme, managed by KEIT, where Mobilint leads the sub-project on on-device AI accelerators for humanoids; no budget, schedule or performance figures were given, and Mobilint plans to carry the platform to service robots, AMRs, appliances and AI cameras.

MEKAWIRE context: the same programme already has Daedong leading a KRW 43.1 billion unmanned farm-robot track with Mobilint and Rainbow Robotics as partners, as MEKAWIRE reported on Sept 9.