Mobilinteu to develop humanoid on-device AI platform, test it with LG Electronics

AI chip company Mobilinteu (모빌린트) said on the 22nd that it is developing an on-device AI computing platform for humanoid robots. The platform is meant to run the various AI functions a humanoid needs inside the robot and connect them to actual control, from recognising the surroundings to judgement and movement. [ 1 , 2 , 3 ]

The work is part of the Ministry of Trade and Industry's "K-On-Device AI Semiconductor Technology Development Program". Mobilinteu is the lead R&D institution for the sub-project "Development of an On-Device AI Accelerator for Humanoids". Seoul Economic Daily said the specialised agency is the Korea Evaluation Institute of Industrial Technology (한국산업기술기획평가원), while ZDNet Korea said the ministry and that institute oversee the program. [ 1 , 2 , 3 ]

The outlets set out why a single AI is not enough. A humanoid must run vision-language models (VLM) and vision-language-action models (VLA), Seoul Economic Daily said, along with object recognition, movement, safety judgement, voice and sensor processing. Chosun Ilbo and ZDNet Korea added that, unlike a data centre, a humanoid has limited battery capacity and internal space, so power use and heat must be kept down and the time from sensor input to AI judgement to robot control must be short. [ 1 , 2 , 3 ]

Mobilinteu plans to develop a high-performance, low-power neural processing unit (NPU) and a small accelerator module, and to optimise chips and software together so that computing resources are shared efficiently when several AIs run at once. It will also build a software environment supporting the conversion, optimisation and execution of AI models. The company said this should reduce the burden on robot makers of connecting chips, AI models, hardware and software. ZDNet Korea added that functional safety and cybersecurity requirements will be reflected in development. [ 1 , 2 , 3 ]

The platform is to be validated in stages against LG Electronics' humanoid scenarios for everyday spaces and its system environment. According to Chosun Ilbo and ZDNet Korea, the checks will cover power use and heat on battery, stability over long operation, and the module's durability against vibration and shock. ZDNet Korea also listed control delay when several AIs run together. Mobilinteu plans to extend the technology to service robots, autonomous mobile robots (AMR), smart appliances and AI cameras. [ 1 , 2 , 3 ]

Park Ju-hyeon (박주현), head of research planning at Mobilinteu, said that in physical AI what matters is not only the computing performance of AI models but whether the AI runs stably inside a real product and whether its judgement quickly leads to the machine's movement. Chosun Ilbo and ZDNet Korea quoted him as saying the company will develop everything from the NPU to the module and software as one computing platform and verify real-world applicability in LG Electronics' humanoid environment. [ 1 , 2 , 3 ]