Nota, an AI model slimming and optimization corporations, will team up with LG Electronics and Mobilint to develop a humanoid (human-shaped robot) based on a domestic on-device (built-in) AI Semiconductor. The goal is to optimize large-scale robot AI models for a domestic Neural Processing Unit (NPU) so they run fast and efficiently with limited compute resources.
Nota said on the 8th it was selected as the lead organization for subtask 3 in the humanoid field of the "K-on-device AI Semiconductor technology development project" promoted by the Ministry of Trade, Industry and Energy and the Korea Planning&Evaluation Institute of Industrial Technology (KEIT). This project is a case where Nota will fully apply its AI optimization technology, accumulated in mobile, edge, and data centers, to physical AI. Optimization to efficiently consolidation AI models, NPUs, and robots has been configured as an independent subtask, and Nota will lead it.
The project aims to develop an autonomous humanoid that recognizes its surroundings in complex everyday spaces and performs human-level tasks by using a domestic on-device AI Semiconductor. LG Electronics will oversee the entire project and handle humanoid system development. Mobilint will take charge of the domestic NPU, while Nota will each handle optimization software that runs AI models efficiently on the NPU and in real robot environments.
Nota will lead the "humanoid AI model optimization" task. It will handle the core software technology that slims down the vision-language-action model (VLA) installed on the humanoid and optimizes it for the Mobilint NPU so it runs fast and efficiently inside the robot.
Physical AI must process video information coming in through cameras in real time, recognize and judge the surrounding environment, and act immediately. If VLA inference slows, it may fail to reflect changed situations in time, reducing the accuracy and stability of actions. Therefore, optimization that increases inference speed while maintaining model performance inside robots with limited power and compute resources is essential.
Nota plans to apply token compression, quantization, pruning, knowledge distillation, and per-layer optimization to minimize VLA performance degradation while reducing model size, computation, and memory usage. With optimization and a built execution environment enabling the VLA to run inside the robot, the humanoid will be able to recognize and judge its surroundings and act without relying on external servers.
Chae Myeong-su, Nota's CEO, said, "Based on our technology that has efficiently consolidation AI models and hardware across various environments such as mobile, edge, and data centers, starting with humanoids, we will expand technology deployment and business opportunities across physical AI, including various robots, automobiles, and industrial equipment that require real-time On-device AI."