[Edaily Reporter Yun Junghoon ] #Nota Inc., a company specializing in AI model lightweighting and optimization solutions, has partnered with LGELECTRONICS and Mobilint to develop humanoid robots based on domestically produced on-device AI semiconductors. By lightweighting and optimizing large-scale robot AI models to run on domestically produced neural processing units (NPUs), the company will supply core software that enables fast and accurate autonomous control even in environments with limited hardware resources.

Nota Inc. announced on the 8th that it has been selected as the lead organization for the third sub-project in the humanoid category of the “K-On-Device AI Semiconductor Technology Development Project,” led by the Ministry of Trade, Industry and Energy and the Korea Institute for Advancement of Technology (KEIT).

This project is expected to serve as a turning point for Nota Inc. to fully expand its AI optimization capabilities—accumulated through mobile, edge, and data center applications—into the “Physical AI” domain. In particular, the fact that “optimization”—the efficient integration of AI models, NPUs, and actual robot systems—has been designated as an independent core task, with Nota Inc. leading it, demonstrates that AI optimization software has established itself as a core technology for the next-generation robotics industry.

The ultimate goal of this project is to develop an autonomous humanoid robot equipped with domestically produced on-device AI semiconductors, capable of independently perceiving its environment and performing human-level tasks even in complex everyday settings. LGELECTRONICS will oversee the entire project and lead the development of the humanoid system, while Mobilint will be responsible for the domestically produced NPU hardware, and Nota Inc. will be in charge of the core software that optimizes AI models for the NPU and the robot’s operating environment.

The “Humanoid AI Model Optimization” task, led by Nota Inc., focuses on streamlining the Vision-Language-Action (VLA) model to be integrated into the robot and ensuring seamless integration with Mobilint’s NPU.

Physical AI, such as that found in humanoids, must process visual information from cameras and other sources in real time to enable immediate situational assessment and corresponding actions. If the inference speed of the VLA model slows down, it cannot respond to changing on-site conditions in a timely manner, inevitably leading to a significant decline in safety and task accuracy. Therefore, optimization efforts to maximize inference speed while maintaining model performance within limited battery and computational resources are essential.

Nota Inc. is deploying a wide range of its proprietary advanced technologies, including token compression, quantization, pruning, knowledge distillation, and layer-by-layer optimization. Through these efforts, the company plans to dramatically reduce model size, computational load, and memory usage while minimizing performance degradation of the VLA model. The company will convert lightweight models—which achieve an optimal balance between accuracy, computational load, and latency—to fit the Mobilint NPU architecture, and verify their effectiveness by integrating them into a real-time inference environment along with workload-linking modules.

When such optimized VLA models are run on-device within the robot itself, the robot can make independent decisions and perform actions without relying on external servers (the cloud) or communication networks. By eliminating the need for data transmission, communication latency is drastically reduced, and the risk of sensitive internal information leaking to the outside is eliminated, significantly enhancing security. Reduced power consumption resulting from computational efficiency is also expected.

Chae Myeong-su, CEO of Nota Inc., stated, “‘This project is a significant example of applying the AI optimization technology Nota Inc. has accumulated to a domestically developed NPU-based humanoid robot,’” he said. “Based on our technical expertise in efficiently connecting AI models and hardware across various environments—including mobile, edge, and data centers—we will expand the application of this technology and business opportunities across the entire spectrum of physical AI that requires real-time on-device AI, starting with humanoid robots and extending to various other robots, automobiles, and industrial equipment.”