[Edaily Reporter Kim Hyun-ah ] Nota Inc., an AI model optimization company, has optimized artificial intelligence (AI) for robots to run on Qualcomm’s neural processing unit (NPU), tripling the operational speed of a real robotic arm.

#Nota Inc. announced on the 22nd that by optimizing its Vision-Language-Action (VLA) model for the NPU in Qualcomm’s industrial system-on-chip (SoC), the “Qualcomm Dragonwing IQ-9075,” it reduced the robotic arm’s task completion time from 36 seconds to 12 seconds.

This demonstration involved a robotic arm understanding a Saramin voice command, picking up a cube, and moving it to a mat on the opposite side.

The inference speed at which the AI determines the next action increased by up to seven times. The actual task success rate dropped by only 1 percentage point, from 93% before optimization to 92%. The company explained that it largely maintained the performance of the original model while increasing speed.

Notably, the AI inference was processed directly on the Qualcomm NPU connected to the robot, without relying on high-performance GPUs or external servers. Both the AI model and the runtime environment were optimized to enable the robot to handle everything internally—from camera image recognition to understanding voice commands and generating actions.

VLA is an AI model that understands video and language and uses this information to generate the robot’s actual actions. Because it requires significant computational power and memory, running it on a robot with limited power and computational resources necessitates optimizing not only the model itself but also the semiconductor and execution environment.

Nota Inc. implemented VLA model lightweighting, NPU computation graph optimization, a multi-NPU execution environment, and action generation acceleration. Subsequently, the system was connected to an actual robotic arm to comprehensively verify AI inference speed, task execution speed, accuracy, and task success rates.

AI inference speed also affects the robot’s responsiveness. This is because if decision-making is delayed, the robot may continue to move based on the previous situation even after the surrounding environment has changed. In particular, for AI models trained on a robot’s continuous movements, reducing inference latency is crucial for the stability of actual operations.

Building on this achievement, Nota Inc. plans to expand the optimization technologies it has accumulated in mobile and edge AI to the field of physical AI, including robots and humanoids.

Nota Inc. is a company specializing in AI model lightweighting and optimization, founded in 2015 by KAIST researchers as a student startup. Starting with technology to reduce typos on smartphone keyboards, the company expanded its business scope to model optimization technology that enables efficient AI operation even on small devices. It was listed on KOSDAQ in November 2025.

Its flagship technology is “NetsPresso,” an automated AI model compression and optimization platform. By reducing model size and computational load while minimizing performance degradation, it enables AI to run even in on-device environments with limited power and computing resources.

The company is collaborating with global semiconductor and technology firms, including SamsungElectronics, LGELECTRONICS, and Naver, as well as NVIDIA, Qualcomm, Arm, and Renesas.

Nota Inc. is also participating in the government’s “K-On-Device AI Semiconductor Technology Development Project” in the humanoid category, in collaboration with LGELECTRONICS and Mobilint. In this project, the company is responsible for optimizing a vision-language-action model for use with domestically produced NPUs and real-world robot environments.

Chae Myeong-su, CEO of Nota Inc., stated, “This is an example that demonstrates how we can go beyond simply improving the size or inference speed of AI models to actually enhance the responsiveness and task performance of real robots,” adding, “We will expand our optimization technology, which spans various AI models and hardware, to physical AI applications such as robots and humanoids.”