Nota (486990.KQ), an AI model compression and optimization company, has run a robot AI model using only the neural processing unit (NPU) of Qualcomm Technologies' chip, without an external server, tripling the working speed of a robot arm.
Nota said on the 22nd that it cut a robot arm's task time to 12 seconds from 36 seconds after optimizing a vision-language-action model (VLA) and its runtime environment on the NPU of Qualcomm's industrial system-on-chip, the Qualcomm Dragonwing IQ-9075 processor. VLA inference speed improved by up to seven times. The task success rate was 92%, one percentage point below that of the original model.
The demonstration involved a robot arm that received a verbal command, picked up a cube and moved it to a mat on the opposite side. All AI inference — from camera image recognition to understanding the verbal command and executing the action — was processed on a Dragonwing IQ-9075-based edge AI board connected to the robot. That differs from the conventional approach of offloading AI computation to an external server equipped with a high-performance graphics processing unit (GPU) and then moving the robot based on the returned results. Instead, the commands were processed on-site using the NPU.
Inference speed in a robot is tied directly to actual task performance, because the surrounding environment and the robot's own position keep changing while the robot AI decides its next move. Robot foundation models such as VLAs, in particular, require heavy computation and memory, so running them inside a robot with limited computing resources and power requires optimizing not only the model but also the chip and the runtime environment.
Nota said it applied VLA compression, NPU computation graph optimization, a multi-NPU runtime and acceleration of action generation to achieve this. The work extends AI model optimization techniques previously used in mobile and edge devices into physical AI, where actual movement is required.
"This result shows that AI optimization technology does not stop at reducing model size or improving inference speed, but can translate into the responsiveness and task performance of an actual robot," Nota CEO Chae Myung-soo said. "We will expand our technology and business beyond mobile, edge and data centers into physical AI as a whole, including robots and humanoids."
Nota also demonstrated a robot arm using the technology at Korea Real-World AI Network (KRAIN) 2026, hosted by the Ministry of Science and ICT on the 11th of this month. Together with LG Electronics and Mobilint, Nota is taking part in a state-funded project in the humanoid field under the government's K-On-Device AI Semiconductor technology development program, working on VLA optimization based on domestically produced NPUs.
