Vision Language Action Model Inference Speed Improved Up to 7-Fold, Robot AI Driven on Edge NPU Without GPU

Nota, an AI optimization specialist company, has announced that it optimized a robot AI model on Qualcomm's industrial AI processor, significantly improving actual robot task performance. Through this, the company demonstrated optimization technology spanning AI models, semiconductors, and execution environments in a physical AI environment.

On the 22nd, Nota announced that by optimizing a vision language action model (VLA) on the neural processing unit (NPU) of Qualcomm's industrial system-on-chip (SoC) called 'Snapdragon IQ-9075', it improved robotic arm task speed by 3-fold compared to the existing baseline.

Performance verification was conducted on a task where a robotic arm picks up a cube and moves it to a designated location based on voice commands. The experimental results showed that the AI model's inference speed improved by up to 7-fold, and actual task time was reduced from the existing 36 seconds to approximately 12 seconds. The task success rate was 92%, minimizing performance degradation compared to the original model's 93%.

AI Execution on Edge NPU Without GPU

A characteristic of this demonstration was that the AI model ran directly on the robot's embedded NPU without high-performance GPUs or external servers.

Vision language action models are AI models that integrate image recognition, language understanding, and action generation functions, and generally require high computational performance. Nota explained that it implemented this in an edge environment through model compression, NPU computation graph optimization, and multi-NPU runtime construction.

The company configured the system to process the entire workflow—from camera image recognition through language command interpretation, action generation, to robot motion execution—within a single system.

Expanding Physical AI Optimization Technology

Nota unveiled a robotic arm demonstration applying the technology at the 'KRAIN 2026 (Korea Real-World AI Network 2026)' event.

The company plans to expand the scope of applying AI optimization technology beyond mobile and edge devices to robotics and humanoid fields.

Meanwhile, Nota is participating in a national R&D project in the humanoid sector of the government's K-On-Device AI semiconductor technology development initiative alongside LG Electronics and Mobilint, and is also conducting research on robot AI optimization based on domestic NPUs.