Nota has released results showing performance improvements achieved by optimizing models for physical AI in the Qualcomm Edge AI environment. This case focuses on increasing speed and maintaining stability while running high-computation models at the device level, suggesting the potential for on-device AI applications in industrial settings such as robots.

Nota announced on the 29th that it has improved execution speed by optimizing the Vision, Language, and Behavior Integration Model (VLA) in the environment of Qualcomm's edge AI device, the Dragonwing IQ-9075. The model is 'SmolVLA 0.45B', which was implemented considering a physical AI environment that includes real-time motion generation.

VLA models are structured to simultaneously process image recognition, language understanding, and behavior generation.

It is generally known that it runs on a server-based system, and real-time execution on a single edge device is limited.

Nota explained that instead of shrinking the entire model, they applied optimization focusing on sections with significant performance improvements. In particular, the recognition and understanding stages were maintained, while the focus was placed on the stage of generating robot movements.

Key technologies applied include 'Real-time Inference Optimization' which reduces repetitive calculations and 'NPU-aware Graph Optimization' which adjusts the flow of calculations according to the hardware environment.

As a result, the processing time for the action generation phase decreased from 218ms to 31ms, a reduction of approximately 85.8%, and the total inference time was also shortened from 505ms to 310ms. The company stated that the job success rate remained largely unchanged at 85%, up from the previous 86%.

Nota unveiled the technology at the 'Embedded Vision Summit 2026' held in Santa Clara, USA.

At the event, a demonstration was conducted in which a model recognizes an item selected by a visitor and generates a robotic arm movement.

This demo is structured so that the AI assesses the situation and generates actions based on input, and is designed to verify the operation of physical AI in an edge environment.

Physical AI refers to technology that perceives the environment through sensors, interprets commands and situations, and translates them into actual actions. As its potential applications expand in fields such as robotics, manufacturing, and logistics, the importance of related models and execution technologies is also growing.

Nota announced plans to further advance AI model optimization technology in various edge and embedded environments in the future. In addition, it intends to expand the scope of on-device AI applications based on cooperation with the semiconductor and device ecosystem.