Nota AI Accelerates Robot AI 3× on Qualcomm NPU, Demonstrating Physical AI Optimization in Real-World Robotics

- Robot arm task time drops from 36 seconds to 12 seconds, vision-language-action model inference accelerated by up to 7× with less than 1 percentage point of performance loss
- All AI inference runs on the NPU of the Qualcomm Dragonwing IQ-9075 processor, enabling language-controlled robotic operation without relying on external GPU infrastructure
- Extends Nota AI’s optimization technology from mobile and edge AI into physical AI, demonstrating platform capabilities across diverse AI models and processors

‍

Nota AI, an AI model compression and optimization technology company, announced that it has increased the task execution speed of a robotic arm by 3× by optimizing both the robot AI model and its runtime environment for the neural processing unit (NPU) of Qualcomm Technologies’ industrial Qualcomm Dragonwing IQ-9075(QualcommTM IQ-9075) processor.

By running a vision-language-action (VLA) model directly on the robot’s NPU without relying on an external server, Nota AI translated faster AI inference into measurable improvements in real-world robotic performance. The results demonstrate the company’s end-to-end optimization capabilities spanning AI models, processor architectures, runtimes and physical robotic systems.

Nota AI evaluated the system on a robotic manipulation task in which a robot arm receives a natural-language instruction, picks up a cube and moves it to a mat on the opposite side. The optimized system accelerated VLA inference by up to 7× and reduced the time required to complete the physical task from 36 seconds to 12 seconds, a 3× improvement. The robot achieved a 92% task success rate, compared with 93% for the original model, representing less than a 1 percentage point performance difference.

A key aspect of the work was enabling a robot AI model that would typically run on high-performance GPU infrastructure or external servers to execute directly on an edge AI board powered by the Dragonwing IQ-9075 processor. Nota AI then extended the gains in inference speed beyond the model itself, translating them into faster physical task execution while maintaining a high task success rate.

VLA models are computationally intensive and require substantial memory, making it challenging to deploy them directly inside robots with limited power and compute resources. To address this, Nota AI optimized the full execution stack, including VLA-specific model compression, NPU computation graph optimization, a multi-NPU runtime environment and acceleration of action generation.

The company then connected the optimized system to a physical robot arm and evaluated the entire end-to-end pipeline from visual perception and natural-language understanding to action generation and execution. Performance was measured across inference speed, physical task completion time, model accuracy and task success rate.

Inference speed refers to the time between receiving visual and language inputs and generating the robot’s next action. Faster inference reduces the time a robot spends waiting before deciding and executing its next movement, allowing it to respond more quickly to changes in its surroundings.

This has a direct impact on both the accuracy and stability of robotic tasks. A robot and its environment continue to change while the model is computing its next action. If inference is delayed, the robot may act on information that no longer reflects the current environment, reducing responsiveness and increasing the likelihood of errors.

This timing is particularly important for robot foundation models trained on continuous action trajectories collected through methods such as teleoperation. If inference latency becomes too high during deployment, the timing of the robot’s physical movements can drift from the temporal patterns the model learned during training. In this sense, robot AI optimization is not simply about making large models smaller or faster. It is also about enabling learned behaviors to be executed at the right time and with sufficient stability in real-world environments.

“This result demonstrates that Nota AI’s optimization technology can go beyond reducing model size or improving inference speed and translate directly into greater responsiveness and task performance in real robotic systems,” said Myungsu Chae, CEO of Nota AI. “We will continue expanding and advancing our optimization layer across a wide range of AI models and hardware platforms, extending our technology and business opportunities beyond mobile, edge and data center environments into robotics, humanoids and the broader physical AI market.”

‍