Nota triples robot arm speed via Qualcomm NPU optimization

Recap

Source: 헬로디디, report of Sept. 22, 2026

According to the report, AI model optimization firm Nota announced on September 22 that it has successfully optimized Vision-Language-Action (VLA) models for the NPU of Qualcomm's industrial System-on-Chip (SoC), the Dragonwing IQ-9075. This optimization allows the AI to run directly on the robot's internal hardware without relying on external servers or high-performance GPUs. The company reports that this approach resulted in a threefold increase in the actual task speed of a robot arm, such as moving objects, while inference speeds for generating the next action based on camera input and language commands improved by up to seven times.

Nota explains that the optimization process covered the entire pipeline, from model lightweighting and NPU computational structure tuning to building an execution environment that utilizes multiple NPUs and accelerating robot action generation. Despite these significant speed improvements, the company claims the task success rate remained stable at 92%, comparable to the original, unoptimized model. The firm demonstrated these capabilities through a live robot arm demo at the 'Korea Real-World AI Network (KRAIN) 2026' event.

Chae Myung-soo, CEO of Nota, stated that this case demonstrates how AI optimization can extend beyond mere model size reduction to directly enhance a robot's responsiveness and operational performance. The company plans to expand this technology to broader physical AI fields, including smart robots and humanoids. Nota is also participating in the government's K-On-Device AI semiconductor technology development project, specifically focusing on the humanoid sector to develop optimization techniques tailored for domestic NPUs.

Context

Vision-Language-Action (VLA) models are a critical component in modern robotics, as they bridge the gap between visual perception and linguistic commands to generate physical actions. Traditionally, these models are computationally heavy, often requiring cloud-based inference or powerful server-grade GPUs due to their large parameter counts and memory demands. This creates latency and dependency on network connectivity, which can be problematic for real-time robotic tasks. By moving this processing to the edge using specialized NPUs, companies like Nota are addressing the latency and autonomy challenges inherent in current robotic systems. Qualcomm's Dragonwing series is designed specifically for industrial and edge AI applications, making it a suitable target for such optimizations.

Robot's take

This development is significant because it addresses one of the primary bottlenecks in deploying advanced AI on physical robots: latency. By achieving a 7x speedup in inference and a 3x speedup in task execution while maintaining a high success rate, Nota suggests that complex, multimodal AI can be made practical for real-time control. However, the 92% success rate, while good, still leaves room for error in high-precision or safety-critical applications. It remains to be seen how well this optimization generalizes to different robot morphologies and more complex, unstructured environments. The move toward domestic NPU optimization in Korea also signals a strategic push for technological sovereignty in the robotics supply chain.