Nota puts robot VLA inference on Qualcomm’s edge NPU

Nota, a South Korean AI optimization company, announced on September 22 that it had run a vision-language-action robot model entirely on Qualcomm’s Dragonwing IQ-9075 NPU, without using a GPU or an external server. The company-reported demonstration centered on a robot arm responding to a spoken command to move a cube. In that test, Nota says the task time fell from 36 seconds in the original configuration to 12 seconds after optimization, while task success was reported at 92%, compared with 93% before optimization.

What Nota says changed

The announcement is an enterprise disclosure, not an independently verified benchmark. Nota says the change was not simply a hardware substitution, but an optimization effort that made the robot AI workload suitable for execution on the edge NPU. The company lists several methods: model compression, NPU graph optimization, use of multiple NPU units and action acceleration. Together, these changes were presented as enabling the VLA model to run locally on the Qualcomm processor rather than relying on a GPU or a server outside the robot system.

The practical distinction matters because a vision-language-action model must connect perception, language understanding and physical movement. In the demonstration described by Nota, the robot arm had to interpret a spoken cube-moving instruction and translate that into an action sequence. The reported improvement therefore concerns the end-to-end demonstration task, not only a single isolated neural-network operation. The reduction from 36 seconds to 12 seconds indicates a shorter cycle from command to completed movement in that specific setup.

Nota also reports inference up to seven times faster after optimization. That figure should be read separately from the threefold reduction in the demonstrated task duration. Inference speed refers to the model’s computation, while the robot-arm task includes additional stages such as command handling, perception and action execution. The two measurements point in the same direction, but they do not describe the same thing. The company’s claim is that optimization improved model execution enough to materially shorten the robot action sequence.

How the optimized setup works

The verified information does not provide model size, robot hardware details, cube arrangement, command wording beyond the cube-moving task, or the original system architecture. What can be said is that Nota moved the VLA workload to Qualcomm’s Dragonwing IQ-9075 NPU and avoided both GPU execution and external server processing. That implies the demonstration was framed around on-device inference: the computation needed for the robot AI task was handled locally by the edge accelerator rather than sent away for remote processing.

Model compression is one of the methods Nota says it used. In general terms, compression reduces the computational burden of a model so it can run within tighter processing and memory constraints. In this case, the relevant claim is not that compression alone produced the result, but that it formed part of a broader optimization pipeline. Because no additional technical parameters are provided, it is not possible to quantify how much of the 36-to-12-second reduction came from compression versus other changes.

NPU graph optimization is another stated component. This refers to adapting the model’s computational graph for the target neural processing hardware. For an edge NPU, that can be important because performance depends not only on the model’s theoretical size, but also on how operations are arranged and supported by the accelerator. Nota’s announcement attributes the result to this sort of hardware-aware optimization, including use of multiple NPU units. The company also cites action acceleration, which connects the inference improvements to the robot’s movement pipeline.

The result was demonstrated at KRAIN on September 11, according to the supplied facts. That setting makes the figures demonstration results rather than a broad evaluation across many environments. The company reports 92% task success after optimization, compared with 93% before optimization. The one-point difference, as reported, suggests that the demonstrated speed improvement did not come with a large drop in this success metric. However, without the number of trials, test conditions or failure definitions, the statistical meaning of that difference cannot be established.

What the numbers prove, and what they do not

The strongest supported conclusion is that Nota reports a successful on-device demonstration of a VLA robot model on Qualcomm’s Dragonwing IQ-9075 NPU, with the shown cube-moving task completed in one third of the previous time. The announcement also supports the narrower claim that the company measured faster inference, up to seven times, in its own optimized configuration. These are company-reported performance figures from a demonstration, not independent laboratory measurements and not a public benchmark suite with disclosed reproducibility conditions.

The figures do not prove that every VLA robot workload would see the same speedup on the same NPU. They also do not prove that the optimized system would maintain 92% success in more varied tasks, with different objects, commands or operating environments. The verified facts do not include comparisons with other edge processors, GPUs, server-based systems or rival optimization platforms. For that reason, the announcement can be analyzed as evidence of Nota’s specific implementation, but not as a general ranking of robot AI hardware or software.

The 36-second original configuration is also not fully described in the supplied material. It is therefore unclear whether the original setup used the same edge hardware without optimization, a different execution path, or another arrangement. That limits how far the before-and-after comparison can be interpreted. The central change described by Nota is the optimized execution of the model on the Qualcomm NPU without external compute. The practical outcome reported by the company is the faster completion of the spoken cube-moving command.

Practical implications

For robotics, the practical implication of such an approach is reduced dependence on remote computation for a class of AI-driven control tasks. If a robot can process a VLA model locally, the system design can avoid sending inference to an external server for that step. The verified facts do not provide measurements for power use, latency breakdown, thermal behavior or cost, so those factors cannot be assessed here. Still, the demonstration addresses a clear engineering question: whether a robot AI model can be made to fit and run on edge NPU resources.

Running on an edge NPU rather than a GPU may also matter for deployment constraints, but the announcement does not quantify those constraints. It does not state battery impact, enclosure requirements or total system bill of materials. The relevant supported point is narrower: Nota says its optimization methods allowed the model to run on Qualcomm’s Dragonwing IQ-9075 NPU and complete the demonstrated task faster while keeping task success nearly unchanged in the reported result.

Nota’s position in the story is also that of an optimization company rather than a robot maker as such. Founded at KAIST in 2015, the company markets the NetsPresso optimization platform and listed on KOSDAQ in November 2025. That background frames the announcement as a showcase for optimization technology applied to robot AI. The demonstration is therefore as much about adapting a demanding VLA workload to edge hardware as it is about the robot arm’s visible movement.

The announcement leaves important questions open, including how the optimization affects generalization, how the system performs across more complex tasks, and how repeatable the result is outside the demonstrated scenario. Even so, within the limits of the company-reported data, the case is technically significant: Nota says it moved a VLA robot model onto Qualcomm’s edge NPU, removed the need for a GPU or external server in that setup, and cut a specific spoken cube-moving task from 36 seconds to 12 seconds.