Neural processing units (NPUs) are moving AI computation from distant data centers directly onto robots and vehicles, enabling faster decisions and lower power consumption. A Seoul-based company called Nota recently demonstrated a robot arm completing tasks three times faster by running advanced AI models on Qualcomm's edge processor instead of relying on cloud servers, signaling a major shift in how physical machines will operate.

What Makes Running AI on Robots So Challenging?

Robot arms and autonomous machines face a unique problem that smartphones and laptops don't: they need to make split-second decisions while moving through unpredictable environments. When a robot arm receives a verbal command to pick up an object, the surrounding environment keeps changing. If the AI system has to send camera data to a distant server, wait for a response, and then execute the command, the robot becomes slow and unreliable.

Traditional robot AI systems offload heavy computation to powerful graphics processing units (GPUs) running in cloud data centers. This approach works for tasks that don't require immediate responses, but for physical robots operating in real time, the delay becomes a serious limitation. Researchers have been searching for ways to compress complex AI models so they can run directly on the robot itself, using far less power and responding instantly to changing conditions.

How Did Nota Achieve This Speed Breakthrough?

Nota optimized a vision-language-action model (VLA), a type of AI that combines image recognition, language understanding, and decision-making into a single system. The company compressed this model and fine-tuned it to run on Qualcomm's Dragonwing IQ-9075 processor, which includes a built-in neural processing unit designed for edge computing.

In a demonstration, a robot arm received a verbal command, picked up a cube, and moved it to a mat on the opposite side of a table. All of the AI processing, from recognizing the camera image to understanding the spoken instruction to executing the physical action, happened on a small edge AI board connected directly to the robot. The results were striking: the robot completed the task in 12 seconds instead of 36 seconds, and the AI inference speed improved by up to seven times.

The task success rate was 92 percent, only one percentage point below the original model running on a powerful server, proving that compression didn't sacrifice accuracy for speed.

Steps to Optimize AI Models for Edge Robots

• Model Compression: Reduce the size of AI models through techniques like quantization and pruning so they fit on devices with limited memory and processing power.
• NPU Optimization: Restructure the AI computation graph to take advantage of the specific hardware capabilities of neural processors, ensuring efficient use of available computing resources.
• Runtime Acceleration: Deploy specialized software environments that coordinate multiple NPUs and accelerate the generation of robot actions, reducing latency between decision and execution.

Why Does This Matter Beyond Robots?

The implications extend far beyond robot arms. Nota CEO Chae Myung-soo explained the broader vision: "This result shows that AI optimization technology does not stop at reducing model size or improving inference speed, but can translate into the responsiveness and task performance of an actual robot," he stated. The company is now working with LG Electronics and Mobilint on a government-funded project to optimize AI models for humanoid robots using domestically produced NPUs.

"We will expand our technology and business beyond mobile, edge and data centers into physical AI as a whole, including robots and humanoids," said Chae Myung-soo, CEO at Nota.

Chae Myung-soo, CEO at Nota

The shift toward on-device AI processing is also appearing in automotive applications. OmniVision recently introduced the OAX7700, an ultra-low-power AI companion processor designed to add always-on security functions to parked vehicles. The chip handles initial detection locally through an integrated neural processing unit, allowing cameras to monitor a vehicle's surroundings without keeping the main processor active. Systems using this chip can consume up to 97 percent less power than existing parking surveillance solutions, while still detecting people, objects, and potential security threats.

What Does This Mean for the Future of AI Hardware?

The success of running complex AI models on edge processors suggests that the future of artificial intelligence won't be dominated by centralized cloud computing alone. Instead, computation will be distributed across multiple layers: powerful data centers for training and fine-tuning models, edge processors for real-time decision-making, and specialized NPUs embedded in robots, vehicles, and other physical devices.

This distributed approach offers several advantages. Robots and vehicles can respond instantly to their environment without network latency. Power consumption drops dramatically because complex computation happens on specialized, efficient hardware rather than general-purpose processors. And manufacturers gain flexibility to update AI capabilities through software updates without changing the underlying hardware.

The OAX7700 processor, for example, allows automotive manufacturers to develop their own AI algorithms for functions beyond security, with the ability to deploy updates over the air. This software-defined approach means that as AI models improve, vehicles can be upgraded without a trip to the dealership.

As robotics and autonomous systems become more prevalent in manufacturing, logistics, and consumer applications, the ability to run sophisticated AI models directly on these devices will become a competitive advantage. Companies that master AI model optimization for edge processors, like Nota, are positioning themselves at the center of a fundamental shift in how AI gets deployed in the physical world.