Arduino UNO Q: Bringing Edge AI to "Execution" – A Complete Standalone Analysis Guide Using unoq-vision-app

Background: The emergence of Arduino UNO Q and its advantages in edge AI

Until now, Arduino has been synonymous with simple and reliable real-time control, exemplified by its "blinking LED" functionality. However, with its mission to eliminate technological barriers, Arduino has now evolved into the phase of "Action" on "Thinking" AI. At the heart of this evolution is the Arduino UNO Q, born from a strong partnership with Qualcomm.

The most distinctive feature of the Arduino UNO Q is its ability to combine high-performance computing with real-time control. Conventional microcontroller boards lacked the processing power to perform inference on large-scale AI models, and it was common to send data to the cloud and wait for the analysis results. However, this configuration has major obstacles such as "communication delay (latency)," "running costs," and "the risk of data leakage."
The Arduino UNO Q solves these challenges with the power of hardware. This board has the power to enable industrial-grade vision inspection and cost-effective predictive maintenance. In addition, its integration with Arduino Cloud enables rapid deployment of IoT projects, while also supporting the execution of Large-Scale Language Models (LLMs) and Vision Language Models (VLMs) in a local environment.
Since LLM tends to evoke images of Qwen or Llama, which are around 10-30 bytes in size, something like the following would be better. It offers the extensibility to run lightweight LLM or VLM in a local environment.

Of particular note is the ability to replace areas that previously required expensive, dedicated AI hardware with the familiar Arduino open-source ecosystem. This gives engineers a powerful tool to instantly bring their ideas to life, from the early stages of development to quick troubleshooting in the field, without relying on "overpriced hardware."

Arduino UNO Q

Preparation: Components of the AI vision app "unoq-vision-app"

​unoq-vision-app was developed to maximize the performance of the Arduino UNO Q and make advanced vision analysis "easy." This application packages multiple advanced AI models and execution infrastructure (runtime), providing guidance for building complex inference pipelines in edge environments.
The main components of this app are the following three:

1. YOLO Model (yolov8n.onnx): This is an AI model for real-time and high-speed detection of specific objects, such as "people," within video footage. It uses the ONNX format and is optimized for efficient inference on edge hardware.
2. VLM Model (SmolVLM2): This is a multimodal model (Vision Language Model) that integrates and understands images and language. It is responsible for analyzing specific image regions (cropped images) detected by YOLO in detail and describing features such as "what kind of clothes are they wearing" and "what are they holding" in natural language.
3. Runtime (yzma framework): This is the foundation for actually controlling and executing the developed AI model. It manages everything from model loading and memory management to inference scheduling, enabling engineers to implement AI functions without having to worry about low-level processing.

1. A user-friendly vision AI app that runs on Arduino UNO Q.
https://github.com/robby-proj/robby-project/tree/main/unoq-vision-app

2. Key components

YOLO model
models/coco/yolov8n.onnx
AI model for real-time, high-speed object detection
VLM model
SmolVLM2 (.gguf)
AI models that integrate and understand/process images and language.
Runtime
VZMA framework
The platform for actually controlling and executing the developed AI model.

Convenience is also a priority when setting up the development environment. By utilizing the Docker container provided in the GitHub repository (robby-proj/unoq-vision-app), complex dependencies and the effort required to set up the environment can be minimized.
The directory structure is as follows:/home/arduino/models Place the trained AI model there,/home/arduino/ArduinoApps/unoq-vision-appA structure in which the application is run is recommended. Regarding model preparation, it is also recommended to run it directly on the board.wget oryzma A specific workflow has been established for downloading and installing using commands.

Practice 1: Implementation results and analysis points using a standalone configuration

This column will focus on the implementation of a fully independent (standalone) inference environment, which is the highlight of this article. The concept behind this configuration is to connect a camera and DISPLAY directly to the Arduino UNO Q, completing the "input, inference, and display" process without any external network involvement.

System Configuration and Procedure: The physical connection is very simple. Connect a USB camera, DISPLAY, KEY BOARD, and mouse to the Arduino UNO Q via a USB hub. POWER SUPPLIES is supplied via USB Type-C. Software is run from the onboard browser (Chromium). http://127.0.0.1:8080 Access the website and display the dashboard. When the user presses the "Manual trigger" button, the following analysis flow will start:

1. Video input: /dev/video0 Captures native camera footage.
2. Object detection: YOLO detects people and identifies their area.
3. Detailed analysis: The identified regions are passed to SmolVLM2 to verbalize the characteristics of clothing and accessories.
4. Display of results: The detected image and analyzed text (Chatbox) are displayed on the screen.

Practice 1: Configuration

Analysis of implementation results:
In the actual demonstration, the camera detected a person and analyzed the clothing the person was wearing. As a result, VLM provided a detailed description: "The main colors are black and white."
Key points of the analysis:
The advantages of this method over conventional cloud-based AI analysis lie in its low latency (real-time capability) and closed network environment.

Native processing: Because it does not use a network, it is completely unaffected by delays caused by insufficient communication bandwidth or packet loss.
Instant Deployment: Even outdoors or in secluded areas of a factory without Wi-Fi or LAN connectivity, advanced AI analysis can begin on the spot as long as there is POWER SUPPLIES.

View image @ Browser (Chromium) screen

In the unlikely event that the system becomes unstable, it can be quickly restarted or rebuilt using Docker commands (such as `docker compose restart`), making on-site debugging extremely efficient.

Practice 2: Realizing "remote control" monitoring via SSH connection

In "Practice 1," we explained a standalone configuration where peripherals are directly connected to the Arduino UNO Q. However, in actual operational settings and presentations, it's not always possible to stay close to the board. Therefore, we recommend a "remote control and monitoring" type of operation using SSH connectivity.
In this configuration, all computationally intensive processes such as AI inference and camera input are handled by the Arduino UNO Q, and the operator controls its operation from their Ubuntu PC via the network (Wi-Fi/LAN).

Specific implementation methods and procedures:

1. Network connection: Connect the Arduino UNO Q and your PC to the same Access point.
2. Login via SSH: From your PC's terminalssh arduino@[ArduinoのIPアドレス](For example: Execute ssh arduino@192.20.10.2) to log in remotely.
3. Controlling the container: After logging in, navigate to the application directory,docker compose restartBy executing commands such as these, you can freely build and restart the inference container from your local machine.
4. Remote Monitoring: Inference results can be viewed from a browser on a PC.http://[ArduinoのIPアドレス]:8080Simply Access the system. The processing results on the board (detected images and analyzed text) will be reflected on your screen in real time.

The flexibility offered by remote monitoring: The biggest advantage of this method is that it enables "smart presentations." There's no need to carry around heavy DISPLAY and KEY BOARD connected to an Arduino; you can smartly present the results of camera footage from a remote location on your laptop. Furthermore, since there's no need to build a complex AI runtime on the PC side, environment setup is extremely simple.

In terms of development, the ability to instantly check logs and make minor code modifications remotely when the system becomes unstable in the field dramatically improves debugging efficiency. This "remote control" approach, which overcomes physical constraints through the network, can be considered a practical operational style that maximizes the agility of edge AI.

Practice 2: Configuration

Summary: Three benefits supporting next-generation edge analytics

The analysis method combining Arduino UNO Q and unoq-vision-app brings three major advantages to future edge AI development.

1. Elimination of Noise and Cost: Because it does not rely on external networks, it completely shuts out the "noise" of unstable communication environments. In addition, you no longer need to worry about expensive cloud usage fees or dedicated server licensing costs, dramatically reducing development costs.
2. Faster Decision-Making: Analysis results are visualized immediately, enabling on-site decision-making. Browser-based dashboards help even non-specialized personnel intuitively understand AI inference results.
3. High Customizability and Scalability: Changes to the recognition target and prompt adjustments are instantly reflected simply by modifying the Python code on the board. Flexible system design is possible to meet your needs, including continuous monitoring with automatic polling every 5 seconds.

Without relying on expensive specialized equipment, the accuracy and speed of vision analysis can be dramatically improved by using an Arduino UNO Q and open-source software. We have moved beyond the era of being plagued by wireless communication problems and can now focus on the essential question of "what to solve at the edge."
Edge AI is no longer a technology of the distant future. With an Arduino UNO Q in hand, you can experience the world of "working AI" right now.

Related information and product introduction

Qualcomm Arduino UNO Q Information:
https://www.qualcomm.com/developer/hardware/arduino-uno-q
Arduino UNO Q product information:https://www.arduino.cc/product-uno-q/
unoq-vision-app (GitHub): https://github.com/robby-proj/robby-project/tree/main/unoq-vision-app
yzma Runtime:https://github.com/hybridgroup/yzma