We’re going to explore how to perform local, low-latency AI classification on the new Arduino Uno Q.

Edge AI has traditionally been defined by rigid trade-offs between processing power and hardware constraints.

We will see if the new Arduino Uno Q is a good choice. Pairing a Linux Debian-capable Qualcomm Dragonwing™ QRB2210 microprocessor (MPU) with a real-time STM32U585 microcontroller (MCU), this board bridges high-level AI computing with low-level hardware control, all while maintaining seamless compatibility with existing Arduino shields.

To harness this dual architecture, we’ll be leveraging Arduino App Lab, the new development environment designed to streamline Edge AI deployments.

Throughout this session, we’ll pit three leading object detection architectures against each other:

• YOLO (You Only Look Once): The gold standard for full-frame object detection and precise bounding boxes, but heavily demanding on hardware resources.
• MobileNet SSD: A lightweight neural network optimized for embedded devices, delivering the sweet spot between structural accuracy and compute efficiency.
• FOMO (Faster Objects, More Objects): Edge Impulse’s ultra-lean architecture that trades bounding boxes for object centroids—enabling blazing-fast, real-time tracking on microcontrollers like the Arduino Nicla Vision or Portenta.

By the end of this talk, you will have a clear framework for selecting, optimizing, and deploying the exact AI model your project needs based on its memory footprint, latency, and power budget

By Jérémie Robert and Danica Woolgar