What can you do with a camera, an AI accelerator, and a handful of screws, nuts, and bolts? For us, it was an opportunity to explore real time edge AI vision and demonstrate what's possible with PHYTEC's phyFLEX®-i.MX 959 SOM.
Meet Screw Loose, a fastener identification and classification demo developed by the PHYTEC and Au-Zone engineering teams. Combining a camera with an AI perception model, the system identifies individual hardware components in real time, even under challenging conditions such as low lighting, reflective surfaces, and busy backgrounds.
To see it in action, simply place a handful of hardware from the junk drawer onto the tray. The system identifies and sorts each part into one of four categories: screws and bolts, washers, nuts, or miscellaneous items.
More than a fun demonstration, Screw Loose shows how the phyFLEX-i.MX 959 development kit can serve as a jumpstart for developing intelligent industrial vision applications.
In this setup, the work is divided between the processors. The phyFLEX-i.MX 959 handles the camera and display pipeline, while the Ara240 focuses on running the AI model. This allows the system to process camera images, identify individual fasteners, and display the results without putting all the processing demands on a single processor.
Working with an AI Partner
A camera can capture a pile of screws, nuts, and washers, but how does a system know what it's looking at? That's where the AI model comes in.
For Screw Loose, PHYTEC worked with Au-Zone and its EdgeFirst platform to develop an object detection model for the application. Unlike image classification, which identifies what an image contains, object detection can identify individual objects and locate them within an image. This is what allows Screw Loose to recognize and classify multiple fasteners in the same frame.
EdgeFirst brings together tools for AI assisted data labeling, model training, and optimization for embedded hardware, helping connect the different stages of AI development. The broader platform also supports more advanced perception applications, including sensor fusion across cameras, radar, and LiDAR.
Of course, developing the model is only part of the process. The team also needed to get the embedded platform ready and collect the images needed to train it.
Preparing the Data and Platform
An interesting part of the project was that hardware and software development could happen in parallel. While one group worked on the BSP, adding support for the EdgeFirst and Ara240 components through the meta-edgefirst and meta-kinara Yocto layers, another group could get started collecting images.
PHYTEC's prebuilt Embedded Linux vision BSP already included tools for working with the camera to display streaming, image capture and video capture. This meant the teams could begin collecting images using the actual demo hardware under different lighting conditions, with assorted fasteners and hands moving through the camera's view. We wanted the models to recognize real hardware in real conditions, not just perfectly staged images.
Once the images were collected and labeled, the teams could move on to training and deploying their respective models. Both approaches ultimately brought the AI processing onto the Ara240, running alongside the phyFLEX-i.MX 959.
So, how does the system actually recognize a handful of mixed-up fasteners?
The EdgeFirst version uses YOLOv8n, a compact object detection model with approximately 3 million parameters. Unlike image classification, which identifies what an image contains, object detection also locates individual objects. That's how Screw Loose can draw boxes around individual screws, nuts, and washers.
Recognizing smaller parts required some additional work. During training, the team used image tiling, which divides high-resolution images into smaller sections to preserve detail around objects. The model was trained on 2,237 images containing more than 35,000 labeled objects.
The teams also tested different input resolutions and found that 960 × 544 provided enough detail to recognize the fasteners without adding unnecessary processing. Before deployment, the model was optimized for the Ara240 using INT16 quantization, which converts its calculations to a 16-bit integer format.
Putting It to the Test
With the model trained and optimized, the next step was seeing how it performed in the live demo. A few reported performance figures:
For an object detection model, mAP@0.5 is a useful way to measure performance. The 0.5 refers to the required overlap between the predicted bounding box and the actual object location.
A project like this is a good example of how a development kit can help bridge the gap between evaluating a processor and building a working application.
Start with an evaluation kit. Get hands on with the hardware, explore the SoM's capabilities, and see how it handles the demands of your application before committing to a custom design.
Once you've tested the platform and know it meets your needs, the development kit becomes a starting point for designing your own system. The software work also carries forward with PHYTEC's Embedded Linux BSPs providing a foundation as your application moves from evaluation to a production design.
Screw Loose started with a handful of fasteners and an opportunity to explore edge AI vision. Along the way, it gave our team a chance to work through the hardware, software, and AI development involved in bringing an intelligent vision application to life.
See Screw Loose in action at Electronica 2026 on NXP's partner wall!