Inspiration

As the world is pushed to the edge of environmental crisis, the notion of sorting your waste has become increasingly crucial. However, as you have probably noticed, many individuals simply don't care enough to sort their trash. Our team has long experienced this indifference firsthand, holding onto the hope that society would eventually recognize the importance of waste sorting. Yet, here we are, balancing on the precipice, surrounded by unsorted garbage and apathy. That's why we decided it was finally time for a change with the AI-powered EDGEbin—a step towards a more mindful and sustainable future.

What it does

The EDGEbin utilizes artificial intelligence and a clever mechanical system to do automatically identify and sort your trash into four categories: trash, recycling, electronic, and organic. All you have to do is toss your trash into the funnel.

How we built it

AI: Employing PyTorch and a custom dataset, we trained, validated, and tested our very own neural network for waste classification. For redundancy, we also made a second neural network where a pretrained model was finetuned to our classifications. These models, coupled with OpenCV, identified and classified waste via a webcam. Then, using the PySerial library, this information was sent to the Arduino.

Hardware: Our physical system, designed in SolidWorks, came to life through the use of laser cutting and 3D printing. Using C++, our Arduino was programmed to control a DC motor that moves an inner tube. Using the Stepper motor library and a motor driver, the stepper motor controlled the opening and closing of the release hatch. These electrical components were meticulously wired and seamlessly integrated with their physical counterparts.

Accomplishments that we're proud of

Creation of a custom convolutional neural network that was trained, validated, and tested in 26 hours

Challenges we ran into

We encountered several challenges throughout the 26 hour duration of UTRAhacks marathon. In terms of software, the lack of a CUDA enabled Laptop resulted in painstakingly slow training, validating, and testing for our AI models. Additionally, understanding the intricacies of convolutional neural network architecture and the mathematical operations for each layer was quite challenging.

On the hardware side, limitations in 3D printing forced us to downgrade most of our intended physical system to cardboard, introducing concerns about durability and accuracy. Adding on, one of our stepper motors unexpectedly stopped working, resulting in a last-minute substitution.