AI Trash Sorting Robot

Styrofoam is the question that started this. Most people do not know whether it is recyclable, so they bin it as garbage to be safe. That instinct scales badly: in Washington D.C. 79% of recyclables end up in the garbage, and roughly a quarter of what goes into the blue bin should never have been there, which costs money to sort back out. Humans are bad at this, so I built something that is not.

The robot is a box with a flat platform on a pivot. You set an item down, a webcam above it feeds the image to a computer vision model, the model calls it garbage or recycling, and a servo tips the platform toward the right bucket.

The classifier is an image model trained in Google's Teachable Machine across five classes (nothing, paper, plastic, can, trash) on a few hundred photos I shot myself, each item captured in several orientations against the same background the robot uses. Whole-image classification was the right call over object detection here: nothing needs to be located, only named, so there was no bounding-box labelling and it runs comfortably on a laptop. The model is exported to TensorFlow.js and served from a local HTTP server, and the browser pushes each verdict over USB serial to an Arduino driving the servo.

The body is cardboard, barbecue skewers and hot glue. The camera mast is deliberately staggered to one side so the webcam looks straight down at the centre of the platform rather than at its own supports, and the platform is folded along its length and skewered internally so it stays rigid while being flung back and forth.

In testing it put items in the correct bin 85% of the time, on only a couple hundred training images. That is already well clear of the 79% and 25% failure rates it was built to beat. A bigger dataset is the obvious way up, and moving the model off a laptop and onto a Raspberry Pi is what would make it something you could actually stand in a kitchen.