Autonomous System for Sorting Objects at the Edge

Abstract

This paper describes an implementation of an end-to-end machine learning-based system for sorting objects. We explore the specific application of sorting recyclables at the edge. This stands in contrast from existing waste processing systems that aim to classify and sort a large variety of items coming from multiple waste streams. Moving the sorting process closer to the point of waste generation reduces the risk of contaminating recyclables and enables local data collection to train the classification system on specific waste sources. This eases the classification task which enables the use of cheap, low-power electronics. Our system uses a low-power microcontroller (MCU) with an on-board camera module and a convolutional neural network (CNN) accelerator for classifying items. The MCU also controls a set of mechanical arms to physically sort objects that move along a conveyor belt. We outline the specifications of the physical system in addition to the development process of our machine learning model.