Today, Amazon fulfillment center workers have to identify products at various points by manually scanning barcodes. As the e-commerce giant explains, the employee has to take the item, look for its barcode, and then scan it.
The company is not satisfied with the performance of this system, which is based on almost 50-year-old technology. “This process is repeated millions of times across a huge catalog of items of different shapes and sizes and is not easily automated,” explains Amazon.
In fact, there are currently no robots powerful enough to automate this task. Also, Amazon’s goal is now to move away from barcodes and adopt a new AI-based method of identifying items.
Amazon wants to replace the barcode with the MMID
To get rid of the barcode, Amazon researchers are working on a new method of product identification called Multimodal Identification, or MMID. Essentially, it involves identifying an item from an image and extracting information about its appearance and dimensions. And while the system isn’t perfect yet, Amazon researchers believe it may one day stop using barcodes in warehouses.
“This vision of using MMID throughout the execution process to accelerate and enable robotic automation is being achieved.”says Nontas Antonakos, one of the initiators of this project at Amazon. “And if that’s the case, it will be another step in our journey to deliver packages to customers faster and more accurately.”
Researchers developed a new approach using computer vision and multimodal identification to ensure packages are delivered more accurately and efficiently. #Robotics #corporate research
— Amazon Science (@AmazonScience) December 9, 2022
It’s not the data that’s missing
First, Amazon researchers photographed items moving on the conveyor belts. These images were then converted into vectors, sequences of numbers. Next, Amazon developed machine learning (AI) algorithms for identification.
In initial experiments, Amazon researchers achieved a match rate (i.e. correct identifications) of 75% to 80%. And today they would be 99%.
This high rate can be explained by the fact that the algorithms do not search the entire Amazon catalog for the item to be identified. An impossible task, according to the company.
“Each item comes from a specific bundle, and each bundle contains a few dozen products. So the algorithm only has to match an item with the contents of a single box.”states the e-commerce giant.
In any case, there is no lack of data to perfect this system, since Amazon only has to position a camera over its conveyor belts. This provides images of scrolling elements to train the AI.
potential and limits
For Amazon, this technology is promising. And it could pave the way for more automation in its warehouses. This artificial intelligence has already proven itself in Hamburg and Barcelona. On the conveyor belts, it detects errors: when an item does not match what is stated in the inventory.
But the e-commerce giant also admits its technology has limitations for now. This is tested on conveyor belts, for example, since both the lighting and the moving speed of the items are relatively constant. In other places, identification might be more complicated.
