Amazon Opens Its First Checkout-Free Store Concept

Amazon announced Amazon Go on December 5, 2016 via a public video, describing a 1,800 square-foot convenience store at 2131 7th Avenue in Seattle’s South Lake Union neighborhood (Amazon’s corporate campus) where shoppers could scan a QR code in the Amazon Go mobile app at an entry gate, take items from shelves, and walk out — receiving an automatic receipt and charge to their Amazon account minutes later, with no checkout lines or cashiers required. The store opened to Amazon employees in December 2016 for internal testing; the public launch was delayed over a year to January 22, 2018, reportedly due to technical difficulties tracking large crowds (more than 20 shoppers at once in early tests created tracking accuracy problems) and disambiguating who had taken items in dense crowd scenarios. The underlying technology — which Amazon branded “Just Walk Out” — combined ceiling-mounted depth cameras (hundreds per store), weight sensors embedded in store shelves, computer vision deep learning models, and the Amazon app entry ID to build a real-time “virtual cart” for each shopper. Contrary to widespread speculation, the system did not use RFID tags on individual products; all tracking was camera and sensor fusion based.

The core computer vision challenge was not product recognition in isolation — identifying a can of soup in a clear image was solved — but multi-target tracking under occlusion: tracking each of potentially dozens of shoppers simultaneously as they crossed paths, temporarily obscured each other from cameras, picked up items from crowded shelves in rapid succession, or replaced items on incorrect shelf slots. Amazon’s models needed to attribute each item removal or replacement action to the correct shopper even when multiple shoppers’ hands appeared near the same shelf simultaneously. Shelf weight sensors provided a complementary signal: weight changes on a shelf section correlated with items taken or returned, providing ground truth when camera attribution was uncertain. The integration of camera and sensor data required a fusion model that reconciled these two sources, handling cases where weight sensor noise or calibration drift disagreed with visual evidence. Amazon’s machine learning teams trained these models on millions of events in the company’s own test stores before the public launch.

Amazon Go expanded beyond the initial Seattle location throughout 2018: stores opened in Chicago and San Francisco, with the first New York City location opening 2019. Amazon Go Grocery (a larger ~10,000 sq ft grocery store format) debuted in Seattle in February 2020. The COVID-19 pandemic (March 2020 onward) increased interest in contactless shopping formats, but supply of the complex sensor infrastructure limited Amazon Go’s expansion rate. Amazon began licensing the “Just Walk Out” technology to other retailers and venues in 2020: airport convenience stores (Hudson News partnership), sports venues (concession stands at sports arenas), and third-party grocery chains. Whole Foods (acquired by Amazon June 16, 2017 for $13.7 billion) received pilot “Just Walk Out” checkout-free sections at select locations from 2021. The system demonstrated that computer vision and sensor fusion could replace a fundamental retail operation — the checkout process — though the economics of installing hundreds of cameras and sensors per store limited deployment to high-revenue, high-traffic environments where labor cost savings justified the installation cost.