iStock.com/doug4537

Beer-slinging robot predicts whether you’ll give that brew a thumbs up—or down

As we pawn more and more jobs off on robots, there are a few you’d think we’d keep for ourselves, like beer taster. But brewers often need an automated way to ensure product quality, and a research team in Australia has developed a cheap method to help them. Their latest study assesses a freshly poured beer’s frothy top. Why focus on the bubbles? Because foam affects people’s enjoyment of beer and sparkling wine even more than taste and aroma do.

To gauge people’s reactions to beer foam, the researchers needed a consistent way to produce the foam, so they employed RoboBEER, a robot they’d previously built out of Lego pieces that pours beer from bottles into a glass. They showed videos of these robo-pours to people and asked them several questions about how they liked the height and stability of the foam, as well as the beer’s clarity and overall perceived quality. The goal was to be able to show people a video and predict these ratings without a long—or any—questionnaire, and without having to serve any actual beer, which slows the evaluation process even more.

To make predictions, the researchers used two types of data. First, as RoboBEER poured, it measured 15 beer attributes including bubble size, beer color, gas release, and foam height and stability. Second, people’s faces were videotaped as they watched the beer videos. Artificial intelligence (AI) analyzed the videos to measure biometric factors such as pupil dilation, heart rate, and emotional expression. For each viewer, the researchers fed 28 pieces of RoboBEER and biometric data into a neural network—another AI algorithm—to see whether the data correlated with the person’s conscious ratings.

The neural network could predict whether someone liked a beer’s foam height with about 80% accuracy, the team reports in Food Control. In unpublished work, the team has found that an AI with just the RoboBEER data can predict a beer’s likability—as rated in sipping sessions by consumers or even connoisseurs—with about 90% accuracy. And it doesn’t need a designated driver.