A form of AI known as deep learning has transformed computer vision in recent years and is rapidly spreading through manufacturing. An algorithm fed many thousands of example images can learn to identify dogs or cats in images, or to spot a particular person in security footage. It can also be trained to spot deviations from the norm in images of screws or circuit boards or screens.
Instrumental was founded by ex-Apple engineers to use machine learning to automate the monitoring of production lines. “Traditional vision systems are not well suited to discover and solve problems, because they’re ultimately rule-based,” says Anna-Katrina Shedletsky, the company’s CEO. “It’s a really good time to be talking about AI inspection, because there are these new pain points.”
Makers of traditional computer-vision systems, such as Cognex, increasingly tout machine learning in their products. Some startups offer off-the-shelf systems that promise to be easy to deploy and use.
At a Toyota manufacturing plant in Indiana that churns out hundreds of cars a day, quality control is crucial. Put the wrong widget into the wrong dashboard and production may grind to a halt. Workers normally scan a barcode on each part to double-check that it’s correct. But the plant is now preparing to deploy a robotic system that moves a camera around an object when an employee holds one out. It peers at the part from different angles and uses artificial intelligence to identify the component before (hopefully) giving the OK to install it.
The inspection robot, sold by Elementary Robotics, a startup based in Los Angeles, doesn’t look particularly futuristic, with a camera that moves horizontally and vertically along H-shaped bars. Place an object in front of the camera and it will inspect it from several perspectives. The robot shows how human workers and autonomous systems may work together on some manufacturing lines.
“Automation is classically a very brittle environment where you design these really complex, kind of kludged-together solutions,” says Carlo Cruz, a senior engineer at Toyota who is overseeing testing of the system. “I think the idea of having a human in the loop becomes fundamentally important in the future.”
Cruz says he would like to deploy the technology in other areas eventually, including inspection and quality control. “We see a lot of potential,” he says.
Elementary Robotics, founded in 2017, has been operating in stealth mode until now; it announced a $12.7 million Series A funding round Tuesday. Over Zoom, the company’s CEO, Arye Barnehama, shows off another version of the inspection system designed to examine ecommerce products for packaging damage and misapplied labels. He also demonstrates a version being used by another customer to examine circuit boards for flaws.
The systems cost “in the low teens” of thousands each, Barnehama says, and they need relatively few examples to be trained. A customer sends a few dozen images of an object to Elementary Robotics, which uses them to train an algorithm. As workers display new objects, the algorithm determines if they are as intended. A worker clicks a button to say whether the algorithm was correct, improving the process for the next round.
More Great WIRED Stories