San Francisco: Google published research this week detailing how its software enables robots to learn from one another. To demonstrate, the company’s scientists showed videos featuring robotic arms whirling inside its labs.
Google’s robotics group built those machines and wanted to sell them to manufacturers, warehouse operators and others. However, executives at Google parent Alphabet Inc. nixed the plan because it failed chief executive officer Larry Page’s “toothbrush test,” a requirement that the company only ship products used daily by billions of people, according to people familiar with the situation.
The verdict came around the end of 2015, just before the Google robotics unit moved to X, Alphabet’s research lab. Roboticists who worked on the project voiced frustration with Google’s caution, echoing sentiment at other divisions outside Google’s core Internet business, like its self-driving car unit, which display technical prowess but have yet to ship products.
“It was still a prototype, but it had a lot of advantages,” James Kuffner, chief technology officer at the Toyota Research Institute who previously led Google’s robotics unit, said about the arm. “The team worked really hard. If it had been entirely up to me I would have shipped it. But it was not.”
Google spokesman Jason Freidenfelds said there are no plans to sell the machines. “We’re using them to do basic research on how machine learning might help robots be a bit more coordinated—a promising field of research, but still very early days.” Courtney Hohne, a spokeswoman for Alphabet’s X, declined to comment.
Google built around 50 of the robotic arms capable of lifting about 10 pounds each, according to one person familiar with the project. They were designed by Meka Robotics, a start-up Google acquired in 2013, according to another person. People inside the robotics division pushed to release the machines at a relatively low price, undercutting other entrants in the small but growing market for light, collaborative robots. The people asked not to be identified discussing an internal project.
One of those companies, Universal Robots, was acquired by Teradyne Inc. for $285 million in 2015. Its models often cost more than $20,000.
The unreleased arm is another sign of trouble at Google’s rudderless robotics division. A slew of start-ups joined in 2013, thanks to an acquisition spree by former Android chief Andy Rubin. He left the following year, and Google ported the robotics teams into X for a reboot. Earlier this year, Google moved to sell the largest of these groups, Boston Dynamics, after tensions arose internally. Google hasn’t yet sold the unit.
Still, frustration with Google’s hesitancy to release unrefined products extends beyond its experimental divisions. Page often encourages Googlers to pursue “moonshots” that are risky but could have huge impact. This sometimes results in smaller, more practical products and services being shelved or remaining research projects for years.
Kuffner compared Google’s approach to that of Amazon.com Inc., which starting selling its voice-controlled Echo speaker two years ago. “When it was first released it was pretty miserable. It was very unreliable,” he said in a recent interview. “But they shipped it. That allowed them to get consumer feedback, to iterate quickly, to rapidly improve it. And now they’ve shipped lots of units.”
On Tuesday, Google released its own wireless speaker that operates similarly to the Echo.
When Google considers releasing a product, it has to protect its brand, which is among the most valuable in the world, Kuffner said.
“There’s risk associated with something that could be sub-par,” he added. “No executive is going to get it right all the time. It’s a hard balance and a hard line to walk because on the one hand you want to launch and ship early, but on the other hand you want to protect your brand.”
The robotics field needs more of Amazon’s approach, where products go out early and are iterated, Kuffner said.
Although Alphabet abandoned shipping the robotic arms, it has not put them to rest completely. Its latest papers, a collaboration between researchers at several Alphabet units, demonstrate a groundbreaking technique called “collective learning” applied to robotics. After the fleet of arms were trained on Google’s algorithms to open a certain type of door, for instance, they were able to replicate the action with doors they had never seen before.
“The skills learned by the robots are still relatively simple—pushing objects and opening doors,” the researchers wrote, “but by learning such skills more quickly and efficiently through collective learning, robots might in the future acquire richer behavioural repertoires that could eventually make it possible for them to assist us in our daily lives.” Bloomberg