Google’s PaLM-SayCan helps train robots for home and work

I’m standing in the kitchenette of a Google office in Mountain View, California, watching a robot at work. It’s looking at items on a counter: sparkling water, a bag of whole grain chips, an energy drink, a protein bar. After what seems like an eternity, he reaches out, grabs the chips, rolls a few feet, and drops the bag in front of a Google employee. “I’m done,” he says.

This snack delivery, which the bot performed at a recent press briefing, may not seem like a particularly amazing robotic feat, but it is an example of the progress Google has made in teaching bots how to be helpful, and not by programming them to perform a well-defined set of tasks, but by giving them a broader understanding of what humans might ask and how to respond. It’s a far more demanding AI challenge than a smartphone assistant like Google Assistant responding to a limited, carefully crafted set of commands.

The robot in question has a white tubular body, a gripping mechanism at the end of its single arm, and wheels. The fact that it has cameras where we have eyes gives it a certain anthropomorphism, but more importantly, it seems designed for practical functionality. It was created by Everyday robots, a unit of Google’s parent company, Alphabet. Google has collaborated with its bot-centric corporate sibling on the software side of the challenge of making bots useful. This research is still early and experimental; Along with tasks like finding and retrieving items, it also includes training robots to play ping pong and catch racquetballs.

A squadron of Everyday Robots bots hang out in Google’s robotics research lab. [Photo: Harry McCracken]

And now Google is sharing information about its latest milestone in robot software research, a new language model called PaLM-SayCan. (The “PaLM” stands for “Pathways Language Model.”) Developed in conjunction with Everyday Robots, this software provides the company’s robots with a broader understanding of the world that helps them respond to human requests such as “Bring me a snack and something to wash it with” and “I spilled my drink, can you help me? It requires understanding spoken or typed statements, unraveling the ultimate goal and breaking it down into steps, and accomplishing them using whatever skills a particular robot might have.

Google’s software breaks down human requests into components and decides which steps to take based on a particular robot’s capabilities. [Image courtesy of Google]

According to Google, its current PaLM-SayCan search is the first time robots have had access to a large-scale language model. Compared to previous software, according to the company, PaLM-SayCan makes robots 14% more efficient at planning tasks and 13% more efficient at carrying them out. Google also found a 26% improvement in bots’ ability to plan tasks involving eight or more steps, such as responding to “I left out a soda, an apple, and some water.” Can you throw them away and then bring me a sponge to wipe the table?

For you or me, choosing from several snacks on a counter is easy. For robots, this remains an achievement of note. [Photo: Courtesy of Google]

Not soon in a house near you

Although the Everyday Robots bots have done useful work such as sorting garbage in Google offices For a while now, the whole effort has always been to teach bots to learn to self-learn. In the demos we saw at the recent press conference, the robot performed its snack-fetching tasks so slowly and methodically that you could practically see the wheels spinning inside its head as it figured out the job step-by-step. . When it comes to searching for ping pong and racquetball, it’s not that Google sees a market for sports bots, but those activities require both speed and precision, making them good proxies for all. kinds of actions that robots will have to learn to handle.

Google’s emphasis on robotic ambition rather than getting something to market immediately contrasts with the strategy followed by Amazon, which already sells Astro, a $999 home robot, by invitation only. In its current form, Astro does little and isn’t much more than an Alexa gadget/security camera on wheels; When my colleague Jared Newman tried one at home, he struggled to find uses for it.

Google Research’s head of robotics, Vincent Vanhoucke, told me the company isn’t quite at the point yet where it’s trying to develop a robot for commercial release. “Google tries to be a company that focuses on providing access to information, helping people in their daily lives,” he says. “You can imagine a ton of overlap between Google’s overall mission and what we do in terms of more concrete goals. I think we’re really at the capability delivery level and trying to figure out what capabilities we can provide. It’s always a quest of ‘what are the things the robot can do? And can we expand our imagination on what’s possible?’

In other words, don’t assume you’ll be able to buy a Google robot soon, but stay tuned.

Comments are closed.