Baxter Meets World: Rod Brooks on What Rethink Robotics Is Learning
At San Francisco’s Fort Mason yesterday, former MIT roboticist Rodney Brooks gave the opening keynote address at Solid, a new O’Reilly Media conference focused on innovation at the confluence of software and hardware. True to that theme, Brooks talked about Rethink Robotics, the company he founded in 2008 to build humanoid manipulator robots, and the fact that Rethink’s flagship robot Baxter is getting more powerful over time—even without any design changes—thanks to improvements in the software that controls it.
Brooks, an Xconomist who is also a keynote speaker at Xconomy’s Boston 2034 event on June 10, showed two videos of Baxter at work, made 12 months apart at Rethink’s Boston headquarters. In the first, the two-armed robot completed a benchmark pick-and-place task—the kind of job a warehouse laborer packing a box with goods might do—in 3 minutes and 20 seconds. In the second video, shot recently, Baxter finished the same task in 1 minute and 33 seconds.
The only difference was that software engineers at Rethink had, well, rethought the way Baxter’s operating system models the physics of its moving arms. (It’s hard to describe, but in the second video Baxter seemed to move in a much more relaxed and human-like way.) Software “has changed the way we think about products in the physical world,” Brooks said. In a sense, he said, a complex, $25,000 machine like Baxter is now almost as upgradeable as an iPhone.
It’s probably a good thing that Baxter can be so easily reprogrammed. Rethink has been gathering a lot of data since Baxter’s debut in September 2012, and it turns out that the way customers are using the robot in real-world situations doesn’t always match up with the company’s original notions.
In an interview after his presentation, Brooks told me about Rethink’s business today and its transition from research and development (Baxter was nearly four years in the making) into sales mode. He said it’s been a bit of a challenge helping customers understand how Baxter can fit into their businesses, and how a robot whose capabilities are largely defined by software differs from the old-fashioned industrial robots people picture in their minds.
It’s certainly helps, Brooks says, that Rethink can upgrade Baxter just by releasing a new version of its operating system, called Intera. “But I think it leads to a question on how you market that, and whether customers are ready to understand that,” he says. “We are used to it with our smartphones—you get the upgraded software and you get something better, except when Apple threw out Google Maps. But I’m not sure, across different industries, that that is the case.”
One of Baxter’s main features, for example, is that it’s easy for non-experts to program. To show the robot how to carry out a task—say, picking up a cup on a conveyor belt and placing it into a box—a worker simply grabs one of its arms and guides it through the needed motions. Baxter also has several cameras that allow it to “see” the objects it’s handling, and over time, it can learn to recognize certain types of objects and apply actions consistently, such as picking up cups from the inside by spreading its pincers.
But some customers are uncomfortable with the idea that Baxter is learning; it means there’s something happening on the warehouse or factory floor that hasn’t been dictated and documented. “It’s a mindset,” Brooks says. “They want to know exactly what the machine is doing at all times, and if something is not right, they want to be involved in it. They don’t want it to just fix itself.”
In response, Rethink added connections to Baxter that allow customers to run the robot from programmable logic controllers (PLCs), which have been used since the 1960s to coordinate the actions of industrial machines. “We were ignoring that, and many customers said, ‘No, we’ve got to have that,’” Brooks says.
While the reaction was understandable, Brooks says it’s partly a relic from the first era of industrial robots, which were big, dangerous, and expensive. Baxter is designed to be the opposite of all of those things, and people haven’t quite adapted yet. He makes that point in this short video excerpt from his Solid presentation:
“I think I underestimated [the change in thinking required] and thought this new category of robots would be easy to understand,” Brooks says. “People know what an industrial robot is and they map Baxter to being an industrial robot, which it is not.”
This may start to change as businesses put more Baxters on the factory or warehouse floor, Brooks thinks. Right now, most customers own only one Baxter or a handful, as they experiment to see how the robot can save them money by making repetitive processes more efficient. If there were 20 Baxters in a shop, managers would be happy to let them learn and adjust on their own. “It’s being held back by people who are not seeing it as a commodity yet,” Brooks says.
Another misguided assumption that people make about Baxter, Brooks says, is that the robot should always mimic a human’s approach to a task. “The mistake they make is, ‘Oh, a person can do a task this way; therefore the robot should do the task exactly the same way,’” he says. “Well, no. For one thing it’s not as dexterous and its vision isn’t as good. It wasn’t meant to replace a person.”
But even if Baxter is a little slower than a human worker at some jobs, Brooks says it may still be a good investment, since it can work longer shifts and endure harsher environments.
A “light bulb” moment at one customer site occurred when factory managers realized they were focusing too much on Baxter’s cycle time for a certain task, Brooks says. “As we improved the software, it got closer, but it was still 10 or 20 percent slower,” he says. But what was far more important was that this particular factory floor was often uncomfortably hot for human workers. “The people didn’t like doing [the task], and the management didn’t want to have people doing it.”
The task itself, though, involved putting the finished products on a pile, not passing them directly to someone else. That buffer meant Baxter didn’t have to beat a human to be more efficient—it just had to show more fortitude.
“That was the realization: that if you put a robot into a line of production with other people, as long as there is an input buffer and an output buffer, it doesn’t need to have the same cycle time,” Brooks says. “They are rethinking a lot of the applications for Baxter, now that they have that constraint removed.”
The industries where Rethink is finding the most customers include warehousing and logistics, consumer goods, and plastics manufacturing. As it zeroed on those market segments and abandoned others, Rethink retrenched last December, laying off about 20 of its 90 employees.
Today, only half of Rethink’s engineering staff consists of mechanical engineers and other hardware specialists, Brooks says. The rest are all software engineers, and they’re working not just on making Baxter smarter but on opening up the Intera operating system to third parties, who could eventually sell their own control software.
In Brooks’s vision of the future, Baxter will evolve into a true platform, open to a variety of vendors selling software modules or hardware attachments for specific applications.
The more open that platform becomes, Brooks acknowledges, the more often customers and vendors will want to do things with Baxter that Rethink never anticipated. And that’s a natural evolution, he says.
“That’s what you always discover with customers when you have a new technology,” Brooks says. “Until you get out there with enough product and enough people, you are never going to learn those things. You can fiddle in the lab for years and not know how it’s really going to shake out.”