Amid late last week’s slow news cycle, someone reached out to ask if I’d like to interview Brooks on the occasion of his receiving the IEEE Founders Medal at the IEEE VIC Summit and Honors Ceremony in May. The pitch arrived along with a five-paragraph summary of his accomplishments. I didn’t need it, and if you’re reading this, you probably don’t need it either. Brooks has been a guiding voice in the robotics field for four decades, dating back to his long tenure on MIT’s faculty. Hell, he starred in an Errol Morris documentary about smart and interesting people.
I gladly agreed to speak to Brooks. Surprisingly, we’ve never spoken directly. Lora Kolodny interviewed him when he appeared at our first Robotics event back in 2017, and at last year’s show, Devin moderated a panel about HRI featuring Brooks and Clara Vu (with whom Brooks had worked when she was an intern at iRobot). We had a good chat earlier this week, which you can read here:
TC: What does your day-to-day look like as a CTO?
RB: Well, if you don’t count the last few days, I try to provoke people.
So, nothing new on that front.
I call what I do “making provocations.” I either write some code or do some design. It’s not what’s going to go into the product, but I I just try to say to engineers, “Why don’t we try to do this?” They’ll say, “We couldn’t do that. It’s impossible.” So, I do a crappy job of it. Then I think about it for a while, and they try to make it better and eventually do something good.
You’re one of the more pragmatic people in the field, so it’s interesting to set a starting point of “impossible.”
If you don’t try something out that’s hard to do, then other people will do it before you. So you have to try hard to do it. But you also have to be realistic about how hard things are and how easy things are. I’ve read some of the stuff you’ve written recently, and you talked about the humanoid race that’s happening. I think they’re totally unrealistic. They have no idea how hard the things are they’re saying they’re going to do in a year or with 20 people. One company I saw recently said they have 100 years of robotics experience between them. Last time I built humanoid robots — and I built more humanoid robots than any other person on the planet — I started with 1,000 years of robotics experience, and we just got a little, tiny way.
My suspicion is that a lot of companies have been around for a while, but the Tesla announcement forced them all out of the woodwork. I’m very much not a roboticist, but to me the hardest part is more the general-purpose part than the humanoid part.
Yeah, and the general-purpose manipulator in particular is incredibly hard. We’ve been working on robot hands since the mid-’60s, and they haven’t really progressed much since then. The most common robot hands are still parallel jaw grippers or suction cups for moving stuff in fulfillment centers. Nothing remotely like general-purpose manipulation.
As someone who — as you said — has built more humanoid robots than anyone, are you still bullish on the form factor?
The argument they’re using is exactly the argument I used in 1992, when I started doing it.
That we build structures for ourselves, so we should be robots that can operate in those structures.
Exactly. I’ve become less enamored with that over time. I saw it was, and that you actually make much more progress in the shorter term — and by shorter term, I mean 50 to 100 years — by building special-purpose machines.
Humanoid robots were part of what Rethink was working on.
Before that, the Cog and other humanoid robots, but none with legs. A few companies are doing legged robots, but they recommend you don’t stand near them. They operate in a very different way from how humans walk and they pump a lot of energy into the system when there’s a slight imbalance. If you’re nearby, you might get hit pretty hard by the legs.
Cobots were a big part of what Rethink was working on — the ability for humans and robots to work closely together.
Right, and we did it with arms, but not legs. We had very quick reactions to anything that wasn’t appropriate, and we just got the servers to suck the energy out of the system really quickly.
I think that’s something that most people don’t realize when they see the robots operating in places like Amazon warehouses: that’s mostly done in cages.
Precisely. You don’t walk around near those moving robots. They have safety systems that shut it down if a person goes onto the floor.
What happened with Rethink?
There were a couple of lessons. The proximate cause was the trade war with China. We were building robots in the U.S. and shipping them to China, and suddenly we got hit with retaliatory tariffs from China, whereas our competitors in Europe didn’t have them. There was a more fundamental problem, which was probably my fault, in that I let us build an expensive robot. Once it was an expensive robot, people wanted to treat it like a regular industrial robot, and they wanted repeatability. It was force-based, and we needed a much better sales and training organization to train the end users about how to use that robot different.
I still see the robots a lot in universities and other research centers.
Yeah, well, I say we were a total artistic success. We just weren’t a financial success. We changed robot shows forever. All robots at shows were in cages, and now they’re not. We had to fight like crazy in Chicago in 2013 to get approval for that. It’s just the way things are now.
Years ago, Melonee Wise told me Fetch was still building Willow Garage–style research robots because they’re a great recruitment tool, but you can’t really sustain a business on them.
At iRobot, we did the Create version of the Roomba that you could buy a 10-pack for $1,000. So that’s the standard people use for teaching robotics. But it was never a major part of the market for iRobot. It was a side thing. We thought that we had to help create more roboticists.
I’ve asked Colin [Angle] and Helen [Greiner] the same thing over years. It took iRobot something like a decade to land on the Roomba. What was the aha moment for you?
For me, it was when I went to Taipei in 1997. I was taken under the wing of a guy who was doing manufacturing in China, and I stuck around with him for a couple of weeks, getting the feel for how to do it. We were trying to build toys, and we knew we had to build them cheaply. We did a partnership with Hasbro, and we built a product called “My Real Baby” that was sold in stores. That was a humanoid. We learned how to build things cheap, and the thing with the Roomba was, “how cheap can it go?” We didn’t try to take a complex design and scale it. Instead we went and tried to ask people how much they would pay for it. When the answer was $200, that was the first piece of design. It had to retail for $200.
There’s been a big push in recent years to develop hardware-agnostic software solutions to help companies deploy robotics. Is Robust AI operating in that world?
Yes and no. We feel that a lot of these companies — because they’re dealing with existing solutions — are missing the boat where silicon is going. Silicon is putting tremendous amounts of processing right next to cameras, doing narrow floating point operations. Now you can do a lot of stuff, where as, I think existing mobile robots are still relying on lidar — a single, one-dimensional laser scanner — and you just don’t get as rich a view of the world. These cameras with these processors are incredibly cheap. You can get and run a lot of neural models and get a very rich 3D and labeled description of what’s in the world.
One of the things I appreciated in revisiting your predictions was its focus on eldercare robots. It’s been a thing in Japan for some time, but I don’t think enough people in the U.S. are discussing it. That makes the most sense to me as far as bringing more robots into the home.
Yeah, and I think it’s going to be a tremendous pull for anything that can help the elderly stay in their homes with independence and dignity. I don’t know exactly what it’s going to look like, but you can say the Roomba helps people stay in their homes, because they can clean the floors without having to [do] much work. There will be a tremendous pull just because of the incredible demographic shift across the world, with less young people to look after older people. Anything that can happen there will be helpful.
I know Helen would staunchly disagree with me here, but the robot vacuum is still really the only game in town when it comes to mainstream home robot adoption. It’s been so long since the first Roomba came out. Why has it been so difficult to repeat?
I wish I knew, because I would go and do it!
Obviously you don’t know what the finish line looks like, but you have as good a grasp as anyone on what makes it so difficult.
Houses are cluttered, and they have steps. Even a one-inch step makes a wheeled robot’s life very, very difficult. I occasionally I see new solutions that can get up one-inch steps, but then they can’t get up four-inch steps. It’s an artifact of how houses are built and modified over time. That’s an incredibly hard problem. Staircases are killer.
I’ve heard a lot of discussion about the ways in which smartphone innovations have led to robotic innovations. I think the same can be said for self-driving cars. The key difference is that, unlike smartphones, we don’t actually have self-driving cars.
You’ve noticed. But we actually have a lot better driver assist than we had. That sensory information is being used and processed, but it’s still letting the executive decisions occur with the human because the long tail of weird cases is just endless. Just trying to have enough data to pull from is unlikely. Whenever we’ve changed the way transportation works in the past, we’ve changed the infrastructure. There was this promise of a one-for-one substitution, and I think that’s held up what could have been a lot of change to infrastructure.
It’s also just understood that there will be deaths with human drivers. The minute an autonomous car kills someone, it sets things back by years.
There are 35,000 deaths per year in the U.S. from car accidents. I’ve seen some technologists argue that if we cut it down to 30,000 with self-driving cars, everyone will think that’s a big success. I say no. The acceptable number of deaths from self-driving cars is probably about 10 per year, and if it gets more than that, we won’t have them. I’m not saying 10 people dying per year is acceptable, but 35,000, that’s a big difference. With self-driving, it’s going to have to be a much lower number, and it’s not rational, but that’s how it’s gonna be.
What’s your take on this latest ChatGPT hype cycle?
I think people are overly optimistic. They’re mistaking performance for competence. You see a good performance in a human, you can say what they’re competent at. We’re pretty good at modeling people, but those same models don’t apply. You see a great performance from one of these systems, but it doesn’t tell you how it’s going to work in adjacent space all around that, or with different data.
Have you seen anything in robotics over the past few years that’s really excited you?
The outcome from deep learning and image labeling. People call it “perception,” but I don’t think it’s the same as human perception. But how well you can label images is a real change, and we’re using it here, because you can do great stuff. As long as you don’t think you’re getting human-like performance, but a different sort of perception. It’s labeling, and as long as you stick with that and realize the limitations, you can build really interesting capabilities into robots, which were pretty unimaginable not too long ago.
|