Industrial technology concept. Factory automation. Smart factory. INDUSTRY 4.0
getty
For a couple of years now, as AI’s uncanny capabilities have exploded, people on both sides of the debate (proponents versus detractors) have been waiting for humanoid robot dexterity to evolve, in order to answer that all-important question: if AI can think like us and see like us, what will happen when it can actually move like us?
Now, an open-source design from Fourier is pushing the envelope in this direction. We’re seeing robots that can shuffle, run and tackle an inclined slope.
There are lots of different takes on this, understandably. People don’t even agree on how popular humanoid robots are, let alone what the eventual outcomes are going to be. We’re still trying to read the tea leaves.
“Humanoid robots sound like something out of a movie, but they’re slowly becoming part of real-world industry,” writes patent attorney Bao Tran at Patent PC. “Or are they? While news headlines might make it seem like humanoid robots are taking over factory floors, the truth is more complicated.”
In support, Tran lays out some statistics: that 85% of all robots currently built are non-humanoid, for one, and that humanoid robots account for less than 2% of all commercial manufacturing sales.
But the consensus is that eventually, humanoid robots are coming.
Imagining a Robotic Future
There’s more from a panel at the Imagination in Action Stanford event in September, where panelists laid out use cases, brainstormed outcomes and talked about the process of evolving humanoid robots.
A Case Study
“Imagine a car manufacturing plant,” said Samir Menon, CEO of Dexterity. “A car comes along, the robot can weld it, drill holes, paint it. The car manufacturing plant manager won’t let one screw be out of place, and because it’s perfectly controlled, you can predict what is going to happen, which greatly simplifies the challenge.”
However, he pointed out, complexity is its own challenge.
“With robotics and physical AI, the crux of the problem that we face is, when you enter uncertain environments, even though they might look very nice, they might look fairly straightforward, you encounter substantial complexities and variations,” Menon said.
Panelist Steve Cousins, a roboticist at Stanford, talked about the challenges of “zero shot” design, where a humanoid robot is doing something that it has never done before. That requires a great deal of predictive design.
“We have to make sure that at a lower level, at a hardware level, that we thought about safety, we put the safeties in place so that the robot can try to do whatever it thinks you said, and have a good chance to do a zero shot success,” he said.
Complexities of Robot Movement
In addition, there are those pesky applications of physics, a product of universal laws of motion that are hard to understand except through experience. Karen Liu, a Stanford professor, talked about this work.
“I will argue that the way you talk about humanoids, like complex robotic systems, any movement requires some level of intelligence,” she said. “It’s not that as simple as you think, because if you’re a humanoid, you have to think about how much torque I need, to exert (to move) my ankle joint, my hip, my knee joint.”
Thus, she suggested, we face a data problem.
“When it comes to humanoid with locomotion, we have no data at all,” she said. “That is why we had to start with something, right? We start with a physics simulation. We started with what we know about motor control. Started with reinforcement learning, so we can run on trial and error. But in computer graphics, we also know that if you just learned from trial and error, it would take a long time.”
Opportunities for Design
The panelists also talked about some of what does seem within the grasp of highly evolved robotics.
“We actually have put together a business case for having these automatic drones, or some kind of robotic arms, embedded within a jet engine,” said Beena Ammanath of Deloitte. So if there was “a simple scratch on a plane of a jet engine, you could have this 3D printer manufacture something that could fix it on the fly, while the plane was flying. This is the imagination part, right?”
Cousins talked about coming out of an “AI winter” with a new focus on what’s possible.
“Something good happened, which is that all the AI people have been thinking about data, and importance of data, and the companies that survived that at all survived because they had learned to do something different,” he said. “They weren’t being so structured with their data anymore. They were able to deal with unstructured data.”
The same dynamic, he argued, might come into play with robot humanoids. “You have humanoids that are amazing [in] what they can do,” he said. “And at the same time, safety of humanoids is going to be a challenge for a while.”
There’s the falling down part, for example.
A Lot of Motors
“I kind of joke, like, what’s in common between humanoids and older adults?” Cousins added. “They fall down a lot. They break easily and are very expensive. You know, we have this problem, and we’re going to have this problem for a while, because there’s a lot of motors in a humanoid by definition, and any motor failing might cause it to fall, and that’s going to be a problem. But I also think there’s a lot of great stuff coming out of the efforts to build humanoids around people. We have much better motors, much more flexible systems.”
Menon addressed other design issues.
“You absolutely do not want your robot to lose power, and accidentally smack you in the face, or … fall on your cat or fall on your child,” he said. “That would be catastrophic. We don’t have great ways to address, fundamentally, the issue of hardware failure in the field, or of the capacity. So when we think about how to make robots usefully, I have a strong bias to being a problem solver.”
Their Takes on Moonshots
At the end, the moderator, Gordon Wetzstein, asked each panelist for a quick idea on important moonshot projects. Ammanath mentioned a robot to do housework, Menon looked to the principle of a robot solving a general problem, Cousins cited the growing number of elderly boomers needing physical assistance as a case study, while Liu referred to a “system to do things exactly the way that I want.”
Look for more in the wide-ranging discussion on robotics, and consider what these designs will soon bring us.