“Stop Laughing at those Clumsy Humanoid Robots”

CITRIS People & Robots Initiative Director Ken Goldberg is featured in the following article on designing humanoids.

Reposted from wired.com |  June 11, 2015 | By Matt Simon

THE HUMANOID ROBOT, built like a linebacker with an oversized head, tiptoes on two feet through the dirt. It’s free of any wires. It’s unleashed—but it’s now wavering. It starts veering right, tilting more and more, tilting and tilting, like the town drunkard, until the poor thing tips over and face-plants. It lies there in a cloud of dust, kicks its feet up a bit, arches its back, and gives in to defeat.

You may have seen these bipedal robots tumbling all over the Internet this week. They were all part of a competition in Pomona, California put on by Darpa, the far-out research wing of the Pentagon. After the Fukushima disaster in 2011, Darpa set out to encourage the development of robots that can assist in similar catastrophes: machines capable of working where humans dare not go. And so the yearly Darpa Robotics Challenge was born.

To explore something like a contaminated nuclear reactor, a robot would have to conquer not only piles of rubble in the facility, but also be able to open doors and climb stairs and ladders. In a human-designed space, the thinking goes, a humanoid robot would be best equipped to handle the job. And indeed, for all their clumsiness, the semi-autonomous robots (human operators still do much of the controlling in the challenge) passed some impressive tests, including driving an ATV. Which, quite frankly, is more skill than I can lay claim to.

“You have these environments similar to the Fukushima reactor meltdown, where robots could only get so much work done because the environments are designed for human beings to navigate,” says Doug Stephen, a computer scientist and physicist at the Florida Institute for Human & Machine Cognition who worked on Running Man, which took second place at the Darpa challenge. “You could imagine a door that maybe can’t be opened all the way, maybe it’s wedged, so a human being can sort of turn sideways and squeeze their way through that.”

The thing is, relief workers have had operational “tracked” robots for 15 years that roll along on tank-like treads. They even helped out in the aftermath of Fukushima, and still run tests there to this day. Bipedal humanoids, on the other hand, have never gotten near an actual disaster. They’re expensive, and you don’t have to be a physicist to notice these robots are top heavy. So are humans, and it took a whole lot of evolution to get Homo sapiens walking upright (likely a strategy, by the way, of freeing up the hands for gathering food, fending off predators, etc.—the type of utility that makes bipedal robots so coveted). If you don’t think walking is tough, ask a one-year-old human. Look at it this way: C-3P0 was built to interact politely with human beings, but R2-D2 was built to do actual work. So why even bother developing bipeds?

Well, when Darpa’s handing out $2 million for first place in its challenge, building a walking robot that drives ATVs must seem like a sweet deal.

‘I Need a Robot That’s Like a Meerkat or a Lemming’
Even if humanoids can overcome their bulk, they still might not be the best option in a search-and-rescue situation. Robin Murphy at Texas A&M probably knows the ins and outs of rescue robots better than anyone. A computer scientist, Murphy helped invent the field of rescue robotics, and she’s quite skeptical of the bipeds. “I have never had a request from any responder—and I have worked with responders in 20 countries, 19 deployments—no one has ever come to me and said, ‘I want a robot that’s like us, a human being.’ I get, ‘I want a robot that’s like a snake, I need a robot that’s like a meerkat or a lemming, I want a robot that can fly and do things like a hummingbird.’”

The majority of so-called “human-habitable” disasters, in places that human beings built, happen in mines. A humanoid bot would really struggle in that environment, says Murphy. Because they’re so top-heavy, pilots have to proceed slowly to make sure the bot doesn’t tip over. If they fall or get stuck, they can really muck things up for everyone. At the Pike River Mine disaster in 2010, for example, a bomb squad robot the size of a golf cart got stuck in a passage. It was too heavy for the team to lift out even if they could have reached it. After a couple of hours, the team was able to reboot the robot, but lost precious time. In mines, or pancaked buildings and other damaged structures, smaller bots have a distinct advantage.

Where a humanoid might actually excel one day, though, is exactly what Running Man pulled off in the challenge: turning valves. It sounds mundane, but human-habitable structures like chemical plants and nuclear reactors have a lot of valves. They’re meant to be turned by humans, of course, so they’re generally placed up high, out of the reach of smaller tracked robots. “The one area where you would be likely to use a humanoid is your chemical and radiological events, where you’ve got this hazmat leak,” Murphy says.

Even the people at iRobot—makers of tracked bomb disposal bots and the hockey puck-shaped, wheeled Roomba cleaning robot—agree. “Manipulation tasks, using their arms to do things like open a door, turn a valve, or unplug and plug an electrical cord are very hard tasks,” says Chris Jones, director of strategic technology development at iRobot. He also complemented the autonomy the Darpa robots showed. It takes some serious coding to get a semi-autonomous bipedal robot to stay on its feet, not to mention detect its surroundings.

If you ignore the bipedal part of humanoid robots, their accomplishments at the challenge were pretty remarkable, however dopey they may have looked pulling them off. And those advances could well be applied to bipeds’ tracked counterparts. Indeed, iRobot’s tracked PackBot, which has been working in Fukushima, now has some autonomous features, allowing it to, for example, automatically flip itself over with its arm if it’s been upended. With Darpa actively encouraging teams to exchange these kinds of findings, advances will spread not just among the humanoid robots, but to other vehicles as well.

Meanwhile, though—possibly because of the rules Darpa applied to this particular challenge—it was the robots with two legs that also had the ability to drive, use tools, and perform other tasks a roboticist might classify as requiring dexterity. The wheeled bots in use today can’t do any of that stuff. After all, Roombas are notoriously bad at opening doors.

Humanoid Hybrids
Now, this is where the philosophy of robotics gets really interesting. Humans have designed the environments that these robots need to work in, yes. They’ve got stairs to climb and doors to open and, most difficult of all for a robot, ladders to somehow ascend. A tracked bot might have ED-209-type problems with stairs, and totally fail with the doors and ladders. A humanoid bot could handle all three, hypothetically. But what if there’s an entirely different design roboticists haven’t yet considered?

Human beings can be a bit … egomaniacal as a species (the $10 word for it is “anthropocentric”). “We’ve been interested in this for millennia, that we want to build something in our image,” says Ken Goldberg, a roboticist at UC Berkeley who did not participate in the Darpa challenge. “That’s a deeply rooted instinct.” While humans may have designed the Fukushima nuclear plant, that doesn’t mean only humans can navigate it. Roboticists could design a machine that looks like nothing else in nature, yet still functions like a person. If you want to fly, you can build an airplane … but you can also build a helicopter.

Goldberg thinks hybrid designs could have conquered a lot of Darpa’s challenges—driving cars, climbing stairs, what have you. “It may not be wheels, but it may be treads or a variety of different locomotive designs,” he says. Indeed, the robot that took first place at the challenge, the Korea Advanced Institute of Science and Technology’s HUBO, was a transformer of sorts. It could walk like a biped, but it had wheels attached to its knees. In a jam, it could kneel and roll around.

Even more ambitious designs populate labs around the world. Tiny flying drones could collaborate autonomously. Groups of small crawling robots could collaborate and share tasks like an ant colony. At the Darpa challenge, a team from the Jet Propulsion Laboratory fielded a bot that looked more like a cross between a spider and a gorilla. The RoboSimian was a brilliant example of nonlinear thinking and engineering. It was also horrifying.

“I think we’ve learned that humanoids are hard,” Goldberg adds. “So maybe it’s back to the drawing board. We have to think a little more about the design. And maybe it won’t look quite human.” If a tracked robot can steal some of a humanoid’s dexterity and autonomy, roboticists could make something that’s greater than the sum of its well-oiled parts.

And all of this isn’t to say that the two classes of vehicles, humanoids and tracks, can’t one day work in tandem. Each has its own strengths and weaknesses. Tracked robots may be indispensable in mines and collapsed parking garages, where humanoids will struggle. But conversely, a biped could one day turn a valve in a nuclear reactor gone haywire. Different kinds of disasters demand different heroes. After all, C-3P0 and R2-D2 always worked better as a team.