The savvy marketers at Boston Dynamics created two major robot-related news cycles last week. The bigger of the two was, unsurprisingly, the announcement of the electric Atlas. As of this writing, the less than 40 second video is steadily approaching his 5 million views. The previous day, the company struck a chord with the local community when it announced that the first-generation hydraulic Atlas would be put out to pasture 10 years after its introduction.
The accompanying video celebrated the old Atlas' journey from DARPA research project to impressively agile bipedal “bot.” But after a minute, the tone changes. In the end, “The Farewell Atlas” is as much a celebration as it is a failure. It's a welcome reminder that every time a robot videotapes its landing, dozens of slips, falls, and sputters occur.
Image credit: Boston Dynamics
I have long supported this kind of transparency. That's something I would like to see more of in the world of robotics. Simply showcasing your highlight reel will negate the effort you put into getting the shot. It often takes years of trial and error to get a robot on camera. Sharing only positive results sets unrealistic expectations. A bipedal robot falls over. At least in that respect, they are just like us. As Agility recently stated, “We all fall down sometimes, but how we get back up defines us.” I would take it a step further and say that learning how to fall well is just as important. I would add that there is.
Pras Velagapudi, the company's newly appointed CTO, recently told me that it's actually a good thing to see robots doing the work at this point. “When robots actually go out into the world and do real activities, unexpected things are going to happen,” he points out. “You'll see some falls, but that's part of learning to run for very long periods of time in a real-world environment. That's expected and a sign that you're not staging things. ”
A quick read of Harvard University's Rules for Falling Without Injury reflects what we, as humans, intuitively understand about falls.
Protect your head Use your body weight to induce a fall Bend your knees Avoid carrying others with you
When it comes to robots, a great place to start is last year's IEEE Spectrum article.
“We're not afraid of falling. We're not treating robots like they're going to break all the time,” Aaron Sanders, Boston Dynamics' chief technology officer, told the magazine last year. “Our robot falls over a lot, but one thing we decided a long time ago was that he [is] This means we need to create a robot that won't break even if it falls. If we can repeat the cycle of forcing a robot into failure, studying that failure, and fixing it, we can progress to the point where it no longer breaks down. But if you build a machine or a control system or a culture around never falling, you'll never learn what you have to learn to make a robot never fall. We celebrate falls. Even if it's a fall that breaks the robot. ”
Image credit: Boston Dynamics
The topic of falls also came up when I spoke with Boston Dynamics CEO Robert Player prior to the launch of the electric Atlas. Notably, this short video begins with the robot in a prone position. The way the robot's legs arc is very innovative, allowing the system to stand up from a completely flat position. At first glance, it feels as if the company is simply using the flashy movement as a way to showcase its extremely robust, custom-made actuators.
“It would have very practical applications,” Player told me. “The robot is going to fall over. It better be able to get up from a prone position.” He added that the ability to rise from a prone position could also be useful for charging purposes.
Much of Boston Dynamics' learning about tipping came from spots. While quadrupedal form factors are generally more stable (as evidenced by decades of failed attempts to kick robots in videos), in real-world situations the Spot robot works. The time is simply much longer.
Image credit: Agility Robotics
“Spot walks approximately 70,000 kilometers a year on factory floors and performs approximately 100,000 tests per month,” Prater added. “They fall eventually. They have to get back up. I wish the fall rate would go down, but that's what we do. I think they fall once every 100 to 200km. The rate of decline has actually been smaller, but it does happen.”
Player added that the company has a long history of treating robots “roughly.” “They have to be able to survive when they fall. No fingers fall off.”
Watching the Atlas outtake above, you can't help but project some human empathy onto the “bots.” It really appears to be falling down like a human, holding its limbs as close to its body as possible to protect itself from further injury.
Digit achieved a 99% success rate in about 20 hours of live demos, and still fell a few times on the ProMat.
Although I have no proof, I believe our sales team made adjustments to talk about Digits' quick-change limbs and durability. #conspiracy theory pic.twitter.com/aqC5rhvBTj
— Agility Robotics (@agilityrobotics) April 6, 2023
When Agility added weapons to Digit in 2019, there was discussion about the role weapons would play in a fall. “For us, arms are tools that help us manipulate and carry things, as well as help us move through the world: getting up after a fall, swinging our arms for balance, pushing doors, etc. in action,” said co-founder Jonathan Hurst. It was pointed out at that time.
I spoke a little with Agility about this topic at Modex earlier this year. A year ago, a video of a digital robot falling over on a convention floor made social media buzz. “Digit fell over several times with his ProMat, even though he achieved a 99% success rate in about 20 hours of live demos,” Agility said at the time. “Although I have no proof, I believe our sales team arranged for him to talk about Digits' quick-change limbs and durability.”
Similar to the Atlas video, the company told me that something similar to the fetal position would be helpful when it comes to protecting the robot's legs and arms.
The company uses reinforcement learning to help fallen robots get back on their feet. In the video above, Agility cut off Digit's obstacle avoidance and forced him to fall. In the video, the robot uses its arm to reduce the fall as much as possible. It can then use reinforcement learning to return to its familiar position and from there it can stand again in a robot push-up.
One of the main selling points of humanoid robots is their ability to be integrated into existing workflows. These factories and warehouses are known as “brownfields,” meaning they are not custom-built for automation. In many cases of existing factory automation, when an error occurs, the system effectively shuts down until human intervention occurs.
“Rescuing a humanoid robot is not an easy task,” Player said, noting that these systems can be heavy and difficult to manually modify. “What are you going to do if you don’t get up off the ground?”
If these systems are truly to guarantee uninterrupted automation, they must work fine and restart immediately.
“Every time the digits collapse, we learn something new,” Velagapudi added. “Falling is a great teacher when it comes to bipedal robotics.”