The savvy marketers at Boston Dynamics produced two major robotics news cycles last week. The bigger of the two was of course the electric Atlas announcement. As I write this, the sub-40-second video is steadily approaching five million views. A day earlier, the company stirred the community when it announced that the original Hydraulic Atlas would be abandoned a decade after its launch.
The accompanying video paid homage to the older Atlas’s journey from DARPA research project to an impressively nimble bipedal “bot.” However, after a minute the tone changes. Ultimately, “Farewell to Atlas” is both a celebration and a blooper. It’s a welcome reminder that every time the robot manages to land on the video, there are dozens of slips, falls and stutters.
Photo credit: Boston Dynamics
I have been advocating for this type of transparency for a long time. I would like to see more of this from the world of robotics. Simply presenting the highlight reel detracts from the effort that went into capturing these shots. In many cases, it involves years of trial and error to make robots look good on camera. If you only share the positive results, you are setting unrealistic expectations. Two-legged robots fall over. At least in this respect they are just like us. As Agility recently put it, “Everyone falls sometimes, it’s how we get back up that defines us.” I would go a step further and add that learning how to do well is just as important falls.
The company’s newly appointed CTO, Pras Velagapudi, recently told me that seeing robots doing their jobs at this stage is actually a good thing. “When a robot is actually out in the world doing real things, unexpected things happen,” he notes. “You’re going to have some falls, but that’s part of learning to run for really long periods of time in real environments. This is expected and is a sign that things are not being staged.”
A quick look at the Harvard Rules for Non-Injury Falls reflects what we intuitively understand about falls as humans:
- Protect your head
- Use your weight to guide your fall
- bend your knees
- Avoid taking other people with you
When it comes to robots, this article from IEEE Spectrum from last year is a good place to start.
“We’re not afraid of falling — we don’t treat the robots like they’re constantly breaking,” Aaron Saunders, CTO of Boston Dynamics, told the publication last year. “Our robot falls a lot, and we decided that a long time ago [is] that we had to build robots that could fall without breaking. If you can go through this cycle of causing your robot to fail, investigating the error, and fixing it, you can get to where it no longer breaks down. But if you build a machine, a control system, or a culture around never falling, you will never learn what you need to learn to keep your robot from falling. We celebrate falls, even the falls that break the robot.”
Photo credit: Boston Dynamics
The topic of falls also came up when I spoke with Robert Playter, CEO of Boston Dynamics, ahead of the electric Atlas launch. It is noteworthy that the short video begins with the robot in a prone position. The way the robot’s legs bend is quite novel and allows the system to stand up from a completely flat position. At first glance, it almost seems as if the company is showing off and is simply using the eye-catching move to showcase its extremely robust, custom-made actuators.
“There will be very practical applications for this,” Playter told me. “Robots will fall. You’d better be able to get up from a prone position.” He adds that the ability to stand up from a prone position could also be useful for recharging.
Much of Boston Dynamics’ insights into falls comes from Spot. While the four-legged form factor generally has greater stability (as evidenced by decades of videos trying to knock the robots over), there are simply many more hours of Spot robots working in real-world conditions.
Photo credit: Agility Robotics
“Spot travels around 70,000 kilometers per year on factory floors and carries out around 100,000 inspections per month,” adds Playter. “At some point they fall. You have to be able to get back up. Hopefully you’ll reduce your fall rate – we have. I think we fall once every 100-200 km. The fall rate has gotten really low, but it does happen.”
Playter adds that the company has long been “rough” on its robots. “They fall and have to be able to survive. Fingers can’t fall off.”
Looking at the Atlas outtakes above, it’s hard not to project a little human empathy onto the “bot.” It actually appears to be falling like a human, pulling its extremities as close to its body as possible to protect them from further injury.
When Agility added weapons to Digit in 2019, the role they play in the fall was discussed. “For us, arms are simultaneously a tool for moving through the world – think of getting up after a fall, moving your arms to maintain balance, or pushing open a door – and at the same time they are useful for moving objects to manipulate or wear,” co-founder Jonathan Hurst noted at the time.
I spoke a little with Agility about the topic at Modex earlier this year. A video of a Digit robot falling over on a meeting floor a year earlier made the rounds on social media. “With a 99% success rate over approximately 20 hours of live demos, Digit still experienced a few setbacks with ProMat,” Agility noted at the time. “We have no evidence, but we believe our sales team organized it so they could talk about Digits’ quick-change limbs and durability.”
As with the Atlas video, the company told me that something like a fetal position would be useful to protect the robot’s legs and arms.
The company uses reinforcement learning to help fallen robots pick themselves up. Agility turns off Digit’s obstacle avoidance for the video above to force a fall. In the video, the robot uses its arms to cushion the fall as much as possible. It then uses its reinforcement learning experiences to return to a familiar position from which it can stand up again with a robotic push-up.
One of the main selling points of humanoid robots is their ability to fit into existing workflows – these factories and warehouses are called “brownfield,” meaning they are not purpose-built for automation. In many existing cases of factory automation, errors result in the system essentially shutting down until a human intervenes.
“Rescuing a humanoid robot will not be trivial,” says Playter, noting that these systems are heavy and can be difficult to repair manually. “How are you going to do that if it can’t put itself into action?”
If these systems are to truly provide uninterrupted automation, they need to function well and come back up.
“Every time Digit drops, we learn something new,” adds Velagapudi. “When it comes to bipedal robotics, falling is a wonderful teacher.”