No big deal, Boston Dynamics’ Atlas can perform backflips now | Hacker News Reader

No big deal, Boston Dynamics’ Atlas can perform backflips now

theverge.com

148 pointshourislate69 comments

Why are so many users in the comments on here fear mongering? Don't you understand we already have dumb solutions for killing and making war? Bombs, bullets, Chemical and biological weapons, non of which need a complex robotic platform. How will human shaped robots change anything when we can build flying, hovering grenades and mass deploy them with COTS components?

You know what does need a complex platform like Atlas? Disaster Rescue, Medicine, Telepresence, Automation etc

To get an idea of progress, here's state of the art 10 years ago. Note that the actuators are pneumatic pistons, not direct drive motors (very hard to control, springy like muscles):

https://www.youtube.com/watch?v=eXwpIQMUikU

Does anyone know what breakthroughs in control systems have gone into making this possible ? Are there any papers describing how Boston Dynamics is programming this ? are they hardcoding the backflip from human data and then overlaying some stability code ? or is the backflip entirely generated by the underlying controls ?

Why is the robot wearing Plimsolls? Compliance and/or friction at the contact surface?

I still wonder why Google sold Boston Dynamics, and no other transportation or defense manufacturer scooped them up. The technology certainly looks amazing. Are the commercialization prospects really that bad?

The video is amazing, but let's be honest: fluid motion without falling is something you can expect from a minimum wage employee. And big robots are physical machines that are likely to cost rather more than a minimum wage scrub to hire/acquire and pay/maintain.

This is a different world from, say, machine learning. It's equally true that we're just now reaching the point where ML is as good as we are at recognizing photos of our children or transcribing our voice mail. But those ML algorithms can be deployed at cloud scale for pennies, so it's worthwhile where physical things aren't yet.

There's one proven application of this technology: entertainment.

> fluid motion without falling is something you can expect from a minimum wage employee. And big robots are physical machines that are likely to cost rather more than a minimum wage scrub to hire/acquire and pay/maintain

Fluid motion without falling, along with a lot of physical activities, turn out to be very hard problems both for robotics and AI[0]. So yes, while the applications may not be as potentially far reaching as with machine learning, as a technological achievement getting a bipedal robot to do a backflip is more impressive.

[0]https://en.wikipedia.org/wiki/Moravec%27s_paradox

Trust me, I know. I was answering the question of why a solution to a very hard problem in one space sees pervasive application, and a similarly impressive solution in robotics gets no interest beyond backfliping robots on youtube.

And big robots are physical machines that are likely to cost rather more than a minimum wage scrub to hire/acquire and pay/maintain

Is that true? And if it is, will it remain true for long?

To staff a position with humans 24x7x365 would cost over $100K/year ($7.25x24x365 = $65K, add in overhead and you get to $100K+ easily).

If you can buy a robot to do the job for $100K, plus $1K/month for a maintenance contract), then you break even after the first 13 - 14 months. Add a year if the robot costs $200K.

I don't know what these robots will cost, an industrial robot arm is in the $50-$200K range (or higher).

Yeah, I guess that they are still looking for the killer app. For a while, it looked like they could be deployed as military mules; but they were too loud (due to the internal combustion engine) for that purpose. So, what can they do?

Meantime, for $7.25/hour, you can hire a human that (hopefully) understands spoken words, and can do jobs requiring fine hand motor skills, e.g. pick-and-pack in warehouse, or serve meals and drinks, stock shelves, flip burgers etc. Boston Dynamics robot cannot do any of that. It can walk over difficult terrain, and flip, but who needs that?

Perhaps they should focus on the "hands/arms" parts (as opposed to legs). Industrial robots don't use legs, they are fine with wheels and rails. Also legs are not really what stands apart - cheetahs and horses and mice can use their legs better than humans, but so what. What matters is hands - opposable thumbs, and so on.

> Boston Dynamics robot cannot do any of that Yet. It's clear that they are not aiming for a commercial product anytime soon, this is ongoing research and the progress over the last decade has been huge and will only accelerate.

Seems like the first use will be working on dangerous environments.

There are already robots being used for mine sweeping, bomb disposal, nuclear reactor inspection, even mars rovers. They all make do with just wheels, or sometimes rails.

It is not clear if adding legs to the above robots would make them any better.

Legs can access an order of magnitude more terrain than wheels.

Think looking for survivors in rubble after an earthquake.

I think a simple drone can make this use case much easier.

Who said their market was to replace minimum wage workers?

You made that up then tried to convince us that it’s a bad idea.

Bingo. Their market is the DoD and anyone said department will let them sell to.

The fair comparison isn't $10/hour workers, it's the entire cost of training, equipping, and employing a soldier (even in non-combat roles) and/or current military tech such as drones.

I did no such thing. I said their market was not billion-plus consumer applications as the case with much ML research, and that this might explain some of the lack of commercial interest mentioned upthread.

I won't take a position as to whether this robot is a "bad idea". But pretty much anyone can see that (relative to the problem difficulty) it's a "worse idea" than, say, photo recognition.

I think Boston Dynamics is more of a huge practical research project that won't have commercialization results for at least 10 years (random number I generated on intuition), whereas companies like Google are committed to justifying their efforts on quarter-by-quarter basis.

If a company owns Boston Dynamics and has one bad quarter, you have to trim the fat or deal with the bloodbath of angry impatient investors

EDIT: That being said, companies like Boston Dynamics are, in my opinion, Star Trek tier, and, as a tangent, is a perfect example of why we need to eventually live in a world that doesn't thrive on currency profit

Boston Dynamics' technology may be Star Trek tier, but its main use in the near future is going to be to kill, or help kill, rebels in Yemen

That's why I think Google dropped them. Google doesn't want to be seen as Saudi Arabia's accomplice in a present or future conflict

It doesn't need to be. Just sell the robots to both sides, and let the robots destroy each other.

Guess which side can afford robots and which one can't

Throughout history military power has been tied closely to wealth and population. Theorists of AI and robotic weapons talk about how that relationship could significantly weaken:

You don't need a large population to build a robotic army, and while you need wealth, it may be that you just need better AI developers (who could be AIs themselves) more than you need more robots. Or maybe you just need to find the right combination of AI, hardware, and tactics first; win quickly and it's too late for the enemy to learn.

Singapore with its high-tech foundation, for example, could become one of the most powerful nations in SE Asia, and maybe even beyond.

It does kind of need to be.. for the same reason we use drones against people and not just infrastructure.

Google is kind of like Skynet, if they had killer robots, the Terminator comparison would be too robust to ignore. Kind of a joke, kind of not a joke.

I heard somewhere that the military wants to love these robots, particularly for hauling stuff into unfriendly terrain, but the main reason they simply aren’t viable is that they are much, much too loud. There were other problems too, complex maintenance in the field and the likes, but apparently that is less of a hassle than getting ambushed because the enemy heard you coming a mile away...

IIRC, a large part of that is down to the great dilemma of the modern age, power storage. The bots can't carry a battery big enough to give them realistic outdoor range, so outside of the test lab they require a small internal combustion engine, which of course is very noisy.

With all the battery research being done for mobile devices and electric cars, I wouldn't be too surprised if there's a breakthrough there in the next couple of decades. Still, the servos alone may be too noisy for military use, I don't know.

If you believe the news, there’s a breakthrough every other week!

Fuel cells seem like a good mix of quiet yet high energy density. Also, human style walking does not take much energy when done efficiently. But, wheels are even more efficient and equally effective in the vast majority of urban settings.

Did you watch their video of Handle? That's where it's at.

"Hauling stuff in rough terrain" seems like the PR-friendly version of what they want this technology for, but I don't think it takes too much imagination to think up other uses.

Also, there is this issue: In what way is the robot superior to a donkey?

You can give a robot precise coordinates and let them go for it.

This gives me an idea: design a small robot to ride the donkey and steer it using, e.g., a carrot on a string.

> I still wonder why Google sold Boston Dynamics

It's a money pit, albeit a cool one. But that doesn't matter. Ruthless doesn't see the coolness, just the dollar signs.

> and no other transportation or defense manufacturer scooped them up

See "money pit." Defense manufacturers are all about how to extract money from the government, and this doesn't do that for at least 10 more years. Transportation is already a cut-throat industry built around fees and doublespeak, so... no joy there either.

Boston Dynamics needs a R&D friendly rich company to leech off while doing their work, which could cost upward of a billion dollars over the next five years. Softbank, from what I can tell, is pretty much exactly that - they're willing to invest in Uber for goodness sake...

Don't be evil? Kind of hard to build robots to help kill people and spruik this motto at the same time. I don't buy the money pit argument, rarely been Google's sole yardstick.

Kicker is in the US, government funded research is primarily through defense spending. If you're going to turn your nose up at that, you're on a long expensive path to monetisation in this space.

> Kicker is in the US, government funded research is primarily through defense spending

Agreed. There are serious consequences to shifting government spending to mostly defense and reducing everything else. Also, it creates a large group with vested interests in continued and increasing defense spending, which includes favoring use of the military over other tools of state power.

Google is an advertising company. I wonder why they bought BD in the first place.

As a former gymnast I can confirm that is indeed a backflip. Pretty good one at that!

In the last animation the robot extends its legs after being already on the ground.

Failing to see a physic justification for it, I wonder how much of these movements is the result of a predefined "script".

to me it's more a bad anticipation of the landing leading to a bad response

The Boston Dynamics robot is extending its legs to return to both a fully stable state and its optimal “home” or resting position state.

As to the question of whether the robot is following a ‘predefined “script”’ (implying, I think, that the developers might have “cheated”) if it were easy to make a “predefined script” that let a robot do this kind of thing, it would have been done years ago. Stability of complex dynamical systems is an incredibly hard problem, and without real active stability control there’s no way to get a thousand pound robot to balance like this. There’s no cheating involved, they simply did the hard things and made them look easy.

Yeah, here's the competition: https://www.youtube.com/watch?v=AEJeIUTValE Not even close.

What's the physics justification for doing a backflip? Obviously it's scripted, the same way a gymnast's tumble is scripted. The robot isn't sentiently deciding that it wants to do some flips today. The interesting part is that it is capable of doing the flip.

The GP means "performing the flip in a pre-programmed manner", rather than dynamically reacting to actual sensor input. However, that strikes me much harder than actually doing it for real, which is still amazingly hard.

Regardless of where we are on the continuum between "move this actuator to this position now (1980s)" and "go jump onto each of those boxes and do a flip off the last one (2030?)", this is damn impressive.

This is a kind of AI/Robotics research that should be regulated. In 10-20 years, they will be quieter, cheaper, and even more capable.

How much resistance do we have against a horde of robotic raptors carrying weapons?

If you disagree, please state why instead of downvoting silently because this is a serious issue and we ought to have rational discussions about it.

> How much resistance do we have against a horde of robotic raptors carrying weapons?

Atlas would suck against self-flying grenade-dropping drones.

I wonder if they’ll start to model the robots “strength” after ratios found in humans. I thought its movement was squarely in the uncanny valley, at least partly because its muscle were probably way stronger than what a human could do.

No reason to, as they're targeting military applications. In fact the robots' super-strength may intimidate enemy soldiers into surrender.

I wonder. Combining strength, speed, precision, and range of motion in a small package can be tricky, and requires trade-offs. It may be no stronger than it needs to be to perform the motions we're seeing.

Fast forward to me being a minor pawn in the resistance, running through the woods from heavily armed Atlas droids.

It's not the backflips that is fascinating to me. It's how human it looks.

Wow. America's future murder machines really are getting spritely.

The dog like one is more chilling to me.

https://youtu.be/kgaO45SyaO4

  video
  paragraph
  gif
  sentence
  gif
  sentence 
  gif

It's a multi-paragraph description of a video that takes longer (and is far more boring) to read than watch. And there are choppy gifs illustrating the paragraphs narrating the video.

I don't understand why this is being done. Presumably it increases engagement?

So you can watch it at work without "sitting down" and watching a video. And you also don't need a headset.

How long before we see Atlas jumping from roof tops in Hong Kong?

The question is how long you live to te...

LOST CARRIER

These videos are always mindblowing to me, but I rarely see anyone ask "how many failures did it take to make the video?" There's a huge difference between this being a 99% performance and a 1% performance.

They should do it on stage.

Wow awesome awesome. Why did Google sell this to softbank?

Google does not want to become an arms dealer

This might be googles IBM moment.

LA council just passed approval for armed police drones. When I look at Atlas, I'm prod how far technology is coming and I am not afraid of this machine; but I am afraid of what our government will eventually do with them.

LA city joins North Dakota and Connecticut states for the armed drones.

It is truly horrifying.

The LA council did not just approve armed police drones.

They're approved strictly for surveillance purposes. It specifically forbids any weapons capability.

"A drone would not be used with any weapons capabilities, including any non-lethal or less-than-lethal systems, according to the proposed guidelines."

http://www.dailynews.com/2017/10/17/amid-protests-la-police-...

https://www.cnbc.com/2017/10/18/lapd-drones-set-to-be-deploy...

Reply on news.ycombinator.com