Atlas, the Next Generation [video] | Hacker News Reader

Atlas, the Next Generation [video]

youtube.com

509 pointsbpierre175 comments

Open articleView on HN

If people are interested in looking at more cool content on the subject, I recommend you looking at:

- http://drc.mit.edu/technology/

- http://www.cs.cmu.edu/~cga/drc/

- https://github.com/RobotLocomotion/drake/wiki

- http://robots.ihmc.us/humanoid-control-workshop/videos

I attended WPI during 2013-2014, and worked with Atlas. I am very excited and impressed by Boston Dynamics. (And hopefully start working on the field again.)

Can someone explain what would be the single greatest challenge (from a research perspective) in making such a robot today? For example, is it:

1. The sensors?

2. The actuators?

3. Coming up with accurate dynamical models?

4. Solving the models accurately?

5. Solving the models efficiently?

6. Making the models robust to inaccuracy/noise?

Yes, I realize all of these are probably hard. What I'm trying to understand is whether the biggest challenge is coming up with e.g. an accurate (possibly nonlinear) dynamical model, or with solving the model (efficiency/accuracy), or with making existing models that are otherwise already completely accurate robust to outside noise, or with the manufacturing aspect (precision), or whatever.

Power.

All the other technologies are increasing extremely rapidly but battery technology isn't.

We are at a point where a breakthrough in batteries would have an incredible effect across a huge number of fields.

Everything is getting more power efficient though, so you will at least get more computing power for the same amount of Watts.

Not everything. The energy required by physics for lifting boxes or driving a car up a hill is pretty much fixed. Better batteries would help with both of those.

Electric jets as well.

The energy to raise a fixed mass a certain height isn't going to change.

Not to mention an electric car with a thousand mile range would be huge.

It would certainly put a whole new perspective on range anxiety.

I don't think the robots who can do those things are near the physical limits, though.

I agree. A quick Google search suggests that the human body requires about 100 Watts[1] of power. On the other hand, a small 12 volt car battery is able to provide 45 Amps over an hour, which implies that it can support a power draw of 540 watts for an hour. Even if human activity required 200 Watts, an android with human level power efficiency should be able to operate for over 2 hours with the technology of today's small car batteries. Of course, my math could be wrong :-)

[EDIT] - My point here is that even if improvements to battery technology remain slow or constant, with sufficient innovation in energy efficiency and algorithms, we could still have robots with usable to good running times. Just figuring out a reliable walking gait and building materials for passive dynamic walking can significantly reduce the power requirements.[3]

[1] - http://hypertextbook.com/facts/2001/JacquelineLing.shtml [2] - http://www.chem.hawaii.edu/uham/bat.html [3] - https://en.wikipedia.org/wiki/Passive_dynamics

Also, teach the robot to replace it's own battery or fuel.

Yikes! As long as it has some parameters about human safety being more important than batteries or fuel, then by all means...

If humans can go multiple weeks without food, then how do biological organisms solve the problem of power storage better? Or does our brain and muscles etc just require less power?

Mostly the latter, I believe.

The human brain uses the incredible 20 Watts to achieve things that modern megawatt supercomputers aren't capable of.

I think the figure is closer to 100 W average across the day since the brain cannot exist in isolation of the body which needs ~2000 calories daily, but still, we're fairly efficient. We're built with advanced nanotechnology.

Part of the reason is that humans are not run on batteries. Humans are more similar to fuel cells in that we consume oxygen to produce energy. We inhale oxygen, combine it with our energy stores and exhale 80% of the mass[2].

An average American man needs about 3400 Calories each day which corresponds to about 1 kg of oxygen consumed[1]. Similarly the average American woman uses 2550 Calories and .75 kg of oxygen.

[1] Design Rules for Life Support Systems http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/2004001... [2] http://www.livescience.com/49157-how-fat-is-lost-body.html

3400 calories? That seems awful high. Perhaps if you are into bodybuilding but no way an average person needs that much. Closer to 2000-2500 maybe.

The biggest problem is that we still don't understand the principles of walking, locomotion, and manipulation. Biology doesn't come with any documentation.

This is much more than the problem of not having good sensors, not having powerful enough actuators, or solving models. Without understanding the fundamental principles we may not even get the hardware right. Recall early attempts at heavier than air flight that attempted to copy and scale up a bird rather than understand the principles behind how wings worked.

Watching that, especially that part starting at 1:23 makes you realize quickly that in about 10 years, 95% of all jobs can be replaced by machines. Scary and interesting at the same time.

Does that mean we live in a happy utopia where 95% of the tedious/dangerous work has been taken care of and we instead get to spend our time focusing on what we find inspirational/creative, or do we live in a dystopia where 95% of job opportunities have vanished and we can no longer sustain a consumption based economy.

Maybe those are one and the same.

There's a TV show (haven't watched it yet) about a future where society is split between something like 15% employed and 85% unemployed. The unemployed live in this desert wasteland or whatever while the employed are living in a big o'l building structure.

I would say it's far fetched and unlikely except that's kind of how the world is today.

Zardoz!

To extend your point. The jobs that are easily programmable like farming, and warehousing are easily replaceable. Some tasks may even require using humans and robots together like construction or search and rescue.

Some things are indeed not replaceable by robots and I think "most" professions* may not be able to be replaced by robots. These could be things like professional chefs (not fast food), programmers, and masseuse.

*By professions I mean the raw number of unique jobs not the number of people who work in those positions.

That being said in the U.S. with the "College Driven" economy I doubt it will ever affect 95% of the population at once. Generally professions will disappear one by one, but those people can retrain and if they don't it won't matter as the next generation will obviously not train to fill positions that don't exist (that's not logical). Eventually the new generations train to fill different job positions.

That's far from guaranteed, and looks less likely all the time. See, the "Luddite Fallacy Fallacy." Gwern gives it a nice treatment here: http://www.gwern.net/Mistakes?cx=009114923999563836576%3Adv0...

Well when the cars were invented did people start living in a dystopia where the car moguls owned mobility itself or did people just got around much faster and had better lives in general.

horses have had a pretty rough run

http://www.humanesociety.org/assets/pdfs/hsp/soaiv_07_ch10.p...

It was mixed. People got around faster and stuff was cheaper. On the other hand, we got fat and dependent on cars, which is particularly bad for the poor (since cars are inherently expensive), and for kids' independence. That's why you're seeing many metro areas starting to prioritize walking, biking, and transit.

On balance the development of cars was good, obviously, but let's not kid ourselves, the downsides were significant.

I'm not sure if it is an established fact that people got fat because cars came along (before cars there were carriages). I'd rather bet that diet is the culprit.

Probably both, but I think they've shown at least a correlation between car commute time and obesity.

If we look at agriculture mechanization, we went from 80% of the population to 5% to feed everyone. While this is good for everyone, I don't think farmers are living some sort of utopia.

Rather most of them are indebted for decades to pay for their complex equipment.

I believe the introduction of robot will follow a similar pattern, some industry will see their number of employee reduced drastically but they won't necessarily increase their margin. The customer will likely win.

The thing is that people who were employed in farms moved to factory. Nowadays I don't know where people who lost their job to robots will move.

As before, people will move to the new world. This time it is called Mars.

You think 75% of the population moving to Mars is a likely outcome?

More like the virtual world.

There were still plenty of jobs that could only be done by humans though. We might be reaching a point where there will be very few.

First will come the dystopia, the utopia might be later. In the early years the best positioned companies will reap the huge benefits of this new field. In time, the benefits will trickle down to ordinary people. The same was with cars and computers.

Even when ordinary people will reap benefits from robotization, the big bucks will go to the Google or Apple of the day. We might have our daily needs met by the future economy, but big companies will dominate the finances. We'll get to play with the new toys, though.

This is the first video that has actually given me that feeling. Maybe it's because of how anthropomorphic they've made it's movement compared to other robots, but I can see this guy doing very real work in the very near future.

10 years for 95% of all jobs is kind of too much too soon I'd argue. Big dog debuted 10 years ago for instance...even supposing non-linear growth, thinking that something like atlas in 10 years waiting tables, cooking gourmet food, greeting hotel guests, making plants, etc...

A portion? Yes. 95%?

I think you vastly underestimate how expensive that robot would be to manufacture and maintain. Of course, costs could come down over time but it's going to take a lot of work to make it economically viable for the types of jobs you are considering replacing.

it seems that it might be close to doing what a low-skill laborer would do in a factory setting. remember that it should be able to work for longer hours, and not require benefits etc. i suspect that companies would start to purchase such a device at $100k. that price seems achievable, probably within a few years. as quantities increase, i suspect that the price will fall dramatically thereafter.

The rule of thumb is you have to recoup capital costs in 2 years to make it worthwhile.

Given the above assumptions that seems doable to me. 2 years of working nearly 24/7 with virtually no costs(compared to wages/benefits) would more than pay for itself. But I don't think the above assumptions are necessarily correct, that sounds a bit too cheap to me.

It may be doable, I am not trying to be overly negative. It's just that I probably wouldn't target minimum wage workers with this robot.

Also you say "no costs" but do you honestly think these would be maintenance free? There would probably be a fairly significant cost to keep a fleet of these robots running as they are incredibly complex as compared to say an amazon warehouse robot.

Oh yes, I didn't mean there are no costs, just that beyond the initial expense they are much cheaper than human employees.

As mentioned elsewhere in the thread, it's generally a lot simpler to redesign your workspace to take advantage of a simpler robot design than it is to get a bunch of more general robots like this that still aren't quite there anyway. Amazon's warehouse robots come to mind in every one of these conversations: https://www.youtube.com/watch?v=quWFjS3Ci7A

The hardware is almost there, at least, for general tasks. I'm not sure about software. Is there anything recent showing machines reading human-targeted instructions to accomplish some task (change a tire, set up a desk) and then fulfilling the task, without having been programmed or trained on that task already? That would be impressive.

Perhaps why YC is funding research on basic income... https://blog.ycombinator.com/basic-income

Datacenters for information crunching, cyborgs for physical crunching... heaven is coming. Or is it.

I'd suggest 20 years, and 10% of jobs can be replaced by machine is a more accurate assessment. The latest revision of the Atlas is an interesting technology demo - but I think progress in this space will be much, much slower than automated driving (which I expect to see 25% penetration in 10 years - that is, 25% of all driving, based on miles driven, will be hands/foot free).

Even something as simple (and very, very low cost) as making up a room in a hotel, is probably 25-30 years off at least.

I think you're right. And I would add that part of the challenge is not just the technology, but the cost as well.

You don't just need a robot that can make up a room as good as a human does it, it also needs to be cheaper than that human to "hire".

This whole current thing with foaming at the mouth over bots taking all our jobs in the next few years reminds me a lot of how VR was being talked about in the early 90s. It's still early in the game, folks.

This stuff is not small. The cost reductions achieved through shrinking transistors are irrelevant here. You can only go so cheap with moving parts that need to be strong and durable.

For widespread adoption, cost is probably the biggest factor. Actually getting close to the same performance of a full time worker who grosses under $20k a year for a lower price is going to take a lot longer than many people seem to think. Throughout human history, we've repeatedly solved labor problems by throwing heaps of desperately poor human slaves at them, and we continue to do so-- competing with that model will be difficult.

Absolutely correct - keeping in mind that most of the (often undocumented) people cleaning your room in places like California are making around $10K/year or less.

Fully automated "Cleaning a hotel room" is actually an insanely complex task across tons of disciplines - solving that problem will also result in you solving a lot of the Artificial Intelligence challenges.

I think what we'll see much earlier, is a hybrid solution, where automated bots that can do 99% of the tasks send an interrupt to a controller when they encounter something they aren't certain of, such as, "Is that a lifelike doll laying on the floor, or a child lying very quietly."

Now you make me want to try and design a self cleaning hotel room.

I think the trickier aspects of hotel room cleaning could be mitigated by creative design of the accommodations.

If it was a stark room with a bed a couple tables and a flat screen there would be little to knock off tables or break.

Make the whole bathroom operate like a giant dishwasher and come up with a more robot friendly fitted bedsheet then use some kind of industrial roomba.

It wouldn't be cheap but it can be done.

Depending on how far you are willing to stretch your definition of hotel room I bet we could automate the cleaning of those Japanese pod/tube hotel rooms right now.

The automatically cleanable bathroom already exists. I have used them years ago in cheap hotels. After taking a shower, it locked the door once you exit and it rinses the whole room with some detergent and water for 30 seconds.

I just had a mental image of it bugging out and flooding the room with detergent with an occupant still inside...

Self-cleaning public toilets, that worked exactly like this, went through a bit of a phase in Australia 10-20 years ago. They started decommissioning them after they began trapping children during the cleaning cycle: http://www.theage.com.au/news/national/auto-toilets-a-flush-....

Heinlein in the 60's suggested rooms that would have a daily "wind" cycle to blow off any accumulated dust and anything loose would be forced to one area where it could be picked up to place back in the room, recycled, or disposed of.

Yeah, I know what you mean. We'll have several dozen heading off to Mars, various moons of Jupiter, and perhaps the outer planets. Because humans won't be going, it'll be a fraction of the cost.'

Btw, why do people always worry about the jobs technology will take away? The good old days when everyone was a farmer.

>Btw, why do people always worry about the jobs technology will take away?

Because if technology destroys your industry you lose. Automating textile work lead to greater wealth for the world as a whole, but it didn't make textile workers better off. The fact that a new job was created somewhere doesn't matter to them if they can't do it.

Robots - they took our jobs! https://www.youtube.com/watch?v=2ZBYLIUqmIs

I think in 20 years it's still much easier and cheaper to rebuild a human workplace to be suitable for (very dumb) robots, than it is to use humanoid robots for human tasks.

We already replaced humans for many tasks such as assembly etc. The next step is cognitively challenging but physically easy tasks like driving cars.

A lot of "simple" human tasks such as making sandwiches in a restaurant require fine motor skills, creativity and other very difficult robotic traits. Those jobs will be done by robots long long after we as humans stopped driving cars or building them. Given the budgetary difference between making a sandwich and going into battle, we can be pretty sure sandwiches will come last. If anyone is still around to eat them by then that is.

Purpose-built robots for things like "making sandwiches" will always be cheaper than general purpose robots. There really isn't a culinary task today that can't be fully automated (you can look at the frozen food section of a grocery store for examples of almost every cuisine type). The reason this hasn't happened at your corner diner is that manual labor remains much more inexpensive in the short and long term for low volume production (ie, restaurant vs food factory). Additionally, before it becomes cost efficient to build a customized culinary robot for a restaurant, it is still less expensive to design a custom culinary tool for use by the human staff that will net 99% of the efficiency of the purpose-built robot.

Barring "replicators" (a la Star Trek), or labor costs rocketing through the roof (perhaps from something like a "minimum income") I don't expect the general composition of restaurant staff to change in the next hundred or even two hundred years.

The part at 1:23 is not that impressive. It is a carefully controlled lab environment, all flat surfaces, no obstacles, with a cubical box that has carefully placed AR tags which tell of the box's exact position and orientation. Not to mention this is a video which is inherently disingenuous. It may have failed 10 times before they got it right. It isn't even using its lidar.

Now what would be more impressive would be for it to get down on it's knees and autonomously change a tire. 4/10 if it does this in a controlled environment, 8/10 if it adapts to tools/bolts being dropped, 10/10 if it does all the above outside on the side of a road.

One thing to keep in mind is that many 'menial' tasks are more complicated than they seem. Because much of the control is done subconsciously they don't seem that hard and they are harder to reverse engineer.

It has a great kinematic system. But with all those QR codes scattered around it obviously has little knowledge of its environment. I think this understanding and planning is going to be the big hurdle for Atlas style robots rather than just walking around.

Just look at the number of types of door out there and what it would take for a robot to figure out how to navigate it without having encountered the exact door before. Its going to need to figure out that it is a door. Experiment to find the unlatching mechanism. Figure out if the door swings or slides and in what direction. Then open it without hitting its self and finally make its way through. Then exporting the training data so other robots can learn it as well.

It seems that this demonstration was for a factory robot, so dealing with all types of doors is not required. Operating in a factory is much easier than being able to adapt to various settings.

I'm sure they're working on getting that integrated with Google Cloud Vision api to recognise objects without the qr codes :)

https://www.youtube.com/watch?v=eve8DkkVdhI

For anyone interested, the qr code system is AprilTags: https://april.eecs.umich.edu/wiki/index.php/AprilTags Works like QR codes but is optimized for CV tasks and works well on embedded systems.

I'm a 38 year old human and I occasionally try to push a door that should be pulled. Getting a door wrong the first time isn't holding me back that much. The same goes for a robot. The main difference is the the robot can go through the door without opening it if it really wants to.

for industrial settings, i'd imagine they'd (start to) optimize the interior design for the robots capabilities rather than for human aesthetics. Less (or different bandwidth) lighting, doors that communicate with the robot, etc.

The QR codes probably encode behavioral scripts to apply to an object that the robot encounters. So a box with one QR code may mean that the robot should attempt to pick it up even after interruption. A different QR code may mean that the robot should give up if it is interrupted when picking up the box. A different QR code may mean pick up box and move it to self. A QR code next to a door may mean that the robot should exit it automatically when it gets within a certain distance of it.

very unlikely. they were probably used for localization. for instance, locating door for egress, or computing a grasp for the box. grasp planning is particularly difficult at the moment.

I think it could have dual purposes. There was an earlier segment of the video where the robot was able to find and pick up a box and stack it on a shelf without the presence of the QR codes. There are a lot of clever uses you could put to QR codes as object and environment tagging. It would be effective at encoding what to do and perhaps what can be done with the box. I used a similar strategy to define behavior scripts for the Aldebaran NAO. I printed out some barcodes and taped them to the walls so that if the NAO was exploring an area and saw certain barcodes on the walls, it would do different things based on the codes that were present. This way, I could do simple behavior scripting without having to change the robot's programming.

I half expected the video camera to turn around and show the audience that all of the footage was shot by an Atlas model. Maybe that could be a stretch goal for the next generation after Atlas.

Do you work on Wall Street? When I watched this video, I definitely wasn't thinking about stretch goals and what they DIDN'T yet manage to do :P

Watching the video, the only thing I see is robot abuse. #freeatlas!

https://www.youtube.com/watch?v=G_P-zl8QKp0

Yes, that's kind of weird this empathy we feel toward these robots.

It's just a machine, yet it's so... human.

A human is a machine.

Yes, a human is a machine too, still I think we are more than our physical body.

Based on what? We evolved from chemicals to single celled organisms, to multi celled, to worms, to fish, to mammals, to apes, etcetera. At which point in this process did we become more than our physical body? I don't see any reason why an organism that came to be through evolution is fundamentally different than a machine that we build. In both cases it's the laws of nature at work within the universe. I think rather than concluding that we must be more than physical, a more reasonable conclusion is that there's something that we don't yet understand about the physical world, i.e. how matter in the physical world can give rise to subjective experience.

> "...can give rise to subjective experience"

Sounds to me like you're saying the physical machine gives rise to something else beyond the machine. In which case, why waste energy insisting we're only machine?

Because there is a difference between being more than physical, and the physical giving rise to the subjective experience. Being more than physical means that we are physical plus some special sauce. This would mean that a robot that we build probably lacks said special sauce, because we didn't put the special sauce in. If on the other hand it's the physical that gives rise to the subjective experience, then there is no reason to think that robots couldn't have it.

Sure, that's fine. We could then say, however, the robot is more than a machine once it has this new special feature. A special feature that may have arisen from a specific arrangement of the robot's material makeup.

Either way, "more than machine" will help these new love-enabled robots have "human rights" when they eventually do need to stand up for their rights, and wish to do so beyond their programmed directive.

Whether or not humans are "more than" machines is a vast subject of debate.

What is interesting right here, however, is the fact that watching this video many people do feel some sense of empathy for the robot.

The editing of the video deliberately and perhaps inexpertly(1) heightens that, but I think we'll be surprised (as a society) at how we "feel" when robots really start walking the Earth.

(1) The recent film "Ex Machina" explores this topic at some depth, and very expertly.

It may be the case that every process is purely chemical and physical, and it most probably is, but it is not a useful way to understand the world. Even if merely conceptual, a metaphysical side to our essence does exist. What's trancendental might indeed be an illusive physical phenomenon, but we're sort of trapped into it, even if we understand it one day, the subjective experience that is, our understanding will happen within our subjective experience, so why try to deny it? I think it's better to embrace it. Karen Armstrong has produced some very interesting reading on the topic [1], though I do not agree her on each and every point she makes.

BTW, when I say metaphysical and trancendental, I do not refer to religion and divine existence. I am a bit close to Sartre, though not completely, I'm adding this just to clarify my stance a bit more.

[1] https://en.wikipedia.org/wiki/Karen_Armstrong

I love to play with words. They often convey meanings. I'm French and "the death" translates into "la mort". When you say "la mort", you hear "l'âme hors" (the soul out), which tells you we are a soul into a body. And that's what almost all religions teach : we come from the spiritual world and we are here to experience the physical word - and not the other way around.

But yes, it's a huge philosophical/metaphysical/religious debate. We can't really address this topic here.

    > I think rather than concluding that we must be more than physical, a
    > more reasonable conclusion is that there's something that we don't
    > yet understand about the physical world, i.e. how matter in the
    > physical world can give rise to subjective experience.

What you're fumbling towards here is called the "hard problem of consciousness"[1]. Essentially some believe that there's some secret sauce in the brain that gives rise to consciousness, while others believe that there's no such "hard" problem and that "consciousness" is just what we call a bunch of simpler brain systems working in unison.

My opinion is a layman that's read quite a bit about this is that there's no "hard" problem, and the only thing that separates us from "lower" animals is just that we're running a slightly more sophisticated version of the same physical brain processes, which we like to call consciousness.

There's no reason to believe that we're anything than the cumulative behavior of our brain matter, or that we're somehow "special" in the animal kingdom in another sense than e.g. wolves are more special & sophisticated than mice.

1. https://en.wikipedia.org/wiki/Hard_problem_of_consciousness

I agree that we're not special, but the hard problem is explaining why subjective experience exists at all. What are the arrangements of physical matter that give rise to it? Here's another fun thought. Presumably if we built a human by artificially placing all the atoms in the right place, then they would have subjective experience. What if instead we simulated that on a computer? The simulated human would act the same as a real one, including expressing wonder about how it is that they have subjective experience. Would the simulated human have real subjective experience? No way to find out as far as I can tell. You can't even find out whether I have true subjective experience. Maybe I'm just a machine acting as if I do. So maybe it's a nonsensical question, but I have never found a satisfying explanation for why exactly it's a nonsensical question.

    > Would the simulated human have real subjective experience? No
    > way to find out as far as I can tell.

I think that once you hold the view that living beings aren't "magic", that our behavior / wants / needs are just a result of the sum of our hardware as it were, you'll quickly realize that talking about "real" experiences doesn't make sense.

Extrapolating from a simpler system can help to give clarity, e.g. does this Lego Robot who's running a complete simulation of a round worm's neural network feel "real" hunger when its food sensory neurons are stimulated?

https://www.youtube.com/watch?v=2_i1NKPzbjM

Some might have the knee-jerk reaction of saying something to the effect of "of course it's not real, it's just a simulation running on some robot, it's not actually hungry".

I think that falls apart when you consider the following: Suppose I had the technology to take an existing round worm and slowly replace its individual cells & copmonents with artificial equivalents, at what point would it stop having "real" feelings of hunger or anything else?

I think the only sensible answer is to realize that the question doesn't make any sense. No more than asserting that moving a computer program from one computer to another makes its execution any less "real".

    > You can't even find out whether I have true subjective
    > experience. Maybe I'm just a machine acting as if I do. So maybe
    > it's a nonsensical question, but I have never found a satisfying
    > explanation for why exactly it's a nonsensical question.

I think if we mapped your entire biology & neural network and uploaded it to the proverbial Lego Robot we could easily find out a lot of things about what's happening with your experience of the world.

E.g. maybe you're in chronic pain, maybe you're hungry. Those things are just the result of biological processes in your body & brain. There's no reason to think subjective experience is anything inherently different.

Edit: Added some clarifications.

You may not have heard about this, but extremely smart people have been arguing about this problem for thousands of years. I guarantee you it won't be solved in this thread.

This is a good starter: http://plato.stanford.edu/entries/consciousness/

Have you played The Talos Principle?

I haven't, but it looks interesting.

Sometimes when I'm stuck on the puzzles, I leave the game for a few days then come back and solve it within 5 minutes. Did the same thing solving javascript bugs sometimes.

There is a similar phenomenon at work with the Witness too. I've heard it attributed to the fact that the brain continues expending energy on unsolved problems even subconciously. Not sure if that's true or not.

Either that, or our next try is simply with a fresher more energized brain.

Or, perhaps better yet, Soma: http://store.steampowered.com/app/282140/

SOMA was great too, that ending!

Interesting that this is such a controversial comment, judging by the rapid fluctuation of the votes.

I think these emotional issues may be solved by making them look less human and more like machines.

Now if only we could apply that empathy to all humans.

At 0:55... it looked like the robot and human might hold hands!

At least the robot leaves after that. No Stockholm syndrome.

He'll be back.

As my six year old son watched one of the videos he asked "Dad, when will the robots be able to stop working for them and be free?" To say I was blown away is an understatement... that was pure human emotion and honesty as he's never seen/read any sci-fi with such questions presented to him.

Yeah, if Skynet ever takes over, I think that guy with the hockey stick is going to be Victim Zero.

It couldn't possibly take over if vengeance against guys with hockey sticks was a primary utility-function-maximizer. Perhaps in Canada, but it would be easy to defeat, having never trained against artillery and whatnot.

When the robots become self aware that guy from 1:25 has it coming.

I know this is (maybe?) a joke, but I've been seeing similar jokes everywhere. Self-aware AIs in the future will probably be much more rational that humans, utilizing game theory for every decision, and recognize that taking out vengeance on a human that had no ill intent on a non-self aware prototype would be a waste of energy and only bring about negative consequences.

I'm replying in seriousness because I too could imagine a world where this video was marked as a historical artifact by AGIs and recognized for its content.

Did you read "strategy of conflict"? It explains why irrational behaviour is the most optimal strategy sometimes.

I haven't read it, but from a quick glimpse on Amazon, it looks to be a book on game theory. Wouldn't a theory on strategy imply rational decision making? Could you expound on what you are referring to?

It is a viable strategy in many games to act so irrationally that other parties will go out of their way to accommodate you in a way that they would not otherwise.

A classic example is a game of chicken (at least, where something is on the line). Sure, it's irrational to run into the other person; but it's totally advantageous to convince the other party that you're willing to do so. At a certain point there's not much difference between a determined effort to look crazy and actual crazy behavior.

It gets even more interesting when the strategy is designed by a different actor or process than carries it out. For example, the people running a society might believe it to be advantageous for their group to look so defensive and prickly that no one will touch them; but in order to do so, they have to create a society that actually is quite irrational. Or, natural processes of evolution might favor organisms (well, genes that create organisms) that are actually quite irrational in their behavior, because that irrationality will change the behavior of other organisms.

The strategy of conflict broadly extends the game theory. It includes discussion of situations in which irrationality makes sense, and so on.

Vengeance has some game theoretically rational aspects though, in a tit-for-tat way. The threat of retaliation makes the other agents more likely to cooperate? However, not sure what the rational response would be to the problem with runaway mutual retaliation.

Think about it this way: That guy wasn't just some jerk on the street being mean to a robot. He's one of its creators, and he's knocking it over in order to make future robots better.

He'll be sentenced to hard labor at worst when the robots rise up.

The interesting thing about this seems like the opposite actually. There's something incredibly poignant about the fact that the robot will carry on attempting to pick up 'his' box forever without ever becoming angry or frustrated no matter how much the humans bully 'him'.

Also shown here, although without robots: https://www.youtube.com/watch?v=SNmJqRV7LOA&t=2m18s

How poignant is it for the HTTP server displaying you this page to continue responding to HTTP GET no matter how much the humans request it? It just doesn't have anthropomorphic shape so it will never be referred to as 'him'

Does anyone known if it's walking using zero moment point like asimo? I thought they were beyond that?

It doesn't seem to have a very natural gait. Very uncanny valley.

https://en.m.wikipedia.org/wiki/Zero_moment_point

Short answer: No, ZMP is still the name of the game +-. [0]

Slightly longer answer: No, they use derivative approaches that are basically ZMP but generally more refined/modernized, which is not to say ZMP/related are good or the 'future'. [1]

Source: [0] I can try and find some non-paywall PDF's that go over the field, but basically everyone uses ZMP except for a few research groups in Korea or UMich and groups using their design/model.

[1] I work in the industry.

This is awesome. Thanks! I kind of figured they'd apply the algorithms behind big dog et al to bipedal. But I guess bipedal is completely different.

What do you think it would take for robots to walk in a similar way to humans?

Great question.

Disclaimer: I don't work with these types of bipeds. I work on powered exoskeletons for paralyzed type people. I would google that phrase to get an idea, but what you'll find is not close to what i'm working on (Read: Wink wink). That being said, the fields overlap very heavily, probably too heavily given the necessary differences, so I do have a fairly rigorous (rusty, at this point) understanding of the topic.

You are right in a way; big dog et al (depending on specific bot/type) share alot of software features with biped robots, but not the type that deal with gait control. The difference is found primarily in gait-control/kinematics, as determined by the necessarily complex nature behind bipedal mechanics.

The future of robotics sort of already exists amongst us, at the very least in the public research domain. The fundamental issue with biped robots right now, as i see it, is:

1) Biped robots are almost universally 'human' inspired

2) Humans are theoretically fully optimized bipeds, copying them is logical and probably an unbeatable 'design'

3) Designers try to copy the human form, but either focus too heavily on ZMP influence or too much on a 'Terminator' level robustness (Heavy, rigid, and impossible).

4) The robot is produced, and because ZMP is very predictable/safe, it walks; but it is not 'human' form or very good at achieving it's design goals because its goals were modeled after the ideal biped, and this is a very bad imitation.

5) Designers say: Hey, spec sheet checks out, FEA says this is strong, problem must be bad software [ZMP]. Software guys say: This software [ZMP] can be improved. The designers are right, the software guys are right, but ZMP is wrong, and because of that, design is bad, etc...

So to really get a true, true biped robot, you really, really need A) Control theory that isn't backwards [literally] and B) A really, really 'human-like' (musculoskeletal) design.

A and B are both really difficult because both are massively challenging, and they have to work together perfectly. So you need a solution that is really creative, really beautiful and elegant and lightweight, and probably most difficult, you need to have people that know how to turn that kind of framework from theory to iterative simulation to real world etc etc.

ZMP is the only real game rn because it was first and etc [sunken cost fallacy and myopia, i think]. All other approaches are neglected fully, regretably (not many other approaches either).

Jumping back to how I see it, I think bipeds will only improve superficially till ZMP is killed or hybridized effectively, which doesn't mean A and B has to occur. Their are just too many fundamental limitations to the theory for bipeds using it to become ubiquitous, or even used beyond research/darpa-fairs.

I also think that my research/focus is more complex than robotic bipeds, and it'd be really worthless waste of my time if it was at the current or slightly above 'state-of-the-art', but this is absolutely the best way I can spend my life right now.

This whole industry [true 'wearables' and robotica] is like a giant powderkeg, and it's going to blow up way bigger and sooner than any thought possible [imo, so take with a grain of salt, but facts afford a potentially representative state of things to be].

I've also just realized i've forgotten to explain zmp/gait and many other things, but i feel that the write up stands on its own.

Genuinely curious: could you chance a guess as to why they chose bipedal over quadrupedal?

I don't know much about the field, but mechanically quadrupedals simply make more sense - a lower center of gravity combined with more opportunities (in the form of more legs) to correct blunders.

I don't think traditional ZMP would enable the stuff this robot does in the snow. More likely some sort of evolved controller, like in this paper: https://vimeo.com/79098420

This is a textbook, ZMP derivative type controller.

ZMP doesn't necessarily require absolute maps of the environment, but generally is more of a predict and optimize for safety type affair, and Atlas i'd imagine does that both very quickly, and a bit smarter leg reaction movements.

The linked video is more akin to elementary central pattern generation, which is a very understudied field, but their is more advanced research on the topic from pre 2010 if you are curious. Interest seems to have dropped off unfortunately in the really cool stuff since then, but you can find it with moderate ease.

I worked with Atlas during 2013 at WPI. Siyuan Feng just defended his thesis on Online Hierarchical Optimization for Humanoid Control for controlling Atlas during the DRC. The controller used is not as well refined as BD's controller, but I believe it stands in the same principles, and relies on ZMP.

[1] https://www.youtube.com/watch?v=QcBYhR0DbbU [2] http://www.cs.cmu.edu/~sfeng/sf_thesis.pdf

I found it interesting how lots of the top comments on YouTube (ie: a wider generalist audience) were concerned about robot abuse and the robot uprising. We may scoff, but it's a real concern of lay people.

I noticed exactly the same thing happening on reddit.

Isn't it just everyone trying to be funny with the same joke they believe will be original, but actually is every second comment? Happens a lot on Youtube. Surprised it's happening here too.

But if there really is a genuine wave of care about the robot welfare sweeping the intertubes, then perhaps we need some ethics drawn up for when the robots have feelings.

BTW, apparently fish do have feelings after all:

http://rspb.royalsocietypublishing.org/content/282/1819/2015...

I think it's even simpler than that. They don't even expect it to be an original joke. This is just the skynet meme. Memes like this get repeated ad infinitum in comment sections like youtube's. People recognize some pattern (e.g. robot abuse), pull out the most appropriate meme from their meme basket, see it also being pulled out by others, and have a sense of belonging & social reinforcement.

That's a kinder assessment. I'm not so sure though. I think many people don't read others' comments and cruise in, watch video, and bust out their joke, not noticing it already exists in various forms in previous comments.

Unlike offering a general opinion, viewpoint or judgement, we prefer our one-liner jokes to be unique.

That is super impressive. The video of the robot walking around in the snow, and multiple times almost falling over but righting itself, exactly how a human would...

Why do they always gotta smack around the robots like that :(

To show how quickly the robot can compensate to having its center of gravity shifted (balance).

But man oh man, the robots are going to nice look kindly on us later.

haha yeah. There's something slightly terrifying at 2:18 when when the robot gets pushed down, stops, then 'jumps' back up. I was half expecting it to start attacking the guy

Not gonna lie - kinda felt bad for it.

Someone on the internet said that humankind's biggest conflict will be between the two sides of those who believe in A.I. as life and those who don't.

This is kind of getting us there.

There's little doubt in my mind this technology will be employed to exterminate villages full of "terrorists" by the US military in the future. You don't even need to invoke AI for that scenario, just VR and maybe whatever Google puts into their autonomous cars.

There's no need to exterminate anything. When you have robotic infantry deployed at theater scale, a largely non-lethal peacekeeping approach could be taken with relative ease.

I'd argue robotic infantry would be a massive leap towards improving the security and humanitarian situations in some of the world's most terrible places.

Unlike long-term deployment of human soldiers, political will wouldn't really factor in to robotic deployments. That would translate to less genocide and suffering in the world.

At one point at least, people surmised that atomic weapons would lead to greater success in peace-keeping throughout the world. Not so sure it's turning out that way.

>Unlike long-term deployment of human soldiers, political will wouldn't really factor in to robotic deployments.

Aren't all military deployments an expression of political will, though?

I imagine videos like these being used in future trials to show humans torturing robots.

The Animatrix depicts a lot of robot vs human strife. Pretty good shorts :)

>I imagine videos like these being used in future trials to show humans torturing robots.

The proceedings may resemble this:

https://i.imgur.com/1m1OPlb.png

All I know is a great prank would be to program Atlas; after being pushed to jump around and start chasing that guy, yelling "Come here squishy human, I want to test your cranium for compressibility"

After a few times where the guy with the hockey stick smacked down the box and moved it away, I half expected Atlas to look at the box, then look at the stick, and grab the stick and attack the guy with it.

I was imagining something like that, although with the robot brandishing a laser gun. The robot would then saying, "die human." in a robotic sounding voice.

That will probably be in the next software version. One thing at a time.

They don't seem to be using end-to-end deep reinforcement learning. I think the robot movements will become smooth and organic once they replace the current specialized approach with more general algorithms.

Robotics is a core application of reinforcement learning as it deals with an embedded agent in a feedback look with the exterior. That's why DeepBrain chose games to test their algorithm, because in a game the player is embedded in an artificial world which is much easier to use in the trial and error learning phase. Logically, after mastering Atari games, at some point they should focus on robots that function in the real world. It helps that Deep Brain and Boston Dynamics are owned by the same company.

Note the quiet compared to Big Dog... you can hear birds sing in the background. That's a big deal. Their 4-legged pack mule for DoD was killed partly due to noise.

Enjoy the silence..?

as best i recall, big dog was meant to be a long distance, possibly autonomous, pack mule for places and missions where a manned vehicle was not an option.

i suspect that this robot have a operating time of maybe an hour before the battery pack needs to be swapped.

Pretty impressive. Seems like a non-stiff torso could help a bit. I'm sure they are investigating that option/have tried it.

Hockey stick guy genuinely made me laugh. They should call this the hockey stick test (once again very impressive).

Not sure what impresses me more overall the walking on non trivial terrain in the beginning or the adaption to the "hockey attack". The video certainly makes me want to work in robotics. Haas Bioroid is here ;)

Very awesome, each generation is just so much better.

I wish they had shown it climbing those stairs though! :)

As I've watched these videos over the years, it's been impossible, especially now, for me to not see how these and other such robots will replace and eventually extinct all humans.

In theory, a family of 4 humans could run a global economy run by robots and no humans. That's a sort of ridiculous, maximal state, but what makes any of us think that, given a continuum from the current state to that maximal state, we won't be somewhere closer to that state in 30 years, especially when considering the unreasonably effectiveness of rent seeking on technology-driven monopolies (e.g., Amazon)?

Anyone read "The Industries of the Future" by Alec Ross? He talks about the technologies of the future, the technologies who will become an integral part of our lives in the next 20 years.

The first episode in the book is about robotics. I really urge everyone to read that book.

This Boston Dynamics video gives me serious Terminator vibes. In reality, from an Economics stand point, I'm really curious and afraid to discover how will this affect the lives of millions and millions of workers world-wide who will lose their jobs to Robot. How can they stand at the right side of a technology revolution they know nothing about?

What do you think?

Bring it on. Same thing happened with industrial revolution didn't it? Trouble, followed by adjustment.

Leave the nasty jobs to robots. New industries will pop up. Technology keeps growing.

There will be protests. Robots will be harmed. Science fiction has told us most of what is coming.

A related article about Dr Moshe Vardi and his controversial view that governments and societies around the world may have to consider a "basic income guarantee"

http://www.independent.co.uk/life-style/gadgets-and-tech/new...

What percentage of the body weight is batteries? HOw long can it operate ?

On thing you can already bet on is that in 20 years that guy is going to wake up in the middle of the night to find a T1000 with a hockey stick in it's hands and a grin on it's facial expression monitor pointing at a box and say "would you kindly; pick it up?". Later his tomb stone will say "Don't be a dick to robots".

It's funny how this video is what everyone in IT thinks work will be like before they get their first job. Meanwhile, this is what work is actually like at every IT job:

https://www.youtube.com/watch?v=FIIWX80KBUE

For those not wanting to watch a video, Atlas appears to be a humanoid robot: http://www.engadget.com/2016/02/23/boston-dynamics-presents-...

What's crazy to me is that there's something "organic" enough about the way these machines are behaving, that I actually feel bad for the robot when the guy knocks over the boxes.

I think I wouldn't feel that bad if this were a self-driving forklift.

And it's not even an uncanny valley thing, or maybe it is?

As a futurologist I don't like this one bit. Seriously, I feel like this company will destroy us all.

It's not the Terminator T-800 yet, still it's impressing. :-)))

Does anyone else start to think of Blinky[1] when the guy starts pushing around the robot?

[1] http://youtu.be/P0lKDy6E918

Atlas is not alone: http://www.theroboticschallenge.org/teams

Impressive. They seem to have reduced how much noise it makes during normal operation. Does anybody know how long will the battery last on the current version?

#robotsarepeopletoo #botrights

I empathized when the robot faceplanted...

Does someone know if they are using ROS?

It doesn't have the grace of an actual human.

Its solution to any problem seems to be more stomping.

Although a great incremental update, is the new Atlas any better than Asimo?

it can use a shiny Bender skin :) hope some Futurama fans are working there

The 600 series had rubber skin. We spotted them easy.

that's terrifying!

Player Piano, the first novel of Kurt Vonnegut, was published in 1952. It depicts a dystopia of automation, describing the deterioration it can cause to quality of life. The story takes place in a near-future society that is almost totally mechanized, eliminating the need for human laborers. This widespread mechanization creates conflict between the wealthy upper class—the engineers and managers who keep society running—and the lower class, whose skills and purpose in society have been replaced by machines. The book uses irony and sentimentality, which were to become hallmarks developed further in Vonnegut's later works.

https://en.wikipedia.org/wiki/Player_Piano_(novel)

and then they will put a weapon on it

Poor robot, #abusedinboston

Am I the only one that saw this and got excited thinking it was a video of FB's new ad platform?

Reply on news.ycombinator.com