MIT cheetah robot lands the running jump [video]
youtube.com
youtube.com
http://www.nature.com/news/robotics-ethics-of-artificial-int...
He calls for a moritorium until we can get a neural net to learn Asimov's 1st Law.
Should the self driving car swerve into you to avoid greater moral hazard ?
Humour aside as researchers and robotocists do we have a moral obligation to obey the 1st Law ?
The self-driving car's first priority should be to protect those inside it. Anything else is more complicated. I read an article a little while ago about this topic and it posed a question along the lines of "should a self-driving car do the thing that'll kill its own passengers, yet save a larger number of others around it?" Think about how difficult it would be for a self-driving car to estimate this. Even more importantly, how would two self-driving cars cooperate in the event of an inevitable accident? Would we allow them to directly talk to each other, or are they only allowed to interpret each others' trajectories?
I think the simplest directive is "save your own passengers". That is what human drivers currently do, and I believe in the long run this would be effective at saving the optimal number of lives. Besides, if choosing between a car with this directive and a car that's programmed to sacrifice you to save someone else, which would you buy?
What are we going to do with all the extra parking spaces? Turn them into parks/wildlife habitat/green space!
> Think about how difficult it would be for a self-driving car to estimate this.
For sure you're right in the near term. But assuming we achieve super-intelligent AI that we manage to corral into a benevolent relationship with us, perhaps difficulty wouldn't be an issue at that point.
> Even more importantly, how would two self-driving cars cooperate in the event of an inevitable accident? Would we allow them to directly talk to each other, or are they only allowed to interpret each others' trajectories?
Yeah, this is getting good. Assuming we have those super-intelligent machines, those cars should be able to share pretty much all relevant data before a collision. So how could the car determine the greater good option? Is one life equal to another? What if one is rich and one is poor? What if one is the president? How could they even collaborate if they don't share the same basis for valuing human life? Could car companies realistically have diversity in those algorithms, or wouldn't that just lead to lawsuits against cars that were overly aggressive or passive? Nobody wants an unsafe car, so would car companies have to compete in their creation of aggressive algorithms? Or would it be a factor of cost... that you need to pay more for a car that guards its occupants more strongly... leading to more and more aggressive cars as they compete for the crown? Honestly, this line of thinking leads me to think we'll eventually need these values to be dictated and homogenized... requiring a strong central government, and aggressive detection and punishment of aberration. Or else a mad max style zero sum game.
Forgive the digression into fantasyland, I just found the ideas and questions here to be inspiringly interesting.
However, your argument omits a couple of interesting details. First off, if a car can have uncontrolled acceleration, it can likely have any number of glitches. It may for example think it's accelerating uncontrollably, and try to throw me off a cliff. Or it may think that the cliff is a small ditch. Or it may think that a bunch of balloons tied to a mail box is a crowd. If you are a programmer, what would you rather debug: code that prevents uncontrolled acceleration or code responsible for killing the driver by recognizing crowds and cliffs?
The other details is whether the metaphorical crowd should even allow me to buy a car that has logic built into it to drive over them. What if instead of a cliff and a crowd it's actually two different crowds of different sizes? What would a human driver do in these cases?
I haven't tried to figure out what the numbers are, but my gut expectation is that vehicles driving into crowds is so rare that there aren't really statistics about it.
As a developer I wouldn't buy a car that would drive off the cliff.
It's a tough problem. Personally I would make a quickly/poorly thought out analysis of how to do the least damage. Being human I'd probably get it wrong.
What does being a developer have anything to do with it?
Non-developers would assume the computer to be (somewhat) infallible.
What if you're in your car with your family?
The best answer is of course: never let the car travel so fast this scenario happens - but given the possibility of an oil spill around the bend -- that might be very slow indeed.
Also, you can enforce cars with directives to save "the most" lives with policy and law. Of course people are going to buy the crowd-plowing car if it saves their own asses. That's why you make it difficult or impossible to buy a car that will do this, instead only authorizing cars that make the better choice to drive the streets.
If you take that dictum, then any action which requires activation energy (e.g. steering, or accelerating to avoid something) is unavailable, and as a result, the vast majority of these ethical decisions are completely moot.
Consider the worst things that governments will do to get an advantage. I mean the downright most inhumane things they have actually done.
Now given this, what chance is there that they will respect a moratorium on research. If violating basic human rights for a small advantage is part of their MO, then it would be foolish to think that a lesser breach of ethical obligations for a greater advantage would not happen.
All we would achieve by a moratorium is a slight slow down in how soon governments access this technology while locking the technology away from access by the general public.
That said, the ability for government militaries to become autonomous and not made up of individuals who have at least some chance of resisting the worst orders is a very scary notion.
We may be approaching the next Great Filter.
The good butler robots had the 3 Laws.
The dreamer hero robot had no laws but had imagination and developed its own morality.
It was demonstrated (via flashback) that the older robots were capable of allowing a human to die, while ignoring a direct order from a human, and might have been able to rationalize law 0. They weren't remotely linked to VIKI, making them harder to coordinate, so it was easier to remove them from the equation so they didn't get in the way.
edit: if you're going to downvote please respond
This means the developers just enabled the functionality, and it was the operator that chose to be selfish.
That said, none of that deals with the difficulty of making time-critical decisions. For that, you should probably read my thesis...
Creating life is what we earthlings do. And yeah, maybe that life form you create goes on to do horrible things. But maybe it goes on to do really good things. You don't know until you try.
Look at what's happening in Yemen. The population there lives under the fear of drone strikes by the US (that includes the civilian population). If governments are prepared to do that with drones, what reason is there to suspect they'll turn down using something that's ground-based to impose control?
Aren't they better off than the Vietnamese who lived under the fear of Napalm bombardment? (or, for that matter, Japanese fearing the atomic bomb) I don't think that robotics has significant negative impact on warfare (and I wonder whether it has a positive impact).
Those are (pretty clearly, IMO) war crimes.
But still, it's useful to be worried about the potential "desensitising effect" of remote warfare. Is someone operating a drone subject to the same psychological pressure to avoid killing other humans as someone pulling a trigger? (Turns out that yes, that person probably is, but that their political superiors possibly aren't).
And the people who make decisions don't really seem to care at all about this sort of killing except insofar as it creates backlash or has some other operational implication, which means the mode of killing doesn't really matter.
In terms of their superiors, isn't that how warfare has always been?
My guess would be that ground troops replaced by robots would likely still have a pilot with a "finger on the trigger" for a long time. If the reaction is this mixed Hacker News, how do you think extremely conservative military commanders would feel about "AI" controlled soldiers?
Incidentally, air traffic control had this same problem in the 90s until too many people started taking the quick way off the tower.
Not really, instead they'll live in fear of Robot bombardment. Which is already happening today in (for example) Pakistan with US-directed drone strikes against civilians.
I'm not wild about drone strikes but the scale of destruction isn't remotely comparable. I mean, just look at the amount of bombing we carried out in Laos over a similar period, even though we weren't even at war with that country: http://peterslarson.com/2010/12/15/us-bombings-in-laos-1965-...
People who get killed by a drone strike are just as dead, whether they are legitimate military targets or unlucky innocents, and likewise the suffering for people who are injured is just as dismal as from other kinds of attacks. But the scales involved are very different and we shouldn't overlook that.
I think we have that, they're called soldiers.
Sure the first generation won't replace all soldiers and they may never replace them all. We will need logistics and support people of course. Its not a perfect solution. But the second it becomes cheaper to put a robot in the place of a human on the battlefield.
They will.
But then why does it kill at all? Presumably because it's been sent to further a human agenda.
If you imagine a noble military purpose - liberating an oppressed population from an aggressive occupying enemy, for example - then your robot soldiers are awesome; they will target only combatants; minimize collateral civilian casualties; they will never loot or rape; they will selflessly interpose themselves between the innocent and those who would harm them. The perfect heroic soldier, better than any human army could be. They will be greeted as liberators.
But to the extent that the underlying human agenda involves pacifying a civilian population, instilling fear, or outright causing terror, there's no reason to think that a robot soldier would not be capable of being far worse than human soldiers. It can't be reasoned with. It doesn't have a conscience. It doesn't matter that it doesn't 'fear for its life' when kids throw stones at it if it's been programmed to respond to that threat with deadly force precisely to discourage other kids from throwing stones.
At the point where we take humans out of the "in-the-moment" decision process, a lot of thorny issues about what's acceptable in specific situations can become less ambiguous. Agreeing on rules is much easier than agreeing on what's acceptable behavior for a person in every situation, because a lot of that depends on state of mind.
You do not have images of flag covered coffins returning home. You have no insubordination. You have the possibility to do economic promotion without being labeled as doing such (regarding international treaties (at least here in the EU)). You do not need all these training facilities.
Just to name a few.
So they will even do it, if it is not cheaper to do (imho). And logistics can be automated as well (delivery by drones, something like fueling done by other drones, and so on).
I believe we will see drone-carriers analog to aircraft-carriers being at least semi autonomous as well within our lifetime.
It's hard to imagine that people will care much if all they see is a pile of scrap metal being shipped home.
It'll be extremely difficult for a government to use robotic infantry whilst having public support for such an action.
Rather than appealing to a sense of "brotherhood/camaraderie against a common foe" in order to support a campaign - propogandists will likely have to exploit the peoples sense of fear to garner support. This is a shift we have already started to see with the "war on terror".
Central African Republic has war with child soldiers using machetes and guns. It's an unpleasant truth: humans are disposable.
This is evidenced by a bunch of stuff - rich westerners have their clothes made by poor people locked into unsafe Bangladeshi warehouses; have many of our goods made by low people in terrible and dangerous conditions; by our lack of interest in garbage pickers or street children or etc.
Consider the initial training expenses, equipment, career-appropriate salary along with possibly reenlistment bonuses, along with (in the worst case) medevac, multiple complex surgeries, and lifelong treatment/disability pension for some injury. Not to mention the logistic costs of transporting them to various places around hte world, and maintaining an acceptable standard of living there.
Per-machine costs would be high, but there's comparatively little training. Maintenance would be expensive, but rehab/repair decisions would be much less politically charged. Not sure how logistics costs would compare - "life-support" is probably lower, but sourcing power & spares might be greater.
I wonder if anyone has done an economic analysis on this sort of thing.
- the think tank "Center for a New American Security", which seems to have some influence on politics and runs studies such as
http://www.cnas.org/research/us-defense-policy-and-military-...
https://en.wikipedia.org/wiki/Center_for_a_New_American_Secu...
- this handbook, which seems to be published by a combination of US and French Army researchers:
http://usacac.army.mil/CAC2/cgsc/carl/download/csipubs/Frenc...
(do you need reading material? http://usacac.army.mil/organizations/lde/csi/pubs - wow)
- This Australian publication discusses costs in an abstract way:
http://www.army.gov.au/Our-future/Publications/Australian-Ar...
All of these deal with robotics and the military. Ethics is discussed widely, but visions and plans are also given (for which a cost analysis should be necessary, but I haven't found details).
Concluding, the current vision of the developed world's military seems to be to replace some humans with robots and have the rest work alongside them, giving them the most dangerous tasks. (A publication lamenting the fact that army operations can be life-threatening is somewhat ironic, by the way.) However, the previous points cited child soldiers, which are neither trained well nor do they receive the benefits you mention. So we're dealing with different contexts, and both statements make sense, IMHO.
I highly recommend this TED talk about how the future of war could return us a feudal world.
http://www.ted.com/talks/daniel_suarez_the_kill_decision_sho...
At the same time, this progress cannot (and shouldn't) be stopped, but we need to quickly 'evolve' our moral and ethical standards in order to accommodate for these new powerful tools. Just like giving a loaded Ak-47 to a monkey is a bad idea, making this tech freely available to everyone is similarly stupid.
And that's because a lot of monkeys are wiser than a lot of humans. Give those humans a button which controls an army of robot killers and they'll press it just to see what happens.
The moral issues in this are no different to those of the use of drones, which are very well discussed.
Unless they invent the Arc Reactor to along with it, those are a long way from the battlefield.
Creating something with our knowledge and the ability to learn ever faster than us is very different.
I think this is an optimistic view of human nature.
I suppose these animal robots could help with same-day deliveries.
This research is funded by DARPA. Not because of its potential to help the elderly, mind you.
You might want to ask those species how that worked out for them. Also, consider how slow the change use to be compared to the life expectancy of the species.
to humans, cats are awesome. to mice, cats are monsters.
ai and killer robits have a strong potential to be monsters.
If it was more rounded and animalish, maybe with a fur coloured coat of paint, I think it'd be a lot less scary.
Longish fur might actually work quite well for those 2 purposes, they're not entirely dissimilar from the requirements imposed on real animals who evolved it.
It would quite possibly make itself rich, and probably morph itself into a distributed system running nodes on millions of compromised machines (+ the datacenters it buys). From here, AI could possibly buy itself access to things like weapons platforms, but ultimately this is contingent on humans selling them.
If a lab has a robot capable of committing violence, and an AI algorithm and it wants to put them together, then yes running the algorithm locally on the robot is probably a bad move. But you could proxy between the AI and the actuators, with the ability to override movements you don't like, for example.
The more likely scenario is that these things will be used as instruments of war, running dumb/normal algorithms and controlled by commanders in the field. In that case, just another one of the many creative ways we have to kill each other.
As for a generalized fear of the emerging technology, that fear is simply your body's way of telling you that it is important that you get it right (other people will feel it too). Same is true for the advancement in laser diodes for example, or recombinant DNA, or lithium ion battery packs. Growing up is scary, but you can't stop it from happening, you have to mature as it goes along or die.
As to a robot army + an evolved computer intelligence... that could be scary. It's easy to imagine something like remove anyone that isn't a "pure" X... where X is a combination of racial and religious dogma. In practice, and with practical limitations on resources, it's hard to imagine such an army that isn't outmatched (at least in numbers) by armed people (lets not weaken the 2nd amendment too much).
Now, when i drive i look further down the road, i plan out how to follow the road, i see the intersection coming up, and try to anticipate the routes of the other cars/bikes and that requires far less energy, than only processing the immediate surroundings.
I've noticed it in cases where I'm a passenger in a friends car also, but I've always chalked that up to the conversation being a distraction.
Alternatively, would it be better if the robots also ran more life-like, or is there a benefit (besides the ease and simplicity of engineering the physics) to the robots running like this? I.e. will they ultimately have the robots running more life-like?
https://www.youtube.com/watch?v=wE3fmFTtP9g
Which is to say, it looks more natural and more balanced with the legs out of sync.
Is it a random result because of the phase of the obstacle in the run cycle?
I am so impatient though - I want to scatter bricks on the landing-side of the jump to see how it copes.
That said, for some reason I am still bothered by the use of laser range finders. To me it makes it all feel like a parlor trick. The easiest way to produce machines that can be used in "Oh, wow! Look at that!" videos that continue to bring in grants.
These robots should use binocular vision and nothing else.
Oh, wait a minute, that's hard, isn't it? Yup.
If you are building a robot for a factory, put limit switches, magnetic sensors or whatever you want on it.
If, on the other hand, you want to build robots to live, work and interact with humans they need to be capable of understanding my world the way I do. Think of a robot interacting with a toddler or a bunch of kids.
This is monumentally harder than scanning in front of the robot with a laser to detect geometry, measure height and approach speed and then plan a jump. Much harder.
Not to diminish their work but the math and physics seem almost trivial. Figure out the x-intercept, width and height of a parabolic path that will give you enough margin of error not to touch the obstacle. Then do the math on the time delay between the front and rear legs based on approach speed. Then plan the gait in order to be able to have the legs at the right point at the right time. So long as you have a robot that can jump it's a done deal.
Again, I know it is more complicated than that, but it doesn't compare to the degree of sophistication a robot would have to have to manage the real world with binocular vision. My 9 year old kid can fly remote controlled model airplanes tooling along at over 60 miles per hour just using his eyes. You don't need millimeter accuracy laser measurement devices, you need to understand the world around you in some context.
Context: I built walking robots (not toys, research grade) over 25 years ago. Today there's virtually no difference in actuation mechanisms and sensors. A lot of these programs are grant-sucking machines that are reinventing the wheel rather than making true progress. Here are some of the things we need:
- True binocular vision systems that can develop an understanding of the environment to various degrees of sophistication (this is hard)
- Better actuators. The artificial muscle has yet to be realized. Make your arm limp on the table. You can't do that with a robot. You can simulate it. But it isn't the same thing. You'll actually consume power and spin gears/pumps very fast to be in "limp and compliant" mode. Real flexibility and real compliance are critical for robots that need to interact with people and animals. Every animal on the planet relies on this to interact with the environment.
- Better programming paradigms. We are still typing "if" statements and "for" loops to program intelligent robots. A far greater degree of abstraction is required to truly advance the art. No, libraries are not a solution. We need to be able to express concepts to a machine in far more efficient terms. How do you teach a robot to tie a knot on a rope on a table and have that robot think about using that same knot on a sailboat or to restrain a dog to a post? Without telling it that these are options?
- Better means of communication with machines. Buttons and knobs isn't how you communicate with your taxi driver or housekeeper. A 5 year old kid should be able to command a machine without having to rock the Linux command line.
etc.
I guess my argument is that we already know how to build "parlor trick" machines. We've known how to do this for quite some time. Any set of decent mechanical engineers can build a decent walking machine given a reasonable amount of time. Making it walk and even jump is almost just as trivial.
Because of the way these departments are funded the truly hard and interesting work might not be done or might not see the same degree of funding. Some of the items I listed above could require 10 to 20 years of solid dedication before the "Oh Wow! Did you see that!" moment is reached. Most of the funding out there isn't smart enough to support these kinds of projects. And so the money goes to the guy who can put a spring-loaded plunger on a little robot with wheels and show it can jump to the top of a building. You know, first year college physics, if not high school physics. This does virtually zero to advance robotics but it sure makes politicians write checks!
One good reason would be to prefer passive sensing, because LIDAR is equivalent to waving a laser across the entire landscape, precisely announcing your own position to anyone watching. Multi-ocular systems can be entirely passive assuming enough ambient light, so they're much harder to detect.
1. Real cheetahs don't run like that. The action of both pairs of legs is staggered for smoother motion and control. Check this out - https://www.youtube.com/watch?v=131wvVGjZUc
2. Looking at this jumping-over-obstacles video, it seems like the robot is lacking flexibility. The real life cheetah's hind legs go quite a bit underneath in preparation for push off. Contrast this with how the robot's hind legs are a little too behind and the robot stutters a bit after every jump. If the robot could put its hind legs further it would make for a smoother landing.
2. How did you come to this conclusion? Have you built a robot that mimics an animal? If so, may you provide some insight into your discoveries?
Robotics programmer/builder here. Always interested in learning about the work of others.
For my conclusion in point two, my insight comes from experience of copying animal movement. When I trained parkour my friends and I would copy animal movement from across the whole animal kingdom[1]. Here's an example of quadrupedal movement[2]. A big problem for us bipeds in trying to run like quadrupeds is that our legs tend to be big, long and inflexible, while our arms tend to be much weaker than our legs. The effect of this combination is to make running properly like a an animal which evolved to run on all fours quite difficult. From my own experience, putting my legs more and more forward past my arms (i.e. copying the cheetah), made for a much easier time.
If you were to try to run like the robot, with symmetrical action, it would be extremely uncomfortable, because the impact is large at every step. Staggering the action allows for a much smoother run. Make any human run like that and they'll eventually automatically switch to the less impactful staggered run.
There's a big difference though between learning valuable lessons & taking cues from natural bio-mechanics, and trying to emulate them.
Watching some of the cheetah videos, I wonder how much of their staggering is compensation. That is, the "ideal" cheetah gait may actually be non-staggered. But as the cheetah is changing direction, dealing with slightly uneven ground and reaching out for prey, it may need to stagger its legs in order to compensate.
When it comes to cheetahs, I bet a million bucks that it has everything to do with impact and smoothness rather than a compensation for changing direction and such. My confidence in this comes from my own experience imitating animal movement in parkour. For an example of one person running in a staggered way - https://www.youtube.com/watch?v=W8oh5Xuy7NA. Running like that is MUCH easier on the joints than running with a symmetrical action. If you make a human gallop in a symmetrical way they'd quickly give it up for staggered action, regardless of surface or directions.
Not sure why they chose a pronk gait, except that they're doing all straight line work. A gallop has more turn options, but roll control is easier for the pronk. Hopping over an obstacle has been done before, by some of the early MIT planar bipeds.[2]
It took about $125 million in DoD funding to get a usable Big Dog and Atlas. It's quite possible to do a lot more in legged robotics, but it is not yet cheap. It's still not clear if Google will achieve commercial legged robots, or decide to "put more wood behind fewer arrows" again and focus on their core business, advertising. The commercial payoff is a long way ahead.
Academic projects rarely have the funding and staff to get beyond the demo level. You need some big assets, like machine shops and skilled machinists. Note that they're borrowing MIT's gym. To get real work done, you need full time access to your own test area. It's hard to get that on a college campus. The typical academic robotics project is one professor and three grad students.
[1] https://www.youtube.com/watch?v=chPanW0QWhA [2] https://www.youtube.com/watch?v=XFXj81mvInc
It's just hard to beat the density and flexibility of a living mammal with a bunch of servos, chips, and batteries. We probably never will.
/sarcasm
But seriously... impressive work :-)
If Moore's law will apply to the progress of Robotics, we have very few years left before we will all be slaves.
- with deep neural nets making ai advances in strides and
- with robots/drones learning to autonomously navigate
we should start including skynet clauses into our source codes immediately!
Irony off.
https://static.nationalpriorities.org/images/charts/2015/tot...
See that Science heading there on the chart? It's hard making the world a better place on a 0.7% budget. DARPA, however has access to that Military heading, 16.3%.
Not that that is ever going to happen.
If you were running a political party, the most logical assumption to make (in the absence of richer feedback) was that the people that voted for you did so because they liked your manifesto and/or believed you'd be more competent than your competition.
With that in mind, an election doesn't really allow any meaningful influence on policy, it has to come in a different form (direct democracy, for example).
I'm not a US citizen, but from what I've read the laws on direct democracy are set at the state level, is that correct? Are there any mechanisms in place to propose policy changes at the federal level (aside from lobby groups in Washington)?
In many states, there is an opportunity for direct democracy to have an effect. Specific initiatives are put on the ballot. In practice, though, specific ballot initiatives are still implemented by the same politicians that tend to bend things to their will anyway.
Seriously.
There's still ways to go.. I still need to finish up my BSc, get a PhD in EE, and finally gain solid experience abroad :)
By developing our own tech, we can avoid depending on foreign suppliers, at least in some areas. This will help with improving national security and decreasing the obscenely high maintenance costs imposed by these suppliers in the long run.
Advancement in military technology will then allow us to apply the lessons learned, supply chains developed, factories built, and so on, to the civilian space. Look at Israel for example. They have an entire startup scene growing around the military. Many of the people who work in the military go on to start their own defense companies. These companies then give back to both military and civilian tech (Microsoft Kinect, for instance).
A crap gif and half a paragraph add nothing to the video imho.
I just watched Ex-Machina and for some reason, I just pictured the robot running out of the building after its last test without the safety harness and setting itself free.
Be nicer my friend. We are all friends here.
It's everybody's responsibility to ensure the quality of the discussion remains high.
I will celebrate AI Day. The day we were surpassed by our creation, in every way.
What a privilege it is to be subjugated by nature's sentient, motile organisms! My only regret is that these magnificant creatures don't sting me more often, but alas, it is the right of any autonomous insect to do whatever it wishes with it's sting, and who am I, but a lowly mammal, to coerce its gifts?
Honestly, I do not understand what your point is. Could you explain? (With less attitude?)
Not to worry, the drone flying ahead will already have mapped the terrain :)
i.e., parity with real life.
The contentious nature of consciousness is in the "makes the decision" part of your statement, suicide is pretty mundane as far as decisions go.