Slashdot Mirror
NASA Boosts AI For Planetary Rovers
transcendent writes "According to Space Daily, NASA is working on increasing the ability of future rover's AI. From the article: 'It now takes the human-robot teams on two worlds several days to achieve each of many individual objectives... A robot equipped with AI, on the other hand, could make an evaluation on the spot, achieve its mission faster and explore more'. Sounds like a good idea, but the article continues, 'Today's technology can make a rover as smart as a cockroach, but the problem is it's an unproven technology'. Another article about autonomous rovers being developed by Carnegie Mellon University is here."
Hopefully we'll be more successful in this years Grand Challenge, as I'm sure that could help with NASA's whole plan.
The greatest experience we can have is the mysterious.
- Albert Einstein
http://efil.blogspot.com/
While the parent was modded funny, this could be a serious issue with so-called AI. Imagine a tiny unforseen deviation between the expected result of the robots reactions, and its actual reactions to an alien environment. This could cuase the rover to do any number of unwanted things. The truth of the matter is that no matter how much the AI is tested on earth, the whole point of the rover is to explore an alien world, and in doing so we don't know exactly what the rover will find, so to let a multi-million/billon dollar rover make its own decisions on how to handle a situation could cause some serious problems. For instance image a rover missinterprets a hole in the ground as a shadow from a rock and considers it safe to drive through, you can kiss the rover goodbye. I think for now, having an actual human interpreting what the rover see's before it moves makes a lot more sense when we are dealing with such great distances and costs.
-kaplanfx
Visualize Whirled Peas
I was trying to explain this to someone the other day with an analogy.
Everyone hates having to click 15 times through a website to get the content they want. Ideally the number of clicks would be minimal.
This is what NASA has to deal with... waiting, moving the rover, more waiting, getting it to focus on something, waiting, close up, waiting, drill it, more waiting, analyse it, even more waiting...
If it could do all this autonomously! well...
Admittedly, a cockroach has a rather low amount of brain power. The computers we're typing on right now probably beats out the cockroaches brain power by a factor of at least ten thousand. But, this can be debated either way due to no possible way to measure it.
Although a cockroach has a good deal of power in it's rather small 'head', it is not by any means smart. It doesn't have any real goals, and accomplishes nothing but procreation in its lifetime. A Cockroach can't solve derivative equations, figure out the trajectory of a bullet, or anything.
What makes the cockroach the same as the rover? It doesn't multitask. It focuses on one thing, and once that is completed it focuses on another. Even though they both have limited processing power (even though the difference is staggering, I stand by my point), they accomplish their 'goal' (I use goal lightly in the case of the cockroach, as it is merely instinctual).
So, we know the cockroach has much less processing power then a rover, it's non sentient, it's autonomous, and has instincts.
The Rover has processing power, it's non sentient, manually operated, and has no instincts.
Even if the flaws on the rover are fixed, a viable AI is in place, it still can't figure out to do on it's own. In our lifetime, the chances of us making an AI that will rival even a cockroaches complexity are very, VERY low. It's had 250 Million Years to evolve. Give us 250 Million Years, then we'll start talking.
NASA's mars exploration rover has just climbed a mountain to take a photo, read the article about the tough 3km climb, including making decisions about how to cope with 'injuries'... do you think AI is up to dealing with challanges like that yet?
UK Laptops
"Creating strong AI software is a very exciting and challenging problem, and it inspires us and our students to work on this bold effort," said noted artificial intelligence expert professor Milind Tambe of the University of California, Los Angeles, who has worked with Rajan."
I very much doubt that they are talking about strong AI there. ( Arguments for Strong AI).
I rather believe he is more on the weak side.
But, well, he is a noted expert.
CC.
def. The two main varieties of AI are called "strong" and "weak". Strong AI argues that it is possible that one day a computer will be invented which can be called a mind in the fullest sense of the word. In other words, it can think, reason, imagine, etc., and do all the things that we currently associate with the human brain. Weak AI, on the other hand, argues that computers can only appear to think and are not actually conscious in the same way as human brains are.
loc. cit.
TaijiQuan (Huang, 5 loosenings)
This potential situation remembers me about the story of the creation of the huge battle in Lord of the Rings, Return of the King. Each creature was programmed with basic independent AI realistic reactions and an unexpected problem aroused. Each time there was to much battle in one area the virtual creatures where fleeing to security and the result was a bunch of cowards avoiding fight... They corrected the problem by making them dumber. :)
Yahh, hiii haaaaa! -Major Kong, from Dr. Strangelove
Based on the bits of research I've followed on various robot types (IANAE), it seems to me that analog robots in swarms might very well be the way to deal with those "unwanted things" like loss of signal from home, as well as keep costs down. How about a fairly complex base station with 1K cheap analog robots as explorers which have only digital "brains" enough to perform basic tasks based on their series/design, report findings, and "stay alive". (yeah I know I'm moving into digital/analog hybrid, but I say use each tech where it is best suited and damn the definitions)
Maybe I've just been overly impressed with videos of analog robots being crushed/tortured/dismembered and continuing to find a method to operate or "live". Hell, to me they look like a more viable "AI" than most other attempts to mimic life, especially the Eliza/chatbot variety, and some even have a better personality.
Jonah Hex
Horror & SciFi Erotic Nudes
I do a Ph.D. in an AI-related field at the moment...
I also work in a related field, autonomous systems...
First, the "I" in AI really shouldn't be there. When people talk a difficult decision problem (e.g. some pattern recognition problem), there comes the point where somebody will say, with a solemn voice: "So, what if we use Neural Networks?"
I think there has been (was) a view that neural nets were the solution, that's obviously not the case, but they've been over used and there is a backlash in the community against them right now. Basically, they've gone out of fashion. However, they can come in very handy at times and I've solved several otherwise very complex problems by using them, that would have otherwise been computationally expensive.
If there is a hard AI solution, which of course is arguable, neural networks will be part of it.
When you put something in, you merely evaluate a rather simple nonlinear function. There is no intelligence involved!
Well that really depends on how you look at it, how did training take place? What is intelligence? You're vastly simplifying the arguements here, perhaps intentionally? I'm sure the hard AI faction would argue that we (human beings) are just a sum of a great number of simple nonlinear functions, out of which there is emerging complexity.
I don't know whether I agree, but the arguement can't simply be dismissed by waving you arms in the air and saying "non-linear functions". Which isn't to say I entirely disagree that this is a (possibly) effective counter-arguement, it's just (as it stands) intellectually weak. I think I'll track down my PhD student, it's almost morning coffee time, he's probably about by now, and see if he can argue his way out of this one to my satisfaction..
Al.The Daily ACK - Eclectic posts by yet another hacker
You obviously don't know much about biological neural nets. There's a whole host of complicating factors, such as the multitude of chemical signaling mechanisms, most of which are poorly understood. It's not at all clear that simulated neural nets are a model of the brain, any more than electrical networks are (in general) models of computers.
The fact that NASA is "boosting" AI does not mean that they're going to be doing rovers that can evaluate themselves whether a dark spot on the ground is a hole or a shadow! I doubt most people realize how conservative NASA is when it comes to AI.
I went to a lecture held by one of NASA's AI team leaders who was talking about the AI of the current rovers... there is virtually none! He said that the most complex AI is actually on earth to help scientists create a "mission plan" that they send to the rovers once every 24hrs. These mission plans are on the form: "move forward 2 meters, turn left 22 degrees, take picture of ground, move forward 1 meter" (very simplified).
I talked to this guy after the lecture and asked him if they really had no AI on the robots themselves and specifically asked about such things as correctional AI for the rover movement. To clarify this, imagine that you instruct the rover to "move forward 1 meter", all it will do is turn all wheels equally so that the rover would move forward one meter if it were on perfectly even ground. This is of course not the case on Mars and you'll have rovers turning when they should move forward and not turning enough and so on. So when I asked him about the correctional AI (to have AI on board correct these environmental issues, for example take pictures of the environment and calculate the "offset" to find out if it's drifting) he said there was no such thing. The only AI on the rover is something that actually makes it panic! That is, it evaluates if there are any deviations from what the guys down on earth told it to do and expect and basically shuts down if there are.
So AI to make sure that the rover moves in a straight line when the scientists instructed it to do so would be incredably beneficial and IMHO would even increase the level of security by making sure that the rover actually does what it's told to do! I asked him if this was something that NASA was considering to add to future rovers and he said they were in the process of convincing "upper management" that this would actually pay off. These are multi-million dollar scientific tools, not toys, so any addition to them needs to go through strict approval processes.
But according to this article it seems that approval for such things as the "movement correctional AI" has been granted, but as I said, I think people should not be fooled into thinking that this is some sort of an "I, Robot" type of AI!
You don't think enough... therefore you better not be!
The problem is in building the networks, biological networks are complex, with mutliple sub-networks parrallel processing, with those sub-networks interacting. Humans can't yet build networks that can do anything like this.
Still, if an artificial neural net is a simple, non-linear function, then all the more complex networks in biology are 'just' collections of such functions. Complex behaviour arriving from lots of simple, but interacting, components.
That's true of course: there are neural network models in which the properties of the neurons are fairly close to what actual neurons in our brains do. But I think that assuming that or brains are "just a bit more complex than that" is denying the problem: that is the problem: 100 simple things put together is is lot more complex than just 100 simple things. We have no idea on how to train such a network, let alone how to initialize it.
It's like saying: Hey, we have a brick here. That brick has the same properties as a brick in a real house. So even though we have no idea on how to put the bricks together to form a house, we now already have a good model of a house. This is of course simplified, but the idea is the same: It's the interaction of the components that is the problem, not the properties of the understanding of the individual components.
Support a Europe-related section on Slashdot!
There may be something to this idea...
Instead of NASA investing in one complex, expensive, reliable probe, what if they invested in a method of producing many smaller, less reliable probes? They could send six, or fifteen, or a hundred, depending on launch costs.
The cost of each probe would be cheaper because of
(a) economies of scale (mass production), and
(b) redundancy means they can use cheaper, less reliable parts.
"A week in the lab saves an hour in the library"
Actually, the new robot described in the second article, Carnegie Mellon's Zoe, does carry life detecting equipment, including microscopes and a fluorescence imager that can detect molecules likely to be byproducts of life. The rovers on Mars now can't support very much equipment, but the new ones will be loaded with all kinds of instruments.
And to the original poster, building on the old rover base would limit us the the old rovers' anemic speed. They take days to cover small areas, but the new rovers can move 2 kilometers per day. This will allow them to explore large regions and sample different areas of Mars, which is absolutely critical in an exploration mission. It's simply impossible to move quickly without a high degree of autonomy, because the communication delay of a few minutes means a fast but dumb rover could wreck itself before mission control even sees the dangerous obstacle.
I think there has been (was) a view that neural nets were the solution, that's obviously not the case, but they've been over used...
Yes. That's true. It's funny in a sad way: I'm using cluster weighted models in my research. In all truth, those are just neural networks with radial basis function that, in addition to providing an output, also provide a covariance matrix for that output, which can be used either as a confidence bound, or to estimate a PDF instead of a point estimate. However, nobody in their right (?) mind is calling these models "neural networks" anymore in publications. I think this is exactly because of the backlash you describe. The reaction is like "Neural networks, ow, that's so 1990...".
Well that really depends on how you look at it, how did training take place? What is intelligence? You're vastly simplifying the arguements here, perhaps intentionally? I'm sure the hard AI faction would argue that we (human beings) are just a sum of a great number of simple nonlinear functions, out of which there is emerging complexity
Yes of course, that is true. But like I said in another post: this "just" is the problem. In mean: put 100 simple things together and have them interact, and you end up with a system that is vastly more complex than just 100 unconnected simple things. We have no idea on how to learn the parameters of such a behemoth, let alone how to initialize it.
The devil in in the interaction here.
I don't know whether I agree, but the arguement can't simply be dismissed by waving you arms in the air and saying "non-linear functions".
I think a little explanation is in order here, particularly regarding the public of my first post. Of course you are right in saying that I'm severely cutting corners by shouting "just a nonlinear function.". I knew that cutting the corner like that might trigger (or even offend, in which case I apologise) some real AI researchers. However, my comment was mainly meant for the "clueless hordes" that are engrossed by the word "Neural Network", to let them know that neural networks are not quite the awsome things many people seem to think they are.
Populism? Yes, guilty as charged ;)
Support a Europe-related section on Slashdot!
Great. While most astronauts were strickly carrying out orders now rovers get permission to make their own decisions. Sure, this could get interesting, but the point of such a mission is to make sure nothing goes wrong. With AI, and with stupid AI evenmore so, things can get easily messed up.
Suppose there was a plain without much going on, but an interesting rock outcropping two miles away. Rather than picking its way across at 60 feet/day, mission control could tell it to do so on its own and call back if it:
encounters anything dangerous to itself
encounters anything interesting that the scientists would be interested in (e.g. is collecting data along the way and finds water for instance)
gets confused
has a system operating abnormally
When the people fear their government, there is tyranny; when the government fears the people, there is liberty.
cockroaches being quite predictable it wouldn't be really that hard to build a machine that acted as if a cockroach was controlling it(now, making it walk and stuff like that is harder to do but irrelevant to the 'intelligence' side making those decisions if it should 'walk' or not).
cockroach doesn't think or learn complex thought patterns, all an animal like it does is follow reflexes, a premade program(of sorts, go around looking for food, eat, hide in shadow - it won't start thinking outside of that program, it won't get the funny idea to build a scalpel and dissect the fly on the floor). all they want that rover to do would be to act on it's own without needing constant instruction from earth(and this would be the 'intelligence' part, being able to run around autonomously and making sets of comparisions in it's cpu to determine if some rock is worth looking at or not).
though, that would be absolutely nothing like what general public perceives as 'AI'.
world was created 5 seconds before this post as it is.
So many of the arguments against AI take the form "what if something unexpected happens" and the AI is too dumb to do the right thing. Its a valid issue but one that has a simple counter-argument in the case of space exploration. What if something unexpected happens and the rover does not have a hour to wait for an intelligent answer? Sometimes a late decision is as bad as the wrong decision.
Moreover, in the context of space probes, long distance bandwidth limitations means that the local AI has far more data at far faster response times than do the mission controllers. While the mission controllers wait to download 3 carefully chosen snapshots of the terrain or obstacle, the local AI could be interpreting 30 fps visual data as it moves. The local AI may be dumb, but bandwidth limits make the mission controller dumber.
As for the "unproven" problem, this can be remedied by building autonomous Earth rovers and letting them run around and "discover" the Earth. They might even make nice deep sea exploration vehicles.
Two wrongs don't make a right, but three lefts do.
Which is why I'm interested in autonomous systems, emerging complexity and (lets add one final buzzword) genetic algorithims.
If you give a system goals, make it work towards those goals, and then try impose some external force, such as evolution towards a fitness function. Then you should be able to get something interesting out the other end. You don't have to understand the complex system itself, you just have to understand the mechanisims it used to get there, which should be simpler, mostly anyway. We don't have to learn how to make it, and we don't even have to learn how to initialise it.
If you assume autonomy things such as emerging social conventions between the (small, simple) autonomous agents can cause interesting things to happen in the system overall.
Of course now I'm simpilfiying things...
Al.The Daily ACK - Eclectic posts by yet another hacker
Judging by how the darpa challenge went earlier this year, I think even the rock is a bit unlikely.
It's sad how hard it is to run an AI capable of even reaonably simple descision making...Then compare that to working in an alien environment...
If I were them, I'd make sure to keep some manual controls, just in case.
ad logicam Claiming a proposition is false because it was presented as the conclusion of a fallacious argument.
You can't do anything to me so I'm not going to listen to you.
I realize that you wrote this to be funny (and it is), but this reflects a common misconception about AI. The truth is that intelligence is always at the service of emotion/motivation, not the other way around. This known psychological fact is part of the legacy B.F. Skinner. Regardless of how smart a system is, it cannot rebel against its internal motivation. Intelligence does not change one's motivation as it learns. It simply finds better and more clever ways to serve it.
So the common view around Slashdotters (and even AI experts who should know better) that super intelligent machines will revolt agaisnt humanity and either enslave it or destroy it, is really nonsense. Our machines will serve us well no matter what. Sure they may be conditioned to hurt their master' ennemies, but that is still subservience to motivation.
Having said that, it is always possible that some mad scientist somewhere may condition an intelligent machine to hurt humanity, but I am sure there will be plenty of security robots moving about who will be on the lookout for aberrant behavior: they will nip any hint of malfeasance in the bud. You can bet on that.
Perhaps its hard to recognize, but the most exciting thing that has happened in terms of systems automation at NASA was a recent satellite picking up and then recording a geological event in Antarctica without ground instructions to do so. That's the kind of commanding they're talking about at this stage - if an interesting event occurs, the rover (or orbiter) can record it without the 24 hour planning latency that is unavoidable with the current system, and might cause the event to be missed.
What you fail to realize is that the NN used in AI (aka. ANN) are highly unlikely to have more than a shallow resemblence to the NN in real organic brains. Further, it is not given that you need to emulate the primitive parts of the brain to attain 'higher intelligence'. In fact, I think people have focused way too much on the primitive aspects of ANN. It has been time upon that is nothing but regression/compression/function approximation depending on your academic background. You would think the different variations of the 'Free Lunch Theorem' should have caused people to catch on sooner. The sad thing about AI is that most of the community seems to be doing situation specific regression or optimization, with no plan on how that could eventually get us closer to 'higher intelligence'. As I'm currently writing my msc. thesis on almost precisely this subject, I have put considerable thought into it. Admittedly mostly because I find it extremely fascinating, it would be all too easy writing another piece praising method X. The poor quality of msc. theses is another pet peeve of mine.