Slashdot Mirror
Modular Robots
levin writes "An article in the latest issue of IEEE spectrum discusses modular robots--robots made of small, identical components or modules. These robots can slither, roll like a tank tread, inchworm, or crawl like a spider. The idea is that modular robots will be not only cheaper to build because the modules are all the same, but will be more able to repair themselves (by shedding damaged modules). Even cooler, each of the 5cm cube modules in Xerox PARC's polybot sports its own PowerPC 555 and 1mb ram."
The 'imagine a beowulf cluster' comments are more applicable here than in most of the articles I read... Imagine billions of robots able to work in tandem, infinitely reconfigurable. An office building/space craft? An automobile/boat. Hello Transformers...
Guvegrra?
By advances, I was talking in general. Is there any type of systems developed for consumers that can make decisions based on outside input and make physical reactions based upon a large set of rules or guidelines. The generalized question was to ask if there were any robots or mechanized devices that can be purchased at affordable prices(
Beyond your average "Programmed to go towards light, then programmed to bark" kind of robot you described.
You'd be surprised how much costs can be reduced when something becomes mass-produced. Most of the construction cost comes from having to build different objects and then assembling them.
And the men who hold high places must be the ones who start
To mold a new reality... closer to the heart
Ok, so I'm probably going to be pointing a giant neon sign at my head that reads "NERD!" but, back when I was into table top RPG's, Battletech in particular, There was a segment of the Battletech universe called "The Clans" which had mechs that had modular weapons. Meaning, all you had to do was supposidly slide in and out the weapons for configuration. Now in actual game play, this didn't really show up (Gameplay being much more cooler when Mech's have an individual shape - Mad Cat as opposed to a Thor for those of you that played the video games) but it is just a little bit of background factoid which runs similar with the story
blah
Yep, imagine module Vaporizer Weapon being detached from competitor A. Then B picks it up and installs it on an open slot. If modular robots could enter BattleBot and do this, imagine the beasts that would compete over the finals. Assembled parts from the fallen enemies, that's very cool..
Buy a Nintendo DS Lite
But imagine what you could do with a 5 cm^3 computer if it were a self-powered cube that could automatically share number-crunching resources with any other cube it got connected with.
Specifically, I address in this off-topic post the feasability of simulating the human brain with current technology.
Before we start, see here for the statistics I am using.
Note especially that:
I'll now interpret this information.
Let's posit for a second (wrongly) that a five hundred megahertz computer ("PowerPC 555" in article, though again the article refers to robots, not mere number-crunching computers) could simulate with each hertz all that a neuron does in one firing. (By contrast, a typical "hertz" in today's gigahertz computers is less than required to retrieve two thirty-two bit numbers, add them, and store the result.)
With this assumption, we'd only need (upper estimate) 200,000 such processors [1] to simulate the brain real-time.
200,000 * 5 cubic centimeters (size of these suckers) is 1,000,000 cubic centimeters, or 100 centimeters to each side of a cube, which is 1 cubic meter.
That's not very big at all, and even if these robots cost $2,000 each, 200,000 of them would only cost $400 million.
The problem, of course, is that no way one hertz on these babies is going to simulate all that a neuron does, even on average, since each neuron is connected to up to 5,000 other neurons, and has a small interaction with each one each time it fires.
Since a 32-bit integer can enumerate ("address") just over 4 billion items, we would need an integer and another byte (we'd only use half) to address each of the other 50 billion neurons. In other words, just to pass information about which current connection we're looking at we need to handle two 4-byte integers and another byte on each end of your dendrite (connector and connectee). If we assume that an "interaction" between two neurons, when one of them fires, takes a hundred real hertz to process (I think this is fair, since the amount of logical information that a neuron stores can be represented by two or three variables, which you'd read, compare, see if a threshold is met, then store), then we'd need not one hertz per neuron but 100 hertz * 5000 dendrites (connections to other neurons with which it transacts). Our 1 cubic meter has just jumped to 500,000 (five cubic kilometers), and our $400 million price-tag has just jumped to $20 trillion.
But $20 trillion will buy you the processing power (not necessarily the io bandwidth) to process as much as the human brain can possibly, ever process, if every neuron is connected to as many other neurons as it possibly can, and each one is firing as much as it biologically can, by the highest estimate anyone estimates, and is connected to as many other neurons as anyone estimates is possible.
Needless to say, your actual costs for doing as much processing as the human brain processes are much, much lower.
Why, if you take simply the fact that the max hertz we calculated as 2000, whereas the "max" is 250-2,000, and the "average" by most estimates is around 20 hertz (a neuron, on average, will not fire more than twenty times a second), you've just reduced your processing time by a factor of 100, going from $20 trillion back down to $200 billion.
Now let's look at the difference between the "processing" that we said we can buy for $2000 (500 megahertz) and the io bandwidth we need.
We estimated 100 hertz per neuron interaction with another neuron, and we said that a neuron was connected with 5000 other neurons, and that the "state" of each connection could be represented (logically) by three 32-bit integers (four bytes each) and another 5 bytes just to address the second neuron, we now need 8 bytes * 5,000 neurons available over the timespan of 100 hertz, where we're looking at a 500 megahertz computer. This means that to get the io bandwidth over one second, we multiply these eight bytes by 5,000,000 (the quotient of 500 megahertz and 100 hertz), and get 40 million / 1024*1024 = 38.14 megabytes/second.
If we forget about the 5-cm cubes (and any semblance of topicality) this actually isn't so unreasonable, since a $2000 computer needing only 500 megahertz shouldn't have any problem with 38.13 megs/second. Or 4 gigs of RAM.
Anyway, let me know where my numbers are off, but it seems I've concluded that, today, $200 billion will buy you everything you need to simulate a human brain real-time, without any compression or special optimization.
So then next time somebody says: "Computer will never think, because only human can think." You can proudly answer:
"Shut your face, ignorant person. Soon as we figure out all the laws of neural interaction and find a way to image someone's brain, $1.57 billion dollars will buy you all the computer processing you need to simulate that brain real-time. [10.5 years from now, or 7 Moore's law iterations -- I divided $200 billion by two to the seventh]. But, of course, if ten and a half years to you is longer than "never", then feel free to remain ignorant, moron."
ac.
of course, I've been known to be wrong. please correct me gently.
[1] this is 50 billion divided by 250,000, since 500 megahertz is 250,000 more frequent than 2,000 hertz.
That is a pretty good question. In the likely uses for these things only a few motors will run at any one time, with the rest of the modules just maintaining a locked position. So they'd better design them so they can lock in position without a continuous power draw.
Power-PC CPU's draw a lot less power than Pentiums (I think), but it's not going to be practical to have portable-powered units (solar, battery, fuel cell, or combination) with large numbers of them. They also built a chain robot with PIC CPU's -- that's an 8 bitter that draws a little more current than a digital watch. It could follow a pre-programmed plan but couldn't re-configure itself, and probably isn't brainy enough to handle unpredictability. So I expect a real world modular bot will be a lot of PIC "muscle" units, plus a few high-powered "brains". The PIC's will have just enough intelligence to follow the plan sent out by the brains, and not burn up much power when idling.
The lab model (powered by a wall plug) might put a "brain" in every unit, to make it easier to work out the basic motions. Then you compile that to PIC object code add it to the list of options the brain can activate.
Here's a guy with even grander ideas along these lines. Crackpot or genius? Take a look and decide for yourself....
"If you're thinking what I'm thinking, you're right." -