Slashdot Mirror
Of Ants and Robots
conJunk writes "The BBC has an interesting story about Ants and their leaderless collective behavior. It goes on to describe these cool little robots called U-bots. They have a super-simple instruction set and if you let them loose in a room full of frisbees it looks, to the casual observer, like intelligent and guided work." From the article: "Being small is going to be a problem. So how can you get a whole bunch of dumb small things doing something smart?"
Each neuron in our brain is dumb compared to our entire brain.
Same thing with these ants and these robots..
And here I tought the fact that complex problems can be broken down and solved by simplistic devices was a founding tenant of computer science.
"Learning is not compulsory... neither is survival."
--Dr.W.Edwards Deming
your brain is a bunch of dumb things (neurons) doing something less dumb
Sort of, except its a lot less random than monkeys on typewriters. They do have a good idea of what they're doing, its just a lot less of one than what's standard to a robot these days. The goal is to distribute the intelligence so they can still work together and get something done with the efficiency of ants.
The reason this is good is as he was saying, its a lot cheaper to make 15 simple robots than to make a couple complex one, because the manufacturing cost per unit decreases with the amount of units created.
Think of that part this way: You want to build either 1000 matchbox cars or a full size car. Lets assume that the total cost of materials is the same to make things simpler. The full size car is made on equipment that costs maybe $100,000 because it has to be custom-created. The matchbox cars can be made on equipment that costs $1,000, and then that whole price is divided by 100, for a manufacturing cost that's effectively $10 per unit. Of course that's not really how much matchbox cars or real cars cost, but the scenario works. Also keep in mind, even if the equipment cost the same, the per unit price would drop with more units because its being divided.
I think that this is a really important avenue of research but can't help wondering why exactly this project was funded.
Robotics is of course great fun and can certainly be inspiring but all this was presented (albeit indirectly by a superficial BBC report) as a valid study in terms of what the miniture robots can achieve.
It doesn't take the 'Milliard Gargantubrain' to work out that all this stuff is better and cheaper simulated on computers. Cellular Automata have in various incarnations been here before (including countless ant based examples) . How does making it physically real advance the subject at all? Aside from the obvious 'it looks cool' and 'it allows us to write in general terms about ants instead of the truly vexing question of how intelligence can function equally well as a distributed system'.
Really, help me out here. Surely any one of us could have created and run 50,000 simulations in the time it took them to solder up the PICs (or whatever microcontrollers it was that they used). I'm not penny pinching here I'm just wondering if this was the best way to go about the problem.
"Myrmecology, noun, - The Scientific Study Of Ants. This has been Roseanne, your guide to the world of facts."
And two more things.
The U-Bots according to the article follow a few rules. However, while one rule contains 'drop item', NONE of the rules contain 'pick up item'. This means that either all N U-bots must have been carrying one item each (a total of N items) in the beginning, which means the place where items get dropped off highly depends on the initial configuration of the robots in the arena. Or, the article is flawed in describing the rules, because they are not sufficient to perform a 'discover, collect and concentrate' algorithm.
But the thing that strikes me most is WHY THE HELL DO YOU NEED TO BUILD ACTUAL PHYSICAL ROBOTS TO DO THIS!!!??!??
You can simulate all of this in a very simple piece of software. Especially when you later decide to increase complexity by building in any 'feromone trail' aspect, using physical robots just seems a foolish waste of engineering resources.
It only seems like a waste until you've tried it.
Implementing real world solutions to simulated problems always brings up quite a number of "interesting" problems. Things you never thought would be obstacles turn out to be nightmares in the real world. And on the flip side. sometimes a quirky solution to a problem presents itself.
I was working with some friends recently in testing a cross compiler for a robotics platform. They had a simulator and their code worked just fine in it.
But in the real robot it didn't work at all. It turns out that there was an mistake in their control logic such that the bot was switching between two motor drive states at millisecond intervals. This worked fine in the simulated bot, but the real motors of the real robot couldn't handle this (obviously), so it behaved poorly until we change the logic to switch more slowly.
This is a simplistic example yes, but also an excellent one, of the types of real world problems you can face.
Another frequently encountered problem is a light sensitive robot that works fine in the morning but falls to pieces in the afternoon when the sun shines through the windows of the experimental room.
Noone who builds robots only in simulators can be trusted to design real world devices. Their implementations will be brittle and practically guaranteed to be inoperative without major tweaking.