Slashdot Mirror
When Robots Play Games
Roland Piquepaille writes "If the theory of evolution has worked well for us -- even if this is arguable these days -- why not apply it to mobile robots?, asks Technology Research News. Several U.S. researchers just did that and trained neural networks to play the Capture the flag game. Once the neural networks were good enough at the game, they transferred them to the robots' onboard computers. These teams of mobile robots, named EvBots (for Evolution Robots), were then also able to play the game successfully. This method could be used to build environment-aware autonomous robots able to clear a minefield or find heat sources in a collapsed building within 3 to 6 years. But the researchers want to build controllers for robots that adapt to completely unknown environments. And this will not happen before 10 or maybe 50 years. You'll find more details and references in this overview, including a picture of EvBots trying to find their way during a game." Read on for a similar robot competition held this weekend in France.
saunabad writes "The annual Eurobot autonomous robot contest for amateurs is held this weekend on La Férte-Bernard, France. This year's theme is 'coconut rugby,' and the robots are collecting small stress balls from the field and carrying them to the opponent's end, or shooting them in the rugby goal, while avoiding the randomly placed obstacles at the same time. Each team has a one main robot and an optional small assisting robot."
Aimbots have been around in CS for years. Is this really news?
http://www.aaplblog.com/ - News about Apple Inc.
You know all they're going to do is run into each other and explode.
find heat sources in a collapsed building within 3 to 6 years.
Yeah, I think the body will be cold by then...
You can't take the sky from me...
until they master jumping around while strafing and shooting..
I am not impressed until I see one jump+crouch and scream 'I pwn j00!'
--- We need more Ron Paul!
If the theory of evolution has worked well for us -- even if this is arguable these days
Do I detect the scent of an evolution denier? And it is interesting that you implicitly question the validity of a theory even as you cite an example of its successful application.
Wikipedia mentioned Quake and UT. Bzflag is also a great CTF game, and a classic.
Robots are cool and all, but why bother building and programming robots to find mines when we already have biological robots that can do the same thing while running off of water and a little bit of food. It seems a bit like a wonderful solution to a problem that doesn't exist - evolution has been doing pretty darn well at doing this sort of thing so far, so I'm not really sure why would need robots after all this time.
"I would give my right hand to be ambidextrous."
...no one asks it to play global thermonuclear war.
I vote we drop capture the flag, and just start up the tic tac toe game right now.
"Infants flesh will be in season throughout the year." -Swift
"After several hundred generations, the neural networks had evolved well enough to play the game competently and were transferred into real robots for testing in a real environment. "The trained neural networks were copied directly onto the real robots' onboard computers," said Nelson. "
As someone who spent a considerable amount of my childhood less interested in 'organized' sports and instead playing this game, it seems the whole point of playing Capture the Flag was to develop strategies in how to win. We had a set of rules that evolved over the years, depending on how many kids were playing, what time sunset (or the first person called back to their house would be), etc. We even had evolving words that were based on nonsense - or the inability of one of the younger kids to say a word (for instance - in some "Steal the Flag" games - the term "electricity" is used to talk about a strategy that involved making a line of kids that attacked from one end - they all held hands in the stragegy so that if anyone was captured they would automatically be "freed" by the "electricity" back to their own side. We deemed this a violation of the intent of the game, so we had a *no electricity* rule some little kids couldn't pronounce right - so it became "no a-la-ca-triss" - or something like that).
The game wasn't about *object avoidance*, it was about kicking ass through completely ad hoc strategies that had to be original because the teams always traded players rapidly, so you didn't want to make a rule or come up with something that would come back to bite you.
In this way - the random nature of our game was more like evolution than the winning was (it shuffled the components and allowed for *mutations*). The fact that the model showed no improvements with greater numbers of computers is not in line with what actually happens. The best games were the huge ones.
This simulation was probably a lot of fun to watch once the program was transferred to the robots though...
I disagree that it will necessarily take even 10 years and it will certainly take less than 50. Pathfinding and object search algorithms are strong even today. With a combination of radar, sonar, lidar, and optical recognition, I think we should be able to create robots which traverse formerly-unknown terrain in ten years or less.
I'm not trying to trivialize the difficulty of the problem, all the stuff we take for granted as we navigate a room is really quite a lot to deal with and it is only through practice that we are so successful, but an awful lot of effort is going into these problems (I know "more than ever before" is cliche and obvious but nonetheless...) and it is a top priority for so many very smart people that I cannot see it taking even a decade for useful robots with these capabilities to be in use.
Of course, it depends on what you want them to do when they get there...
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
The introduction makes it sound like training neural networks is evolution. Neural networks and genetic algorithms are two very different technologies, although they can be combined.
Table-ized A.I.
Also, as tasks become more and more complex, it becomes much more difficult to "evolve" systems that behave exactly as you want them to. There are a number of stories of neural nets being trained to recognize some feature from a set of training inputs, and instead keying in on some completely different and irrelevant detail.
Hmmm. That describes my boss pretty well. I think i'll check for a port on the back of his neck.
Table-ized A.I.
- they picked the rats because they're too light to set off the mines and are single-minded enough that they work better/cheaper than sniffer-dogs.
- The article describes using cables/tethers to restrict the rats to a line of interest. Hopefully, you can extend this concept to multiple rats on parallel lines and see how that'd allow efficient mine-sweeps of areas of concern.
- The rats live 6 years and can be bred, travel lightly, etc. This is EXACTLY what the parent poster meant when they talked about evolutionarily handling a cool problem rather than expecting rapid results (cheaply) from robots.
- How little do you figure you can make your smart robot for? A few grand? And where will Afghani's (or third-world citizens anywhere, especially those recovering from the economic impact of the very wars that placed these mines) get that money, a steady source of repair parts, etc? Instructions on training, a pair of rats, and fifty yards of string/wire and a clicker could let any small village have their own demining capability. Somehow, I don't think robots are gonna be as versatile or cost-effective.
Seriously, the parent poster on this should have considered posting it as a story (unless it's old news). It sure seems to me to be a great blend of nerd-interest factor, news, and stuff that matters. Props to the parent poster and the involved researchers. Within my life, we'll likely have cheap devices with artificial noses or GPR or another solution. But abandoned mines are too wicked to wait that long.Even discounting these things, worrying about the ethical implications of hurting an animal by training it as a mine-sniffer ignores the huge ethical implications of going the other way: if nothing is done, people die or are maimed. We've had this argument: using animals to save human lives is not taken lightly, but it is ethically tenable.
There is a basic fallacy in this sort of research - that evolution will necesarily develop some kind of intelligence to solve problems. Evolution will do "what it takes" to solve a problem - and no more. If you attempt to use evolutionary techniques to, for example, solve mazes, you will end up with a system very good for solving mazes - and nothing else.
This happened in computing in the 70s. Intel found it convenient to solve the problem of calculator design by buoilding the 4040 - the first microprocessor, But this was in no way *necessary* - Intel could have continued down the old line of discrete logic.
Evolution is a powerful tool - but not a panacea
Consciousness is an illusion caused by an excess of self consciousness.