We all love robot pancakes.
October 31, 2010 1:55 PM   Subscribe

You've seen them here before: serving ice cream, pole-socking, with teddy bear heads, climbing trees, and sporting hands. But now robots are truly Metafilteranean, because they want to know: Who here likes pancakes?

TUM-Rosie and PRJames of the Munich-based Cluster of Excellence Cognition for Technical Systems work together to prepare and serve pancakes.

Want more robot-pancake interface? Check out a robot learning to flip pancakes and robots picking pancakes (the latter mentioned previously).

("Who here likes pancakes?" "I love pancakes," is a running gag spawned from a comment first posted by aaron in 2001 and, if MeFi had one, would be its call-and-response song.)
posted by jocelmeow (15 comments total) 6 users marked this as a favorite
I love pancakes, but isn't this kind of a double?
posted by elizardbits at 2:46 PM on October 31, 2010

"Do you like pancakes?"
"Yes we like pancakes..."
posted by ejfox at 3:07 PM on October 31, 2010 [1 favorite]

I was at a hotel in Seattle recently which had a pretty good breakfast buffet. Instead of the typical bank of waffle irons which patrons could use, it had what we called the pancake printer. Push a button, and a couple of minutes later, a pancake came out of a slot in the side, EVER so slowly. It was truly a pancake bot of the primitive R2D2 order. Very amusing.
posted by hippybear at 3:26 PM on October 31, 2010 [2 favorites]

But can the robots get the rabbit to sit still for a picture?
posted by kuujjuarapik at 3:40 PM on October 31, 2010

"Who here likes pancakes?"

I like flapjax at midnight :)
posted by puny human at 4:11 PM on October 31, 2010 [1 favorite]

Something about its expression makes it always look so curious.
posted by codacorolla at 5:17 PM on October 31, 2010

which hotel in Seattle? Enquiring minds want to know.
posted by warbaby at 5:38 PM on October 31, 2010

Red Lion Inn on Fifth Avenue downtown.

Great rooms, kind of okay rates for a downtown hotel, great breakfast buffet which was included in the room price... TERRIBLE wait service for the happy hour and in the restaurant under the lobby. Terrible. As in, bad enough that I wrote letters to the company complaining.

I recommend the hotel, just avoid the restaurants.
posted by hippybear at 6:31 PM on October 31, 2010


Socking (to hit or strike forcefully) might be be better expressed as gloving (to cover with or as if with a glove.) Yeah I know, but hey, it's Metafilter.
posted by StickyCarpet at 7:23 PM on October 31, 2010

But they're not gloves; they're socks!

If you just can't live with "socking" as an ambiguous word, I propose "pole ensockening."
posted by nebulawindphone at 9:13 PM on October 31, 2010 [1 favorite]

This is my kind of comedy!
posted by breath at 11:48 PM on October 31, 2010

How do you positively reinforce a robot's behavior if it can't feel pleasure or reward? Is it simply programmed to "want" to flip a pancake the right way? This robot operant conditioning is very interesting.
posted by tamagogirl at 7:49 AM on November 1, 2010

It turns out that this sort of "conditioning" (well, it's usually called "machine learning," but I guess "machine conditioning" would have been an appropriate name too) doesn't require pleasure or desire at all. What it does require is expectations — or, well, predictions, at least.

The machine has a mathematical model of how pancakes move. The model lets it make predictions. ("If I move the pan like this, these equations say that the pancake should move like this.") Of course, the computer doesn't care if the model's predictions are right. It doesn't "enjoy" being right. But it can still recognize when the predictions are right and when they are not. ("The model says the pancake should move like this, but it actually moved like that.") So you can program it to make certain corrections to the model's parameters whenever a prediction turns out wrong. ("The pancake spun faster than I predicted. My programmer says when that happens, I should change this angular momentum parameter here...")

Conveniently, the machine doesn't have to "know why" it's making those corrections. It just has to make them, following the program it's been given. Hopefully the programmer knows why he's calling for the corrections to be made in a certain way. But the machine is like a very dumb lab assistant in a physics lab: "My boss says I should work out this equation and try this experiment; and then if it doesn't work, I should try this instead; and if that doesn't work.... — No, I don't know why he told me to do that. I'm just following directions."

If the machine keeps repeating this process — make a prediction; test it; adjust the model's parameters if the prediction was wrong — it will eventually arrive at a very precisely tuned model indeed. Then you can query that model.
"Find a set of input motions that you predict will make the pancake bounce. Now carry out those motions."

"Okay, what input motions will make the pancake fly across the room and hit my annoying labmate? Carry out those motions."

"Okay, now what input motions will make it turn over once and land without bouncing?"
And so on. The better-tuned the model, the more accurate the answer. And of course, if its answer to your query turns out to be wrong, the machine will change its model parameters yet again in response, and so on until everything is just right.

It will look like the robot is "trying" to follow your instructions, like it "wants" to do a good job. But really it's just making and testing predictions, making and testing them, until the model is tuned properly and its predictions match with reality.
posted by nebulawindphone at 9:10 AM on November 1, 2010 [2 favorites]

It will look like the robot is "trying" to follow your instructions, like it "wants" to do a good job. But really it's just making and testing predictions, making and testing them, until the model is tuned properly and its predictions match with reality.

I think that is probably how humans work when they're learning tasks, too. Like, threading a nut onto a bolt, or playing the piano.... you predict how your muscles will move and what the feedback will be, you attempt the task, it works or doesn't, you try again... Eventually you get to where the muscle motions and the feedback work every time to achieve the desired outcome.
posted by hippybear at 4:41 PM on November 1, 2010

This is ridiculously awesome. People are betting a lot of money that domestic robots that can accomplish a variety of household tasks are possible -- the Japanese government, faced with a rapidly declining population, spends $25 billion a year on robotics research. Given the amazing progress I've seen just in the last 2-3 years, I think we'll have robots that can (say) make a meal or do the laundry in my lifetime. Whether they're cost effective or reliable for routine use is another matter entirely.
posted by miyabo at 1:37 PM on November 2, 2010

« Older Drácula   |   Not Fit for Human Consumption Newer »

This thread has been archived and is closed to new comments