3 ms·

I think the story got it wrong - that it's the "thinking" that is harder to automate (as opposed to "doing"). Once we have AI capable of managing a fast food re

by lgieron 13y ago

I think the story got it wrong - that it's the "thinking" that is harder to automate (as opposed to "doing"). Once we have AI capable of managing a fast food restaurant, the workers will probably be long gone anyway (replaced by robots).

waps 13y ago
Having done some automation with actual robots, I must disagree. Determining what needs doing is a policy problem. Sure there may be variables, but it seems to me essentially you have a priority list, a list of workers and other resources, and essentially a bucket fill algorithm. Further enhancement might be to verify that the store is actually operational during all the different times (e.g. at 6pm, refill the fries before you do the cash register counting, even though the cash register stuff is more important in the end). Add to that that you can make significant mistakes and not have real consequences. In fact you could screw up a few things and still have better overall results than a human planner. And manna like planners are pretty normal in Doesn't sound like it's a huge problem to me. Having a 3 segment robot arm pick up a salt shaker requires balancing 7 codependent feedback loops, while constantly checking if you're hitting anything and requires constant evaluation of forward and inverse kinematics. As if that problem is not bad enough, you have to do it not just fast, but in real-time (as in, you have to calculate all this at ~50hz at least. If a single decision takes 100 msec, you've just damaged or lost the robot). State of the art in robotics is picking up and unboiled egg 10 times without breaking it, and most robots can't do it. If you start to use those robots to hand people burgers, expect broken bones at least, casualties at worst. Replacing a burger flipper requires solving not one, but actually most unsolved AI problems. You need voice recognition + understanding. Okay that's one. Then you need to understand the meaning in what was actually said. You know, the problem that google's 35000 employees from amongst the "world's best" don't quite know how to solve. I mean, they're far ahead of the pack, but nowhere near open-ended conversations with a human, never mind doing so with social grace. If your robot can walk around you need to manage an unsolved 30-dimensional feedback loop problem. State of the art for that problem is pathetic as well[1][2]. If it can't walk around you have other problems, and it's still at least a 14-dimensional feedback problem (two six-dimensional arms + 2 grippers). There's the mechanical problem of creating a robot that stays robust when working with "gooey" stuff like food (it will build up inside the robot, then fail or cause overheating). Then you need to solve all the problems that the "manna" system solves, or there's no point having a robot in the first place. [1] http://www.youtube.com/watch?v=vjgkV77gVcs http://www.youtube.com/watch?v=vjgkV77gVcs (note the obviously reduced dimensionality that wouldn't fly in the real world, and that the robot sees no problem moving through itself. It's stability under disturbances is also not at the point that it would keep operating for long periods. Do that in real life and the thing will topple over) [2] And before you say Bigdog, note that 4 feet locomotion is a much simpler problem. Bigdog is solving for 8 dimensions, and that's state of the art (and it is not sufficiently reliable to operate a single day without assistance). 2 feet locomotion in a human is ~140-dimensional, the minimum you can reduce it to while still having a nice mobile robot is 12-dimensional.

by lgieron 13y ago

waps 13y ago

Having done some automation with actual robots, I must disagree. Determining what needs doing is a policy problem. Sure there may be variables, but it seems to me essentially you have a priority list, a list of workers and other resources, and essentially a bucket fill algorithm. Further enhancement might be to verify that the store is actually operational during all the different times (e.g. at 6pm, refill the fries before you do the cash register counting, even though the cash register stuff is more important in the end). Add to that that you can make significant mistakes and not have real consequences. In fact you could screw up a few things and still have better overall results than a human planner. And manna like planners are pretty normal in Doesn't sound like it's a huge problem to me. Having a 3 segment robot arm pick up a salt shaker requires balancing 7 codependent feedback loops, while constantly checking if you're hitting anything and requires constant evaluation of forward and inverse kinematics. As if that problem is not bad enough, you have to do it not just fast, but in real-time (as in, you have to calculate all this at ~50hz at least. If a single decision takes 100 msec, you've just damaged or lost the robot). State of the art in robotics is picking up and unboiled egg 10 times without breaking it, and most robots can't do it. If you start to use those robots to hand people burgers, expect broken bones at least, casualties at worst. Replacing a burger flipper requires solving not one, but actually most unsolved AI problems. You need voice recognition + understanding. Okay that's one. Then you need to understand the meaning in what was actually said. You know, the problem that google's 35000 employees from amongst the "world's best" don't quite know how to solve. I mean, they're far ahead of the pack, but nowhere near open-ended conversations with a human, never mind doing so with social grace. If your robot can walk around you need to manage an unsolved 30-dimensional feedback loop problem. State of the art for that problem is pathetic as well[1][2]. If it can't walk around you have other problems, and it's still at least a 14-dimensional feedback problem (two six-dimensional arms + 2 grippers). There's the mechanical problem of creating a robot that stays robust when working with "gooey" stuff like food (it will build up inside the robot, then fail or cause overheating). Then you need to solve all the problems that the "manna" system solves, or there's no point having a robot in the first place. [1] http://www.youtube.com/watch?v=vjgkV77gVcs http://www.youtube.com/watch?v=vjgkV77gVcs (note the obviously reduced dimensionality that wouldn't fly in the real world, and that the robot sees no problem moving through itself. It's stability under disturbances is also not at the point that it would keep operating for long periods. Do that in real life and the thing will topple over) [2] And before you say Bigdog, note that 4 feet locomotion is a much simpler problem. Bigdog is solving for 8 dimensions, and that's state of the art (and it is not sufficiently reliable to operate a single day without assistance). 2 feet locomotion in a human is ~140-dimensional, the minimum you can reduce it to while still having a nice mobile robot is 12-dimensional.