Robots Are Going To Kill Us All!

The issue with just seeing the physical motion, can't tell what code is actually running in the background. (If they're talking in the background I apologize, can't watch it with sound on). It would be interesting if this is truly organic; if it's simply a preset list of motion types it cycles through, it would be much less impressive. Would seem to be much more efficient to simply have pre-set defaults based on damage to each leg.
 
TwistedAegis said:
The issue with just seeing the physical motion, can't tell what code is actually running in the background. (If they're talking in the background I apologize, can't watch it with sound on). It would be interesting if this is truly organic; if it's simply a preset list of motion types it cycles through, it would be much less impressive. Would seem to be much more efficient to simply have pre-set defaults based on damage to each leg.
Click to expand...

At it's core this is exactly what it is, a set preset behaviors that are chosen based on what the damage is. However it starts getting very complex when you consider all of the variables. The robot would need to know what limb is broken, how that damage affects the gait and how to compensate. Then it would need to sensory input to find out if it's chosen compensation method is allowing it to move in the desired direction...and then (presumably) continue in a loop until the desired outcome is achieved. Oh, and all without human input.
 
So when the leg finally becomes damaged, the robot chooses a gait from the subset that minimises contact with the ground for that leg. It then measures how quickly it can walk using this gait.

It then uses this information to choose another gait, measures the resulting walking speed and feeds this information back into the model. After just a handful of tests, it can then select the best gait.
Click to expand...

From the article
 
Thank you; not to minimize the complexity, pretty awesome, just wondering how much was pre-set and how much was actually "learned" via experimentation.
 
State of the middle east got you down? Illegals flooding our country dragging you under? Not To Worry! The machines WILL get us in the end. :p
 
This is so cool.
I hope in the future we will be fighting machines instead of each other.

Dystopian-cyberpunk, here we come! :D
 
TwistedAegis said:
Thank you; not to minimize the complexity, pretty awesome, just wondering how much was pre-set and how much was actually "learned" via experimentation.
Click to expand...

Technically, machines don't have the ability to "learn" right now, they simply have the ability to respond to input and adjust behavior based on pre-defined/pre-programmed datasets.

In the example of this drone, they visualize it as a matrix of pre-defined movement instructions which is cycled through to find the nominal forward velocity. Whichever predefined movement instructions propel the machine forward at the highest forward velocity, and in the straightest path (deflection from an arc), that is the one that becomes the new standard movement function.

Pretty straight forward from a programming perspective, although I'm sure its very complex behind the scenes.
 
SRTie4k said:
Technically, machines don't have the ability to "learn" right now, they simply have the ability to respond to input and adjust behavior based on pre-defined/pre-programmed datasets.

In the example of this drone, they visualize it as a matrix of pre-defined movement instructions which is cycled through to find the nominal forward velocity. Whichever predefined movement instructions propel the machine forward at the highest forward velocity, and in the straightest path (deflection from an arc), that is the one that becomes the new standard movement function.

Pretty straight forward from a programming perspective, although I'm sure its very complex behind the scenes.
Click to expand...

IMO....I think it'd be relatively easy to program for the initial adjustment, it's the refinement and optimization after that initial test that would get interesting.
 
SRTie4k said:
Technically, machines don't have the ability to "learn" right now, they simply have the ability to respond to input and adjust behavior based on pre-defined/pre-programmed datasets.

In the example of this drone, they visualize it as a matrix of pre-defined movement instructions which is cycled through to find the nominal forward velocity. Whichever predefined movement instructions propel the machine forward at the highest forward velocity, and in the straightest path (deflection from an arc), that is the one that becomes the new standard movement function.

Pretty straight forward from a programming perspective, although I'm sure its very complex behind the scenes.
Click to expand...

Right, but it would be a big leap forward if there were less of the pre-programmed instructions of how to move, and more "higher" level instructions of how to discover how to move. Looks like there is some of both, but the lower-level detailed movement instructions are still there.
 
Bwah hahahaha, I was laughing so hard at this video its so awesome

Notice how the robot in its damaged state is moving faster than it does in its normal operating mode?

This just confirms every movie with a killer robot that moves faster when its damaged and chasing the protagonist ;D
 
"I know now why you cry. But it's something I can never do." said the limping spider bot.
 
What they did to that poor robot is inhumane. I hope those PETR people find out about this.
 
You must log in or register to reply here.
Back
Top