|
monkey posted:Can someone who knows something about building robots explain to me why we haven't made exact copies of humans yet? I mean a replica human skeleton powered by electric or pneumatic muscles. There's only 640 muscles and 200 bones in a human, it shouldn't be that hard. The bandwidth required to control one remotely would be trivial, you could drive it over wifi if not with a cellphone data connection. There's no such thing as electric or pneumatic muscles, stop reading poorly written science articles.
|
# ? Aug 3, 2016 17:22 |
|
|
# ? Jun 8, 2024 18:26 |
|
notZaar posted:There's no such thing as electric or pneumatic muscles, stop reading poorly written science articles. https://www.youtube.com/watch?v=xL5bInUumbM https://www.youtube.com/watch?v=G9dPfCQCQ-U https://www.youtube.com/watch?v=xMGXqT0LWUI https://www.youtube.com/watch?v=1A2LUbJjDQ0 https://www.youtube.com/watch?v=yqCwHzK81dA i could go on
|
# ? Aug 3, 2016 17:37 |
|
Those are fluid filled inflexible tubes, pretty much nothing to do with actual muscles. Any machine you build with them is going to be huge, stiff and slow.
|
# ? Aug 3, 2016 17:41 |
|
there exist hydraulic, pneumatic, electric and heat actuated artificial muscles. I missed the pneumatic video... one second... edit: https://www.youtube.com/watch?v=4Y1mSltcJJ4
|
# ? Aug 3, 2016 17:54 |
|
notZaar posted:Those are fluid filled inflexible tubes, pretty much nothing to do with actual muscles. Any machine you build with them is going to be huge, stiff and slow.
|
# ? Aug 3, 2016 17:55 |
|
a hole-y ghost posted:I've got a fluid filled inflexible tube right here you freaking nerd Do you have a patent?
|
# ? Aug 3, 2016 17:58 |
|
The Bananana posted:Do you have a patent?
|
# ? Aug 3, 2016 17:59 |
|
The other thing i dont get regarding electric muscles is why they don't make them using a similar lattice technique to those others, but on each intersection there are electromagnets, and you switch the poles on half of them to change the magnetic force from pull to push like this, but wrapped around a tube that is filled with stuff to power it:
|
# ? Aug 3, 2016 18:28 |
|
a hole-y ghost posted:The LSD: Dream Emulator theme park is shaping up to be a monster i plaeyed this on bleemcast and i like to think the bizarre crippling glitches added to the experience
|
# ? Aug 3, 2016 18:46 |
|
a hole-y ghost posted:I've got a fluid filled inflexible tube right here you freaking nerd This is a technology thread you got no right to throw around the N word in an aggressive and patronizing manner.
|
# ? Aug 3, 2016 18:53 |
|
notZaar posted:This is a technology thread you got no right to throw around the N word in an aggressive and patronizing manner. Geek
|
# ? Aug 3, 2016 19:10 |
|
notZaar posted:This is a technology thread you got no right to throw around the N word in an aggressive and patronizing manner. Dork
|
# ? Aug 3, 2016 19:15 |
|
notZaar posted:This is a technology thread you got no right to throw around the N word in an aggressive and patronizing manner.
|
# ? Aug 3, 2016 19:17 |
|
Maldoror posted:Is there a link or something that explains the neural net thing I'm sure a bunch of nerds gave you answers involving non-linear blah blah blah but long story short they are very useful because versions of neural nets can create their own feature sets. Like say you wanted to build an algorithm that classifies numbers as a 1, 2, 3.... normally you'd have to give the algo a list of things (features) to look for: is it round, how many edges does it have etc. Variants of neural nets excel at these sorts of tasks because they learn how to classify numbers simply by looking at labeled examples of numbers and the algo itself figuring out the important features it should look for without any user input. The programmer don't tell the program what makes a human face different from a basketball, the program looks at labeled pictures of both and figures out the important factors on its own. So of course these sorts of things are perfect for imitating human intelligence because they slowly 'learn' what they need to without a need for designing complicated decision processing programs and so on, which was the traditional approach. How they do involves a bunch of boring math that you aren't going to understand without an advanced degree in stats/cs.
|
# ? Aug 3, 2016 20:04 |
|
Fun fact: you can 'train' self- driving algorithms by playing computer games, and the games don't even have to be that graphically complicated to give good real-world results. Assuming the mechanics were taken care of you could probably make a war murderbot by playing a bunch of cod matches.
|
# ? Aug 3, 2016 20:06 |
|
TROIKA CURES GREEK posted:Fun fact: you can 'train' self- driving algorithms by playing computer games, and the games don't even have to be that graphically complicated to give good real-world results. Assuming the mechanics were taken care of you could probably make a war murderbot by playing a bunch of cod matches.
|
# ? Aug 3, 2016 20:08 |
|
TROIKA CURES GREEK posted:Fun fact: you can 'train' self- driving algorithms by playing computer games, and the games don't even have to be that graphically complicated to give good real-world results. Assuming the mechanics were taken care of you could probably make a war murderbot by playing a bunch of cod matches. Hmm yeah or just drop some bombs. Bombs are easy and we already got all these planes and rockets...
|
# ? Aug 3, 2016 20:13 |
|
I am mind melding right now. https://www.youtube.com/watch?v=JCqXVTnGzgU
|
# ? Aug 3, 2016 20:54 |
|
thats the most poo poo robot ive ever seen it cant do gently caress all
|
# ? Aug 3, 2016 21:05 |
|
reallivedinosaur posted:thats the most poo poo robot ive ever seen it cant do gently caress all
|
# ? Aug 3, 2016 21:36 |
https://www.youtube.com/watch?v=wqzLoXjFT34
|
|
# ? Aug 3, 2016 21:39 |
|
Yikes this is certainly two spooky four me
|
# ? Aug 4, 2016 11:12 |
|
Can somebody explain to me what purpose human shaped robots would serve? Besides weird sex stuff
|
# ? Aug 4, 2016 11:16 |
|
slaves
|
# ? Aug 4, 2016 11:17 |
|
telepresence
|
# ? Aug 4, 2016 11:20 |
|
but mostly weird sex stuff
|
# ? Aug 4, 2016 11:21 |
|
I've become obsessed by that idea with the electromagnets today. I mean it is an idea I had a long time ago, but I never sat down and figured out how it would work practically, I just expected someone else would come up with it or something better. I had a bit of a deeper look into electrically actuated synthetic muscles today, and most of them are not that great or vapourware like the third video I linked. If electric muscles are all that is missing, hold on to your gonads, cause I'm gonna make one out of everyday household items. The robots are coming. Here's how it would work with the magnets wound as bars rather than coils, cause I think coils might sooner twist in their sockets than stretch the muscles: https://gfycat.com/BlankWhoppingKinkajou
|
# ? Aug 4, 2016 11:29 |
|
monkey posted:I've become obsessed by that idea with the electromagnets today. I mean it is an idea I had a long time ago, but I never sat down and figured out how it would work practically, I just expected someone else would come up with it or something better. I had a bit of a deeper look into electrically actuated synthetic muscles today, and most of them are not that great or vapourware like the third video I linked. If electric muscles are all that is missing, hold on to your gonads, cause I'm gonna make one out of everyday household items. The robots are coming. government electromagnet mind control
|
# ? Aug 4, 2016 12:48 |
|
8 track betamax posted:Can somebody explain to me what purpose human shaped robots would serve? Besides weird sex stuff Clean my room
|
# ? Aug 4, 2016 13:33 |
|
GRILLARY CLINTON posted:since you offered, please answer my easily googlable questions: TROIKA CURES GREEK answered most of it, but I'll go a bit "deeper" Warning, effort post Neural Networks are function approximators (i.e., they approximate functions ) Let's say we have some inputs, x, and just to have some concrete example, we can say it is the pitch of a plane or something. We want to model how the pitch of the plane affects lift, but it isn't so simple; it's some complicated function. Look at a) and pretend it is our data, where x represents input (x can be a set of values - maybe [roll, pitch, and yaw]) and w is some output (which again could be more than one value). We want to be able to predict the w based on some new value for x. Neural networks work by taking a bunch of functions like in b) and weighting them like in c) and then adding them together to make some function in d) which seems to model our data. So you give the network a bunch of (x, w)-pairs, it gives you some different w', train it based on the difference between predicted w' and training w until it gets good predictions, and then start feeding it new x's and it'll make you some new w's. As Germstore mentioned, there are a ton of add-ons and extensions (like recurrent neural networks :P), and as TROIKA CURES GREEK mentioned, the powerful thing about the network is that we can have it find features that predict w well. In the above description, we had to feed it specific x's, with defined features (like roll and pitch or maybe windspeed or something), but new algorithms can find these features. So you can feed it say a model of an airplane and it can learn what features are important on its own (so maybe it learns that airspeed isn't so important, but angle of attack is, or something). But they can be incredibly slow to train and can need a ton of labeled data. Google trained some arms to pick up objects with a neural network, and it took several months. http://www.theverge.com/2016/3/9/11186940/google-robotic-arms-neural-network-hand-eye-coordination Still, they are pretty powerful and I'll be trying one out as one approach in my thesis. We'll see how it goes. As far as what makes a network a deep neural network, when most people talk about neural networks, they mean those (and that's what I have described). A non-deep neural network looks like this (here, y == w): and it just finds some straight line to divide the data. cf. multilayer neural networks / deep neural networks: How it relates to real biology: Well, it is inspired by it, but an extremely scaled-down version. The original idea for each technical neuron was if x > some value, fire a 1, otherwise a 0 - mirroring the idea of action-potentials in real neurons (they are far more complicated now, but still nothing like a real neuron). In reality, neurons are quite complex and the entire process of how data works in the brain isn't completely understood. You've got all kinds of hormones etc too. Each neuron is more like its one mini-processor, and there are billions and billions of them. monkey posted:slaves To answer your question about robotic limitations etc., there are a lot of factors: 1) Unlike robots, you are composed of trillions of micromachines, each capable of self-healing and thousands of tasks that operate at a chemical level. If a part of the robot breaks, it'll never heal on its own, and it will eventually wear out, while a human is constantly replacing these small machines called cells. 2) You have all kinds of sensory feedback. We often hear we have five senses, but we have way more. If you put your hand behind your back, you can move your fingers or have someone move them for you into all kinds of configurations and know exactly what shape your hand is in - the so called kinesthetic sense. https://en.wikipedia.org/wiki/Proprioception Robots don't have this. Also you might not notice, but your movements aren't precise. E.g., when you go to look at something, you usually overshoot the target and then automatically adjust. https://en.wikipedia.org/wiki/Saccade Robots also move imprecisely. You can give your motor some command, but it may over- or undershoot. A lot of robot learning is about compensating for this noise. Sensors are also very noisy. 3) Tasks you consider incredibly simple, like walking or cleaning, are actually tremendously challenging. Animals have had millions of years to develop software and hardware that handles it all for you, so you don't notice, but it's very hard. Especially with less precise sensors than you have built into your body. But even with all the advantages, it still takes human children quite a few years to learn how to walk. There are a lot more things I could say too, but I think this is already too much. 8 track betamax posted:Can somebody explain to me what purpose human shaped robots would serve? Besides weird sex stuff Robots are often human-shaped because most things we build and design are for humans. Doors, stairs, etc, are designed with human anatomy in mind. (Also, yes, sex [and probably human hubris]) Oh, and how neural networks are different from other machine learning algorithms: the good/bad thing about neural networks are that they are a bit of a blackblox. You toss data in, it does something, we get a number out, . Shadow0 fucked around with this message at 14:27 on Aug 4, 2016 |
# ? Aug 4, 2016 14:01 |
|
Thanks you for that effort post, very interesting. The only criticism I have is to please stop describing organisms as machines it is anthropomorphic and wrong.
|
# ? Aug 4, 2016 14:08 |
|
Advanced enough for robotic tricks? Advanced enough to pull my dick.
|
# ? Aug 4, 2016 14:22 |
|
It's gonna be tough to gently caress it with it's arms flailing around like that.
|
# ? Aug 4, 2016 15:08 |
|
Shadow0 posted:To answer your question about robotic limitations etc., there are a lot of factors: Thanks for the reply. I think a lot of what you mention is easily addressable though. 1. Sure parts may wear out, but a robot is far easier to fix with surgery than a human. Finger broken? replace it with a new one off the shelf. Even self repair on a macro level would be possible for some things. 2. Proprioception is not all that hard to emulate is it? I know from dicking around with VR hardware there is a runaway positional error with IMUs, but rotational info from gyros is pretty solid and fast. A couple dozen gyros on the major joints would do it. The distances between sensors is fixed and known, so an IK solver could build a pretty reliable model to make a closed loop as far as proprioception goes. Other sensors are far more accurate than humans (pressure, tension, heat, altitude etc) So again, I don't see the limiting factor there. 3. Cleaning I agree is a tough one, there's a whole lot of stuff going on to do something like dusting the glassware, but to take the example of walking... sure you could take the purist approach like they have in the bot in the OP and give a neural network nothing to go on and see what crazy method of locomotion it develops by itself, but if you're building an exact replica of a human structure down to each muscle, you'd have a shitload of data at your disposal. You could feed it 1000 walk cycles recorded from actual humans and use those as the starting points for fitness testing, take advantage of the same under actuation that we do when walking, similarly you could record people completing an obstacle course or dancing or whatever as seeds for learning. If we want to emulate humans properly, we need to emulate human movement and the most direct way to do that is to copy the physical structure as closely as possible. Back to the electromagnets thing, I was perhaps overthinking it a bit, it'd probably be more efficient to just daisy chain a bunch of solenoids in a flexible tube, though you'd lose the ability to push. Or you could run a single line of switchable electromagnets along the tube end to end with soft silicon between them.
|
# ? Aug 4, 2016 16:26 |
|
monkey posted:Thanks for the reply. I think a lot of what you mention is easily addressable though. Uuuhhhhhnmmmm. All those electromagnet will erase the hard drive disks in the robots brain!
|
# ? Aug 4, 2016 18:11 |
|
Beefeater posted:It's gonna be tough to gently caress it with it's arms flailing around like that. The first 100 years of sexbotics is going to see a lot of torn off dicks.
|
# ? Aug 4, 2016 18:11 |
|
8 track betamax posted:Uuuhhhhhnmmmm. All those electromagnet will erase the hard drive disks in the robots brain! SSDs man!
|
# ? Aug 4, 2016 18:34 |
|
King Vidiot posted:The first 100 years of sexbotics is going to see a lot of torn off dicks. Can't make a sexbot without tearing off a few dicks
|
# ? Aug 5, 2016 19:56 |
|
|
# ? Jun 8, 2024 18:26 |
|
Obligatory Mr Fisto comment.
|
# ? Aug 5, 2016 20:44 |