I'm a contract researcher in manufacturing robotics with a mechanical engineering degree, do you guys think its worth learning neural nets? My programming is ok.
I'm a contract researcher in manufacturing robotics with a mechanical engineering degree...
Dominic Miller
Zachary Cook
Neural networks fall inside a family of algorithms and techniques called machine learning. There is also a more recent subset of algorithms and techniques called deep learning that focuses on a class of algorithms known as deep neural networks.
Machine learning in general is a topic that lives at the intersection of artificial intelligence and data science (depending on who you ask they will classify it as one or the other).
It may be useful to you to learn some basic machine learning but it is probably more useful to get acquainted with AI in general.
I recommend you use those free online courses first to get a general overview and then pursue it further if you find it interesting or useful.
Andrew Ng used to teach a machine learning course at Stanford. There was a theoretical mathy version and an applied engineering version of the course. Back in 2011 he taught the engineering version online and shortly afterward he founded coursera with that course available. It is still available and in my opinion the best introduction to the topic out there.
The Udacity guys also started at Stanford at around the same time and their first course was an intro to artificial intelligence. It is also still available and gives a good general overview to the ideas, techniques, and concepts of artificial intelligence.
As for deep learning. Don't worry about it for now. Basically it's machine learning on steroids and it's at the bleeding edge of the cutting edge. Look into it if you find machine learning interesting or useful.
Hunter Perry
Robots have little to do with neural nets.
Unless you're working on something weird, like a mail sorter.
But go ahead. It's fun an relatively simple.
>deeplearningbook.org
Colton Barnes
i think in robotics they have been using a lot of temporal difference learning which is part of reinforcement learning. if im not mistaken this is how the atlas robot from Boston dynamic robot learned to walk. (educated guess) there might be much more ways also using neural networks or combinations.
Benjamin Young
Yes actually, you if you're also responsible for for control it's quite helpful, in fact it's surprising that you haven't used it so far
Joshua Thompson
they have a lot to do actually, you can learn behavior to reach certain targets.
Christopher Thompson
Atlas did not learn to walk. BD doesn't use much machine learning. Walking algorithm Atlas uses is just zero moment point.
Pic related, there is a horrifying monstrosity inside that thing.
Robert Kelly
ME here(I think), you should fuck around with neural networks. The CS guys have made some cool libraries that make it possible for any idiot to fuck with neural networks.
Right now I'm fucking around using neural networks as a convenient means of representing structures so I can evolve them with a GA. It's fucking stupid and inefficient, but computing power is cheap!
Aaron White
Neural nets are incredibly easy if you know even the littlest linear algebra, so - go for it.
Hudson Allen
it's trivial really
just some matrix arithmetic and deriving
Tyler Martinez
why do they make humanoid robots?
a spider design seems incredibly more useful
Josiah Parker
how do you know this?. they keep it secrete how it work
I have my reason for believing what i said. since people have been using walking algorithm for ages now in Asian countries and the best they can get is those stiff robots that can almost walk up a staircase.
Then some small company from boston ( near MIT) comes and revolutionize the field and make this perfect robots that walk almost exactly like humans. there is no way to preprogram how atlas walk. MIT also has been teaching this machine learning techniques for robots. for me its clear that there is a connection there.
what do you think?
Brayden Nguyen
they are based on calculus since its basically performing stochastic gradient decent to learn, but most people don't really understand how they works, they just multiply matrices like you.
John Cook
Neural nets won't really help with robotics, aside from computer vision. But on the other hand, the basics of neural nets can be learned in an afternoon (gradient descent with backpropagation).
Wyatt Baker
I know ressearchers who use machine learning to teach robot hands complex grasping patterns.
They gather data with a glove that measures joint angles on a human, perform the task multiple times, apply a learning algorithm and then use it for the robot hand.
Same could be done for walking. Let humans walks, measure joint angles etc. then transfer to robot.
Austin Murphy
the same researchers also have a robot arm with weird joints that can basically be modelled as having a joint at every point on the arm.
Using inverse kinematics for that would be a pain in the ass.
Instead the robot explores and learns it's own work space and how to reach points in it with machine learning.
Jaxson Perry
this
i don't get why retards on sci that have no clue come into these threads without knowing shit. saying huge shit claims like
>Neural nets won't really help with robotics
its laughable
Gavin Roberts
"i saw neural nets on youtube once for computer vision and thus i declare that is their only application."
aka "i know nothing but i think my opinion matters"
Joseph Bell
Because that's not the approach Raibert uses for his robots. IE build a state machine that cycles through different simple controllers. I would recommend reading the publications for the robots linked here:
ai.mit.edu
Most of the MIT leg lab people went to found Boston Dynamics.
The only advantage boston dynamics has over those asian robots is that they use hydraulics which allows them to react faster.
>> MIT has also been teaching machine learning techniques for robots
Citation motherfucker. They certainly didn't use much machine learning on their DARPA challenge Atlas.
Oh look it's their github for some of the software they wrote:
github.com
You gonna post a citation on that bucko?
>> using inverse kinematics hard
HAHAHAHAHAHAHAHA
We can calculate the IK of systems with huge numbers of joints in parallel and series faster than real time. It's a lot easier to do IK than it is to train a neural net to do the same thing.
Elijah Collins
Not that guy, but
>We can calculate the IK of systems with huge numbers of joints in parallel and series faster than real time. It's a lot easier to do IK than it is to train a neural net to do the same thing.
Absolutely correct, yet there are two classes of problems. In one case, you know very clearly what the end result you want is and must calculate the mechanical inputs to satisfy that result. In the other, the end result is more of a fitness goal, and you want to be potentially surprised by novel solutions to achieve that goal. Neural networks are for the latter case. You could mix them, but in many cases it would be unnecessary complexity.
Brandon Gonzalez
Good luck convincing the industry that they should use that stochastically trained crap (99.9% recall rate dude!) instead of the deterministic inverse kinematics.
>Oops, my robot arm killed a worker! Maybe we should have used 32 hidden layers instead of 31.
Dominic Lewis
i've been looking through your links. i cant find anything that is referring to the Atlas and how it walks. a static programmed algorithm for walking will ofc not work in the long run over all sorts of terrains. the links you gave example the github do have code for the kind of reinforcement learning i wrote about which is interesting.
Brayden Turner
Yeah well good fucking luck making something that is better than IK, it has been tried. No it's not unnecessarily complex, now you can just output a 3d vector from your neural network instead of joint angles and do more interesting shit than moving an arm to point A.
plus that way it generalize to robots with different kinematics.
Owen Baker
>Oops, the robot arm I was programming was capable of reaching workers! Now the designer is being sued! Good thing that's not my problem unless I was involved in the design and that stupid.
Not every machine in manufacturing has to shake hands with each worker before doing its job.
Chase Allen
>Oops, the robot arm I was programming was capable of reaching workers!
Well, just change it to
>Oops my robot arm destroyed thousands worth of material
Jacob Cox
Here m8
drc.mit.edu
Lucas Carter
this is how people that learned something the old way will try to grab on to his knowledge and refuse new techniques. im sorry your expertise in IK is not needed no more.
Henry Anderson
>Yeah well good fucking luck making something that is better than IK
Did you read? Different is not better, it is for completely different situations.
>No it's not unnecessarily complex, now you can just output a 3d vector from your neural network instead of joint angles
You don't seem to understand what neural networks do. There is absolutely no fucking benefit to putting your input through a process with a cost of O(n^2) before adding constraints. Thus, unnecessarily complex. For much the same reason that we do not compute inverse kinematics to move our limbs.
Sebastian Jenkins
>Oops I'm still a shitty designer with no clue what I'm doing, it's a miracle I got a degree
Please stop with the rhetorical bullshit, it's clear you have no idea what you're talking about and are thinking about this for the first time today.
Luis Nguyen
i don't get it. why bother giving the robots mechanical engineering degrees?
Michael Smith
thats the last thing we will ever do
Jonathan Scott
With incompetent people you know who's accountable, while with NN you incorporate possible failure as normal operation
Asher King
look at what lasa does
Ethan Lewis
salut fellow epflien
Jace Torres
salut jean michel
Jaxon Garcia
Do you even know what a neural network is, how one works, and, perhaps most importantly, what an appropriate application of one might be ? I'm thinking not. You really shouldn't spout off about things you don't understand. It makes you look like an idiot.