Это видео недоступно.
Сожалеем об этом.
Arduino robot controlled by Neural Network
HTML-код
- Опубликовано: 2 дек 2009
- My first attempt in neural networks. I connected 1 adruino mega, 3 servos, 1 analog infrared receiver, 3 contacts and 2 encoders from an old printer. I use a genetic algorithm that trains the 5-7-2 neural network (5 inputs, 7 neurons in the hidden layer and 2 in output). The 5 inputs are the 5 areas that the IR scans using the third servo, and the 2 outputs are the speed of each wheel servo.
Enjoy :)
mpravo re magga...
exeis arxisei kai ginesai professoras...
Awesome work, keep it up. I'm not going to ask what code you used because if their's one thing I've learned its to do it my self and it feels better that way.
very clever - and amazing music!
well done! Been thinking about doing this myself. Did you code your neural network from scratch or interface it with matlab or something like this?
@MadTown314 It's Amon Tobin. I don't know if they used it for the soundtrack of the film
Awesome work bro
@yamandudacosta you can train neural networks using genetic algorithms.
fascinating
So, I understand how back propagation works to train neural nets but how does this work? How does the robot know that its desired output is to avoid running into objects?
Maybe the genetic training algorithm uses a score based on the wheels speed and it tries to maximize the speed, so it learns to avoid things that cause a speed drop (e.g. collisions).
This is not back propagation training. There is a genetic algorithm that trains the neural network. The network is a 5-7-2 and the fitness function is:
F=V*(1-sqrt(Du))*(1-i)
V, is the mean speed of the wheels.
Du, is the speed difference between the wheels.
i, is the distance of the nearest object in the IRradar range.
So the robot learns to avoid objects and run as fast it can but not in circles.
planmix I see. I understood that it was not back propagation. I was just confused because back propagation (my only exposure to neural net learning at the time) is based on matching a given output to a given input... but this obviously doesn't know what the desired output... it just wants to approach some sort of ideal input... I think... that being that there is not something in the way...
Μπραβο Planmix ... πολυ ωρεα κατασκευη , αν μπορεις ποσταρε μια παρουσιαση του ρομποτ (λιγες φωτογραφιες και λιγα λογια) στο Grobot... Και Παλι μπραβο πολυ ωραιο...
A NN can learn. You show it some examples of correct behavior and it learns to generalize to new situations. GA is an optimization technique which can be used to train the NN.
To improve reading put your sharp IR ranger vertically if you scan horizontally.
amazing! i'm currently making an analogue neural network trained robot in the form of an orangutan!
im new to robotics , so can u explain(even a bit) abt what this neural and genetic algorithms are? ive worked only with normal robotic stuffs where based on certain conditions it takes some actions (sensors and stuffs alone with other component's) so whats the difference between this and the one used here?
awesome!
hi there i am trying to do something similar for a university project but im having difficulties. how did u implement your architectural?
Many thanss
Will you be releasing the source code for this at some point? I think a great project for getting people into robotics would be to make a robot that isn't just "remote controlled", but to build something that can learn. I think using some shield that exist now for driving motors and servos, a kit could be put together for pretty cheap.
the objective is to make a robot that self teaches to go around objects. thought that was kinda clear
what was the robot doing after 6th minute that it didn't do before???
impresionante entrenas la red neuronal con algoritmos genéticos el resultado es excelente. lastima q es tan complicado de implementar amazing :)
Actually, genetic algorithms are pretty easy once you get the hang of them.
HI i am trying for the same can i have more info about this project thank you
Call me crazy, but this is the music they used in the Italian Job,
is it not?
Nice, but I think genetic algorithm is not the same as neural network.
Απιθανο!! Ο γενετικος κωδικας μεταλλασεται μεσα στην μνημη του ρομποτ ή με software στον υπολογιστη ασυρματα (bluetooth)??
Μπραβο!!!! Αντε και με ποδια την επομενη φορα!!!
Στον arduino γίνονται όλα :)
I can't access the link. May i ask for it? Thanks you
virtual brain makes it a virtual mice :D
You used the real robot to evolve an artificial intelligence solution. To spare the real robot to some stress you can use a computer model, and then copy the neural network to the real robot. This is offline training. But you can also have such a virtual model in real time, where the robot tries alternatives in the virtual world before applying them.
You can check am example in assemblino com » Database » public » Tests » Neural1_RoboticArm
By any chance, could you give me that code? I have almost same setup and i really want to make it
hey bro , did u make it finally ? i need that code so bad :p
can you share the code....i'm learner of neural networks for robots....it will be very much helpful for me
Im very interested in the source code to study it, can you sent it to me?
Why not just make an obstacle avoiding robot by using an hc-sr04 (ultrasonic sensor) and tell it that if an object is too close, turn away.
@arminask not banging into stuff
?! What was the objective?
a good robot doesn't need any switches
POST THE CODE! AMAZING!
chromosomes? you mean neurons?
bro so gd can u email me the source code
This is not legit, the video speaks of long training of chromosomes, yet the description speaks of a nn with only 7 in the hidden layer - which would be pointless to even train since you could just choose the outputs based off k-mapping in a minute or so on paper. Something does not match up here.
wait...why does the video says chromosomes?
chromosomes? neurons don't have chromosomes!