Arduino Neural Network Robot Part 3: Running Neural Networks on an Arduino

This is the final part of a 3 part series where Sean Hodgins is designing an open source Arduino based robot that will navigate using a neural network. Part 1 involved taking components to make a prototype for testing. Part 2 took the custom circuit board and populated it with components, adding some test firmware to make sure everything worked. Finally, in this video, Sean will discuss different methods of making the robot navigate using the photoresistors, and how the neural network works for navigation. By the end of this series you will be able to create your own neural network robot using the resources provided.

Instrumentals Produced By Chuki

[SeanHodgins]

I talk quite a bit in this. More than any other video I have made. Hopefully what I'm saying makes sense to everyone! Lets see more Arduino neural network robots out there!

[lucassilvapascoal]

I really liked this project, I've studied neural network but I have not seen how to make it practical, that is great, I just love how Arduino and Raspberry Pi with this great community can make very difficult thing to really happen, I am an electrical engineer and nowadays I am seeing a lot of unthinkable stuff being done by regular people because of this strong community knowledge that is being shared.

[Antbiotix]

Just discovered this video! Great stuff. It’s neat to see Adruino doing the heavy lifting of training a NN, although to save time you can pre-train your NN in a more powerful PC and just copy and paste the weights into your adruino code.

[rverm]

i saw this video a few years ago and thought it was cool. now that im taking a deep learning ai course it makes the video so much cooler. im going to make videos on neural networks and arduino

[harmonic]

A solar charged version would be quite cool. One that finds the light rather than runs from it.

[nitanshnagpal]

Very good explanation You bridge the gap of theory and practical

[joebarhouch]

Since u started doing this mini series, i watched every single one of ur videos sean! Ur truly an amazing guy keep up the good work and i hope i could be someday like u +Sean Hodgins

[hiddenotebook]

This is a great job thank you. One robot with ultrasonic sensor will be nice proyect for me. Thanks to you I'm starting to understand this

[Bereft]

That program was elegant! Thank you I need this for a touch array in order to change balance.

[blkfngrs]

Best ml example explained, period. great work!

[springwoodcottage]

Nice series, interesting in that the neural net which I thought would be slower especially on the low spec micro controllers was fast & reactive to the light. Kinda reminds me of least squares fitting to a curve in its operation. Thanks for sharing!

[bogdanpashchenko]

So... would there be a part 4 with something that could be done with NN but can’t be done with own algorithm? Kinda itching to see that.

Thanks for this video. With that idea of NN being a mapper of inputs to outputs I finally starting to gain understating of what a NN is.

[streetpunkstar]

Could you train a NN off the Arduino using whatever (something with a GPU!) and preload the weights onto the Arduino? :)

[nikodembartnik]

That’s great project and super detailed explanation! I really like it :) I made a lot of Arduino projects and learn some about neural networks and genetic algorithms I am thinking about implementing neural net in a line follower project. I know that PID algorithm can be simpler faster and better but just for scientific purposes and because of my curiosity I want to try that. What do you think? Thanks again for this series!

[azizkavas]

It was excellent work sir. Thank you. However, can you share power consumption statistics between first and trained version. Besides, how long it was working with a certain charge. Sincerely.

[VLS-Why]

Does a Neural network kind of work like an automatically tuned PID function?

[Nono-hkis]

Great project. I really appreciate that you laid out the entire process, but also got down into the details. This was perfect intro to neural nets for me. Thank you.

One thing that wasn't clear: during training, the code doesn't actually look at sensors, correct? If so, could the process be reorganized so that the training were done on the host computer rather than the arduino?

[kmachine]

What is the minimum system requirements for this kind of projects? Thanks for answering.

[himselfe]

I've been thinking for a while it would be fun to build a neural network with AVRs, but instead of doing it all on one chip, use a bunch of chips (maybe those super small ATTinys), with each chip acting as a neuron or set of neurons.

[planktonfun]

more hidden layer = more layers = more number manipulation (basically each layer is add, then multiply) = more closeness/percision.
basically more hidden layers, the more complex logic it could handle.
for example:
summation = no hidden layers (its to simple you just add)
multiplication = no hidden layers
powerof = at least one hidden layer
pythagoran theorm = at least four
complex equations = more hidden layers