So I bought a Petoi robot dog for my dad this Christmas, and we have had a lot of fun playing with the thing, but on top of that I realized that I can use python to send serial commands to its arduino over wifi to directly control motor movement, and this has really blown my mind wide open.

I had to go fishing through a storage unit to find an arduino kit that I had bought to reflash the firmwareof my 3D printer long ago, but I never realized how powerful this little board actually is. Seeing how easy it all is, robotics I mean, I am now thinking that it wouldn't be too hard to train a neural network for balance and motion using RL in a simulated environment, then do the final training stage on the real robot that is apparently very easy to build with some cheap motors. Of course an arduino memory is very small, but apparently a R.Pi can be used for such applications and I have one of those in my storage unit too.

So fun times abound. My robot idea is a single arm that can balance upright. Maybe I can get it to hop around. I'm sure some R&D will be needed in both the robot and NN architectures.

Anyone else dabbling in such things?
What are you working on?
Any pitfalls I should watch out for?
Servos or steppers?