I hooked a neural network up to my Roomba. I wanted it to learn to navigate without bumping into things, so I set up a reward scheme to encourage speed and discourage hitting the bumper sensors.
It learnt to drive backwards, because there are no bumpers on the back.
It learnt to drive backwards, because there are no bumpers on the back.
This is Fenton the Roomba. He's a Roomba 650 with a Raspberry Pi 3b brain running my own neural network code on Windows 10 IoT. Seen here bouncing gently off a table leg. 
I wish I'd taken more video before I took him apart for a rebuild. This is Fenton exploring a corner after a couple of hours of exploring. He's learnt to slow down and turn away when he approaches a wall, but he still gets a little stuck in corners.
