Digitalopus

https://www.youtube.com/watch?v=Z-H3EuWYY8A

This has come a long way since my first attempt to train the guy to walk. Over the Christmas break I trained a tumbling brain so he can get up quickly if he falls with some velocity. Next steps: try to get this working with moving and accelerating platforms.

I have always wanted a keyboard controlled, fully physically simulated powered-ragdoll that would walk, run, jump, stand using forces and torques (just like a real world person). With the accessibility of deep learning APIs, this appears to be becoming possible. The current version can stand, walk, run, getup (not very robust yet), jump (do we really need to be able to land the jump?). Still very much a work in progress prototype, but it is almost usable.

Learning to Jump!