Build (various?) awkward shaped robots and make them learn how to move using reinforcement learning. Something like this. You give it some joints, you have some sort of roller that gives positive reward when it's rolled forward, and you use pretty much standard RL.
