You decide what to reward and what to punish. The AI learns how to walk. Train legged robots with reinforcement learning right on your PC's CPU, then pilot the AI you raised in obstacle races and sumo bouts. Keep playing while it trains.
Anime spinning-top battles in an endless underground passage. If something is wrong, turn back. If nothing is wrong, beat the blader who waits and walk on. Spot the anomalies, win the battles and reach the exit 8... if it exists.