I wanted proper hands-on experience with reinforcement learning, so I gave myself a hard, very visual task: bounce a ball on a plate held by a robot arm.
The simulation runs in MuJoCo using the official MuJoCo Menagerie robot models. I wrote a Gymnasium wrapper around it, trained policies with Stable-Baselines3 and added domain randomization so the policy has a chance on real hardware.
It hasn't touched the real robot yet. That's next.