I was very interested in athletic robots, so I left my job in software to build a quadruped and make it backflip [1]. I made the hardware at home (the design was open-source [2]) and wrote the framework to generate jumps, flips, etc. I started with basically zero robotics knowledge, so it was a lot of work, but it was incredibly fulfilling and super fun. I personally learn best by building things.

I'd say start playing with 3D printing and motors to make mechanisms that move (pendulums, simple walkers, ...). +1 to tinkering with the 3D printer itself (that someone mentioned) as you can learn a lot about how the machine works and figure out ways to improve it.

You've mentioned the physical aspect, but simulation could also be a good starting point. You can now very easily train RL agents in sim to say, walk, and transfer to real hardware afterwards.

Finally, if you like flying stuff, there's drones you can buy that are controlled programmatically (e.g. Crazyflie). I haven't played a lot with these, but I've seen them used to teach control and they seem like a nice, inexpensive platform to experiment with.

[1]: https://github.com/rkourdis/trajopt?tab=readme-ov-file#backf...

[2]: https://open-dynamic-robot-initiative.github.io