> Let's say I would like to be able to train a robot to "Open the fridge and take out a beer". That would be something "cool" and guess possible with current technology. Especially after looking what boston dynamics can do...

Well yeah, but you realize that BD has at least 10s or more of engineers work on such their projects.

I think with your background I would opt for a pre-made platform (there are many, don't have experience with any of them) where you worry mostly about the software.

Maybe, just use your laptop webcam or a Raspberry Pi to do computer vision, somewhat related to your DS/ML experience.