Robotics is really broad field, and just by being a software engineer and also with your computer vision projects you've likely already started building a portf

Robotics is really broad field, and just by being a software engineer and also with your computer vision projects you've likely already started building a portfolio that a roboticist will be impressed by. I think that building a physical robot is extremely valuable to anyone who wants to get into robotics. There's so many problems to solve and things to learn that just won't come up in simulation. That being said it might be better to first build your robot in simulation first, which is also extremely valuable to be able to do. So, if you want my suggestion I would try and do all three of the things you listed. Pick a problem that sounds easy to start. And try to build a robot in simulation to do just that. Having this simulated environment set up will be invaluable for future endeavors too. Once you finish with that I think it will be extremely valuable to actually build your robot and try out the software there. Almost certainly it won't work the first time but that will be a valuable lesson too. Another note, walking robots are in general very difficult to do well, quadcopters are more straightforward but still much more involved than the car to build from scratch. The car will be easier to get going, and will be more forgiving mechanically and with the controls. But that means you can quickly get into doing the higher-level motion planning if that's what you're interested in, and there are still interesting control and estimation problems to do there. And believe me even though it sounds very easy you will run into a lot of challenges you probably are not expecting.