Any hardware yet?
ROS is on my bucket list, too, started to learn about it a couple of years ago, after accidentally taking a uni classes (from YouTube) about autonomous mapping (SLAM -
https://en.wikipedia.org/wiki/Simultaneous_localization_and_mapping - not sure if it was this exact playlist
or maybe it was some other SLAM courses on the same channel, but from a more recent year).
Meanwhile nVidia has a very cheap platform, nVidia Jetson nano, ~$75 total, with hardware acceleration for AI/vision kind of stuff. Their robotics platform "Isaac" can work with ROS (random example
https://developer.nvidia.com/blog/building-collaborative-robotics-using-ros-and-isaac-sdk/ ). I expect such a setup to be a 10-100 times faster than, for example, a RaspberryPi (that does not have any AI hardware accelerators).
About going from learning only to performant robots, I can tell you even the smallest robots can cost as much as a car. I know this from a former coworker that built a sumo robot for competitions.
His university was heaving an annual sumo robots competition at the national level. I was (as a spectator) at one of these. The event was half a day long, and I thought this would become repetitive and boring pretty fast, but in fact it was a fascinating experience.
Some robots were moving so fast that the phrase "if you blink it, you miss it" was for real. Low inertia motors, special tires, rare earth magnets underneath the robots to increase the apparent weight and thus the tires friction (the sumo stage was an iron platform), battery packs and drivers capable of hundreds of amps, low latency sensors, and so on.
And the show was also great. The robots were autonomous, so the algorithms used to detect the other robot, push it and evading from its attacks was playing a big role. Some were acting dumb and funny, some were fierce. A few robots beat themselves by slipping outside the arena because of their own inertia, others went in smoke, other were disabled by impact, and so on.
The funniest thing was that the 3rd place was conquered by a team not so experienced. Their robot was in fact bought as a kit instead of using a custom build, and their main contribution was to put on it a nicely colored skirt/cover. The team didn't work much on the software side either, and while others were doing very complicating maneuvers, their robot was just blindly turning in circles, and that was bamboozling most of the other robots, because they were assuming much complicated strategies.