"Second, [the final training algorithm] adds Simultaneous Localization and Mapping (SLAM) maps, which robots use at execution time, as a source for building the roadmaps. Because SLAM maps are noisy, this change closes the “sim2real gap”, a phenomenon in robotics where simulation-trained agents significantly underperform when transferred to real-robots."
Very nice work. They are using simulated environments to train the navigation system (in a very sample-expensive regime already), and on top of that, going to the trouble of using the noisy maps from SLAM in the simulator, to mimic real noise characteristics.
Although I have done work in this domain, I had not heard the "sim2real" term before. It sounds like a generic phenomenon for systems trained on models but later used in the real world.
This is no doubt exciting, but I am curious how well they deal with unexpected obstructions (such as cars at intersections behaving in unexpected fashion). Remember, we have a lot of issues with self-driving, and the kind of problems (not quite handling environment correctly) are translatable to this use case as well...
