Author here, was pretty surprised to see this on HN when browsing over my coffee this morning. Your interpretation is correct, you use an encoder-decoder model to figure out what the dimensions of the task best for learning are.
The drawback is you can only learn tasks which are relatively similar (any time you restrict what motions are possible to improve learning, you obviously restrict what tasks are possible). The benefit is that you can learn tasks which do fall within the learned motion ranges a lot more quickly.
The best analogy within 'classical' control is task space control, where you do control in cartesian dimensions rather than the joint positions. But this has its own drawbacks in that you have to define these controllers manually, and Cartesian space is not sufficiently expressive / appropriate for many tasks.