Hierarchical Reinforcement Learning is suited to tackle many complex problems such as robotics manipulation. Sticking to this example, the basic idea is that instead of learning a sequence of robotic joint motions, we can learn a sequence of high level actions (e.g. grab, release, flip, etc.) and low level implementations of those actions.
Yet, I do not see much work being carried in this direction and I don't know of any impressive success story. What are the current limitations stopping hierarchical reinforcement learning to be more widely applied?
My current guess is that Hierarchical Reinforcement Learning requires some manual problem-specific work in building the high-level state space, which limits the scalability and reuse of the solutions. Still, this does not feel to be a serious issue.
