Basically, when deep reinforcement learning works, it's like magic, but unlike supervised learning where the default expectation is that it works without a hitch right out of the box, the default expectation for most new tasks through deep reinforcement learning is that it will fail, and you will need something to fix it.

For example, the high dimensionality of robotics makes it very difficult to apply deep reinforcement learning to it, although it definitely can and has been applied (and is, IMO, the future of robotics).

Another example is that simple supervised learning often outperforms DRL for many arcade games.