> If you have a very hard problem for which you have a simulator, our results imply there is a real, practical path towards solving it.
Are there that many domains for which this is relevant?
Game AI seems to be the most obvious case and, on a tangent, I did find it kind of interesting that DeepMind was founded to make AI plug and play for commercial games.
But unless Sim-to-Real can be made to work it seems pretty narrow. So it sort of seems like exchanging one research problem (sample-efficient RL) for another.
Not to say these results aren't cool and interesting, but I'm not sold on the idea that this is really practical yet.
There seems to be a bunch of work in this area, but I have no idea how you measure progress in this area, it's not like you can do evaluations on a shared task.
And it's clearly not solved yet either - 76% grab success doesn't really seem good enough to actually use, and that with 100k real runs.
I don't really know how to compare the difficulty of sim-to-real transfer research to sample efficient RL research, and it's good to have both research directions as viable, but neither seems solved, so I'm not really convinced that "just scaling up PPO" is that practical.
I'm hoping gdb will be able to tell me I'm missing something though.
Are there that many domains for which this is relevant?
Game AI seems to be the most obvious case and, on a tangent, I did find it kind of interesting that DeepMind was founded to make AI plug and play for commercial games.
But unless Sim-to-Real can be made to work it seems pretty narrow. So it sort of seems like exchanging one research problem (sample-efficient RL) for another.
Not to say these results aren't cool and interesting, but I'm not sold on the idea that this is really practical yet.