L478

[studywolf]

https://github.com/tcstewar/2015-Embodied_Benchmarks/blob/master/paper/paper.tex#L478

This line is about the controller, not about the accuracy of the simulation, right?

Suggested

"Of course, we need confirmation that a controller trained on the minimal simulation defined above will work on a real physical system as well."

[tcstewar]

Yup. And I think this is a good opportunity to also slip in a comment to the effect that when we test on a real physical system, we can only test on one (or two, or a few) physical systems, just as a practical matter. So the benchmark run in simulation is actually more useful, since it covers a larger space of possibilities. But we still want to run on a physical system as well, just to see. But the results from running it on that physical system only really tell you how good the controller is on that one physical system.

[tcstewar]

Hmm, and one slight rephrasing. I'm not sure I want to think about using the minimal simulation as a way to train a controller. That's slightly different, I think, from using a minimal simulation to benchmark a controller. But I want to think more on what that difference is....

[tcstewar]

I ended up reworking this and adding a whole new paragraph:

Of course, we also want an indication that the minimal simulation defined
in the previous section is reasonably representative of the sorts of real-world
situations in which we might want to use these controllers.  Importantly,
this physical instantiation does not have to exactly match one particular
parameter setting of the minimal simulation.  Rather, we want a physical system
that shares basic functional similarities to the minimal simulation defined
previously.

For example, we want the inputs to the system to act like $u$, in that a
positive number will increase some velocity $v$ which will in turn increase some
sensor value $q$.  We want there to be some sort of external force applied that
affects $q$, and we want that external force itself to be a function of $q$.
We want there to be communication delays and noise in the sensory and motor
system, and we want all of these effects to be somewhere within the extreme
ranges covered by the minimal simulation.  While this cannot prove that hardware
that is successful in simulation will always be successful in any similar
real-world task, it at least gives an existence proof that there is at least
one real-world task where it also performs reasonably.