What is the common representation used for the state in articulated robot environments? My first guess is that it's a set of the angles of every joint. Is that correct? My question is motivated by the fact that one common trick that helps training neural nets in general is to normalize the inputs, like setting mean = 0 and std dev = 1, or scaling all the input values to $[0, 1]$, which could be easily done in this case too if all the inputs are angles in $[0, 2 \pi]$. But, what about distances? Is it common to, for example, use as input some distance of the agent to the ground, or a distance to some target position? In that case, the scale of the distances can be arbitrary and vary a lot. What are some common ways to deal with that?
$\begingroup$
$\endgroup$
2
-
$\begingroup$ and by articulated I mean robots with joints, as opposed to, say, a wheeled robot. $\endgroup$– Mei ZhangCommented Feb 15, 2019 at 8:25
-
$\begingroup$ Welcome to SE:AI. We sometimes have issues with overlap, since the general field of AI is so broad. We do like robotics questions, so long as they involve the decision-making aspects! :) $\endgroup$– DukeZhouCommented Feb 15, 2019 at 22:08
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
This paper might provide some answers https://arxiv.org/pdf/1810.05762.pdf
For the observations / states they used not only angles, but also velocities, heights and positions (Table 2).
In 4.2 Learning algorithm you can see that they mention this, which is related to your question about normalization:
Additionally, for stability we whiten the current observations by maintaining online statistics of mean and standard deviation from the history of past observations.
