A natural way in which Finsler metrics appear in Riemannian geometry is as averages of Riemannian metrics (e.g., the average of the arc-length elements
$\sqrt{dx^2 + dy^2}$ and $\sqrt{2dx^2 + 3dy^2}$ is not the arc-length element of
a Riemannian metric on the plane). A simple application of this idea is the construction of a class of (Finsler) metrics in projective space for which geodesics are projective lines: 

Consider the group $G$ of projective transformations on projective $n$-space and let $\mu$ be a compactly-supported Borel probability measure on $G$. 
Define the length of a curve $\gamma$ to be the expected value of the (canonical) length of the curve $T(\gamma)$, where $T \in G$. 

**Claim.** The preceding construction defines a length structure on 
$\mathbb{R}P^n$ for which projective lines are geodesics. 

This is essentially the same construction as in [this MO question][1].

We can consider the following abstract version of this construction:

Consider a length space $(X,d)$ and a continuous group action $G \times X \rightarrow X$ of a locally compact topological group $G$ on the space $X$. If $\mu$ is a compactly-supported Borel probability measure on $G$, 
define the length of a curve $\gamma$ to be the expected value of the (canonical) $\mu$-length of the curve $T(\gamma)$, where $T \in G$, and define 
the $\mu$-distance between two points $x, y \in X$ as the expected value of 
$T \mapsto d(Tx,Ty)$. 

**Questions.**

**1.** *Is the length of a curve in $(X,d_\mu)$ equal to its $\mu$-length?*

**2.** *Is $(X,d_\mu)$ a length space?*

**3.** *If $(X,d)$ is a Riemannian torus with distance function $d$, $G$ is also the torus acting on itself by translations, and $\mu$ is  the normalized Haar measure on $G$, is it true that $d_\mu$ the flat metric induced by the stable norm?* 

Let me sketch why I think this last question may have a positive answer: a characterization of the stable norm by D. Burago states that if $d$ is a periodic metric (the lift to $\mathbb{R}^n$ of a length metric 
on $\mathbb{R}^n/\mathbb{Z}^n$), the stable norm is the unique norm satisfying
$$
\|x-y\| - C \leq d(x,y) \leq \|x-y\| + C
$$
for some constant $C > 0$ independent of $x$ and $y$. I hadn't noticed this before, but this means that *the set of periodic metrics sharing the same stable norm is a convex set* and, therefore, an average of periodic metrics with the same stable norm $\|\cdot\|$ (if it is a length metric) will yield a periodic metric with stable norm $\|\cdot\|$. Averaging the metrics $d(x+z,y+z)$ as $z$ ranges over all points in the unit square should then yield a translation-invariant metric whose stable norm is $\|\cdot\|$. With some luck, this translation-invariant metric is induced by a norm and by Burago's theorem it would have to be the stable norm.

**Disclaimer.** I really haven't thought all these things through, but I thought it would be helpful to write it down to make things clearer and perhaps someone here has already considered this averages. They are quite natural.


  [1]: http://mathoverflow.net/questions/119668/