TITLE: Wander distance of self-avoiding walk that backs out of culs-de-sac
QUESTION [11 upvotes]: Suppose a self-avoiding walk on $\mathbb{Z}^2$, with random steps each one unit long,
backs out of culs-de-sac, but retaining the lattice points on which it stepped
marked as unavailable for future steps.
For example, in (a) below, a cul-de-sac is reached. Unwalking two steps allows
the path to escape and walk in another direction in (b); but note those two
points are marked. Eight steps further another cul-de-sac is reached (c).
And so on.

 


The walk can continue forever, because once it unwinds to a point
on the bounding box, it is free to step outside that box.
My question is: 

What is the growth rate of the distance of the path endpoint 
  from the origin, with respect to the total length of 
  (number of steps in) the path?

Perhaps this model has been studied?
Here's a longer example:

 
 
 
 
 


(Added.) I can't resist one more, extending the above to 11,371 steps:

 
 
 
 
 


(Added 29Agu14.) Permit me to point to the new paper j.c. found and includes in a comment to his (knowledgeable) answer. That paper calls these walks SKSAWs: Smart Kinetic Self-Avoiding Walks.

REPLY [6 votes]: This process is quite natural and has in fact been studied under the name of the infinitely growing self-avoiding walk in a paper by Kremer and Lyklema (see also this paper by Birshtein, Buldyrev and Elyashevitch who called it the "infinitely prolonging walk").  
Later Weinrib and Trugman wrote a paper introducing smart kinetic walks (which are IGSAWs which can close up on themselves) and connected the whole family to walks around critical percolation clusters - thus your process should be very much related to the critical percolation exploration process.  See also this paper by Coniglio, Jan, Majid and Stanley.  I think similar arguments are described in sections 0.4 and onwards of Lawler's book "Conformally Invariant Processes in the Plane".
Coniglio et al quote results from percolation which imply that if $N$ is the number of steps taken which don't go into the cul-de-sacs, then the root mean squared end to end distance scales like $N^{4/7}$.
Here's a picture of the percolation explorer on the hexagonal lattice, from this page:

The critical percolation explorer is conjectured to have SLE${}_6$ as its scaling limit, and so that in some sense should be the scaling limit of your curves as well.  This was famously proved by Smirnov for the hexagonal lattice and is still open for the square lattice.