One way to get one of the many statistically possible paths is to start from the end and work backwards. This works with the maze puzzles on kids' menus, but also with proteins. To find how a ligand enters an active site, dynamic undocking uses tethered MD to pull a bound form out. Likewise for a protein folding starting from a folded state and lowering the attraction term of Lenard-Jones or gradually increasing the reference temperature used by the Berendsen thermostat in an MD run (as done for DNA annealing simulations).
This assumes that the system is reversible (not true for protein with a maturation step like GFP), but it is avoids getting stuck in local minima or dead-end paths in the case of maze puzzles.
It also assumes that peptides start folding only once they are finished being synthesised, which is not true and actually circular permutations are an interesting case as many the permutation windows/spokes are unfolded.
However, the path down a folding funnel is more of a hike/roll down a canyon than a maze puzzles on paper as there are many possible routes to the solution. So having a bunch of trajectories makes a nice video, but does not say much...

A seemingly similar concept, but very different, is the pathway reconstruction for the evolution trajectory that leads to a multi-mutation variant of interest. This is well studied because the intermediates are not transient but have to persevere in the population until another mutation appears. If they have a too high fitness cost they will not be around long enough for a compensation mutation (cf. epistasis).