Thanks, I can see the problem more clearly now!
First of all, this skirt setup is fantastic. I can’t see how it can get any more stable and well defined than this with Ragdoll, other than adding more shapes and more constraints. That said, in this particular setup, there doesn’t appear to be enough level of detail to cope with those intersections.
Floor
So for the floor, I can see… 5 options.
- Hide legs
- Keyframes to compensate
- Increase the resolution (amount) of your Markers
- Post-solve with e.g. nCloth
- Solve entirely with nCloth
For (1), the floor shot could probably get away with just keyframing visibility on those legs at the time of the first intersection. The leg is still felt through the contact with the skirt.
For (2), depending on how often this happens and assuming this shot is an outlier, then some manual intervention might be feasible.
Options 3-5 are more work up-front, and again would depend on how often these things happen. If you’ve got 100 shots and 80 of them needs this kind of manual intervention, then perhaps this is a task better suited for an actual cloth solver.
Grab
Now, for the grab, I would not go any further in simulation, this is already excellent! What’s missing here is a post-deform step. Talking from a CFX perspective rather than a Ragdoll perspective, so for anyone else looking to learn about Ragdoll, look away now. 
Here’s what’s happening.
- On the frame where the grab happens, f127, the skirt mesh is duplicated
- The duplicate is sculpted to fit the hand
- The sculpt is wrap deformed onto the hand, such that it follows
- The wrap deformed sculpt is then blendshaped back onto the skinned skirt
- The blendshape is painted, so as to only include the part where the grab happens
- The blendshape is animated on when the grab happens
- Profit
The only thing I’d consider here is to add more resolution to the skirt to make a better sculpt. That could happen on-top of the already skinned mesh without issue; so long as your pipeline is OK with a different topology (a subdivided version) being sent for rendering.
Even with nCloth or Vellum, I would pat myself on the shoulder for the simulation here already; trying to simulate the grab itself would be wasteful and painful. And for Ragdoll, this is about as high-res as I imagine one can get.
EDIT: Scene file if needed.
Grab_withBlendshape.7z (5.5 MB)
