When this turned up on e2e my belief was (and still is):
whether or not "UDP Lite" makes sense depends critically
on wireless link error patterns.  Lots of things make
sense depending on error models, but many papers proposing
wireless-related things use simplistic and/or counter-factual
error models (sometimes because they are subject to easy
closed-form mathematical analysis).

As I see it, for "UDP Lite" to pay off, you need to
*frequently* get errors, and they need to *frequently*
come late in the packet.  They need to not nail the IP
header, they need to not nail the "UDP Lite" "checksum
this much of the packet" field, and they also need to
not nail critical application-level data (if you want to
treat errors in I and P and B frames differently, how
can you do that if the frame-type bits get nailed?).

You also need to get the link layer to ignore checksum
failures, which is trivial in theory but hard in practice.

The question (unanswered on e2e if I recall) is "exactly
which links are known to have such an error pattern, and
why can't the problem be corrected by existing link-level
error coding techniques?".  Do the papers you cite provide
pointers to detailed symbol-level error syndrome traces?

Dave Eckhardt