Fixed.

This should be fixed as part of the issue132 branch. (The issue is not related
to issue132, but the fix was included in that branch by accident.)

Setting this to "testing". Gabi, if you can verify that this now works properly,
feel free to set it to resolved.

Universally quantified effects are a very rarely used feature in our benchmark
set, and nontrivial type hierarchies are not that all that common either.

I think this one place (the only place, I think?) where our grounding
algorithm's implementation differs a bit from how it is described in the paper.
In the long run, it may be a good idea to change it to the way it's described in
the paper, which is conceptually cleaner and should be no less efficient. Can
you add a TODO to the code?

Yes, the universally quantified effects are the problem:
The instantiate-methods get a type-to-objects dictionary which is
used to identify the possible instantiations of the quantified variables.

This dictionary did not contain entries for supertypes, yet.
This is actually very easy to fix (like 3 lines), but I am surprised
that nobody encountered this bug before.

BTW: Thank you Patrik for your really helpful bug reports.
With such small examples it is almost a pleasure to fix the bugs.

Wonderful, thanks! I'm surprised about the supertypes issue. Does this have
something to do with universally quantified effects? I saw some unexpected
weirdnesses in the instantiation code there yesterday.

Case #1 is the result of two bugs:

First, the invariant synthesis is not correct (already solved in issue132 but
not yet in the trunk).

Second, supertypes of types were not considered when instantiating the actions.
This lead to action (FaultSensor-sense_fault-__SPONTANEOUS__-fault FS1) being
removed.

Solving these bugs also resolves case #2.

Malte, as soon as the solution of issue132 is in the trunk, I
can branch from that version and fix the second bug.

Update wrt case #2: VAL (version 4.2.04) fails to validate the plan with cost 2
(soln.2). I'm pretty sure this is a VAL bug though.

There are potentially two separate issues here.

Case #1: tg-test1 (-domain/-problem) is solvable (solutions found by at least
three other planners) but the fd translator reports it as unsolvable (goal
simplified to false).

Case #2: test1 (-domain/-problem). This problem has a solution with cost 2 (I
believe that to be optimal), but fd search with options "bo" (i.e., blind
optimal search) finds a solution with cost 10.