Well, it's still using A* and state exploration to validate the plans generated. But, it's generating plans nondeterministically so it is possible it could run in O(\inf).
Why didn't they give more details about the Sokoban puzzles being solved? The standard benchmark is the 90 puzzles in XSokoban and there's a Large Test Suite that has a few thousand but many of them haven't been solved by any search or planning system. Even the 90 puzzles in XSokoban were only solved a few years ago (2020?) by Festival (using FESS, a feature-space search in addition to domain-space, state-space). Before that it had been about 20 years of only 57 or so levels solved via search.
I see that they measure trace lengths but I would have really liked to see # of XSokoban levels and how many states were examined, to line up with existing research.
Why didn't they give more details about the Sokoban puzzles being solved? The standard benchmark is the 90 puzzles in XSokoban and there's a Large Test Suite that has a few thousand but many of them haven't been solved by any search or planning system. Even the 90 puzzles in XSokoban were only solved a few years ago (2020?) by Festival (using FESS, a feature-space search in addition to domain-space, state-space). Before that it had been about 20 years of only 57 or so levels solved via search.
I see that they measure trace lengths but I would have really liked to see # of XSokoban levels and how many states were examined, to line up with existing research.