I had been avoiding to look into details of Relax NG thinking it to be just another schema proposal until saw some good posts from Tim Ewald and Aaron Skonnard. A good introduction of RelaxNG can be found here at http://relaxng.org but the articles (1, 2, 3) which really throw an insight as far as comparisons are concerned are ones from David Mertz at IBM developerWorks. I found part 1 & part 2 very interesting. David provides the code fragments to assert the differences and way both (RelaxNG & XSD) express the same XML instance.
RelaxNG syntax is definitely way simpler than XSD & provide uniformity in describing elements and attributes. I would definitely agree with its creators when they say you need 30 minutes to understand it.
David mentioned a valid case where XSD scores is that of specifying number of occurrences of elements. RelaxNG is limited to <zeroOrMore>, <OneOrMore> & <optional> where as XSD can use any range for cardinality. In part 2, David mention about infoset augmentation, siting an example of how in DTD and XSD, default values are inserted in XML instance and James Clark thinks it to be a bug in DTD and XSD specifications rather than a feature. RelaxNG does not support infoset augmentation. Well, in my opinion its definitely a feature and do not buy the argument of consequences of schema document being not available because of network. Network argument can be used on all parts of schema documents.
Except for 2 features described above (cardinality & infoset augmentation) RelaxNG is better than XSD in almost all ways for eg. Uniform treatment for Elements & attributes, modular & extensible data type system etc