A couple of years ago (see how long it takes me to blog this?) I translated two programs from languages I didn't know to one I did. The destination language in both cases was one of my pet Lisps. I was translating code into it to test its expressiveness, which was painfully instructive: there a number of places where my language couldn't do what I wanted, and some of them were hard to fix. But most of what I learned from the exercise had nothing to do with my language.
One of the programs was in obfuscated Perl. It (ab)used enough obscure parts of the language that I didn't translate it properly. That's fine. I learned a lot of Perl in the process.
The other program was in Groovy, a sort of functional Java. It's a nice example because it involves networking, XML, and a graphical user interface, so it tests a language's libraries in ways most toy examples don't. It was probably chosen to make Groovy look good, and it does - but my Lisp version came out about half the size of the Groovy one. Unfortunately for my bragging rights, this wasn't because my language has greater expressive power. It's because Groovy uses Java's libraries, and their interfaces are limited to things that can be used from Java (archaic Java, in many cases). Groovy is a much more expressive language than Java, but the libraries take no advantage of that expressiveness, so half of the program was spent working around limitations Groovy doesn't have. If Groovy had its own libraries, its program would be about as small as the Lisp one.
Well, maybe not quite as small. The Lisp translation had the unfair advantage that I invented many of the libraries as I went, so they were particularly convenient to that program. And since I didn't implement them, I may have missed some ill-definednesses. I don't think they were unrealistic, but in reality the program would probably be a bit longer than the one I came up with. Whenever you can reimagine your language to fit the task at hand, translation will flatter it. This can be useful - there is nothing as encouraging to a proglanger as seeing your language beat Perl at text processing, Unix at file management, and Haskell at data structure crunching. I've done all of these, and let me tell you, it feels good! But it doesn't mean much.
I didn't learn as much Groovy as I did Perl, probably because the Groovy program did not depend on many obscure details of the language. I treated it as a functional language with Java syntax, and that was all I needed to know to understand it. The Perl program was harder, of course; I had to learn a good deal of syntax to see through the line noise to the underlying structure, which was not as exotic as it looked. Like APL, Perl gets much of its apparent strangeness comes from its abbreviated syntax; the language behind the syntax is only a little weird. But in this case there were more layers of language-dependent obfuscation...
This isn't a bad way to learn a language, as long as you enjoy being puzzled. Just make sure the program you're translating is a bad one. Code that makes its meaning clear won't help you, because you'll understand it without having to think about how it works. For once, that's a bad thing.