To be more precise, F# is based on the non-OO parts of Caml, with a completely different object system so it would be be compatible with .Net objects from other languages such as C#.

IIRC, the original research was going to be into making Haskell for .Net, but there were enough issues (in particular, being purely functional makes interop with C# and the existing .Net libraries radically more difficult) that ML seemed like a better fit.