I would probably stick to AES (Rijndael). It's well studied, lots of implementations are available, it's quite simple, and is small and fast on 8-bit CPUs. I don't see any reason not to use it unless you're seriously constrained in one parameter or another. Many micros you may choose to target will have a coprocessor for it too, which you might be able to make use of at one point or another in your system.
TEA/XTEA/XXTEA are simpler and maybe faster/smaller but much less well studied (and have more serious known weaknesses) and for modern CPUs probably don't offer a huge performance gain over Rijndael.
My understanding from the AES process was that RC6 was both larger in RAM and slower than Rijndael and possibly difficult/slow to implement in constant-time (at least on small 8-bitters) without exposing timing attacks due to variable rotation and multiplication. This might be well out of date now, from Schneier's papers during the AES competition.