The place of the L1-norm in robust estimation