Take a look at this Hal Varian paper:
Many papers in applied econometrics present regression results in a
table with several different specifications: which variables are
included in the controls, which variables are used as instruments, and
so on. The goal is usually to show that the estimate of some
interesting parameter is not very sensitive to the exact specification
used. One way to think about it is that these tables illustrate a
simple form of model uncertainty: how an estimated parameter varies as
different models are used. In these papers the authors tend to examine
only a few representative specifications, but there is no reason why
they couldn’t examine many more if the data were available.
I would also add that the effect may change when you alter the covariates or the sample, but it should do so in a predictable and theoretically consistent manner to be called robust.
There are other sense of robust that are often used and are somewhat related: robust to heteroskedasticity or autocorrelation, outliers, and various assumption violations (like error distributions).