> The binding is on a layer that humans don't see, and it makes it fragile.

And that seems to be part of the point of this project. From the description:

> In Selenium, you need to use HTML IDs, XPaths and CSS selectors to identify web page elements. Helium on the other hand lets you refer to elements by their user-visible labels.

The example script uses a bunch of user-visible labels, not classes, ids, ancestors, etc. It is things like "click('Sign in')" which you would tell a user you were walking through an interface.