In Illusie's original article, the de Rham-Witt complex is defined for a smooth scheme over a perfect characteristic $p$ base $S$, without reference to $S$. Some 25 years later, Langer and Zink defined the de Rham-Witt complex for a smooth morphism $f:X\to S$ (where $S$ is a $\mathbb{Z}_{(p)}$-scheme, though I am mainly interested in the case over $\mathbb{F}_p$).
Question. Why was it difficult to come up with a relative version? Where is the catch?
I have to admit I haven't studied the existing literature very well yet (in particular, I have only a vague understanding of the work of Langer-Zink). To Illusie the need for a relative theory must have been obvious, but as far as I know he does not give hints why the naive ways of defining it do not work. By the time the second paper (Illusie-Raynaud) came out, there were many ways of constructing the de Rham-Witt complex:
As the universal de Rham $V$-pro complex,
$W_n\Omega^q_X := R^q u_{X/W_n *} \mathcal{O}_{X/W_n}$ where $u: (X/W_n)_{cris} \to X_{Zar}$ is the projection (there is a way to get $R$, $d$, $F$ and $V$).
If $X$ has a lift over $W$ together with a lift of Frobenius, then we can identify the de Rham-Witt complex as a suitable completion of a subcomplex of the direct limit of the de Rham complexes of the lift over the lift of Frobenius.
Why don't they easily generalize? For example #2 seems okay if we replace $/W_n$ by a (formal) lift of $S$ over $\mathbb{Z}_p$ which is possible (locally on $S$) if $S$ is smooth.
Not to mention the obvious idea of taking the quotient of $W\Omega_X$ by the pull-back of $W\Omega_S$...
Log-motivation. The need for a relative version is more apparent in the logarithmic setting: if $X$ is a proper log smooth scheme over the log point $S=Spec(\mathbb{N}\to k)$, there are two interesting de Rham complexes: $\Omega^\bullet_{X/S}$ and $\Omega^\bullet_{X/k}$ (in the latter the base has trivial log structure). We need both to define the Gauss-Manin connection, i.e. the monodromy operator on $H^\bullet(X, \Omega^\bullet_{X/S})$. If $k$ is a perfect characteristic $p>0$ field and we are interested in crystalline cohomology, we have even more choices for the base: (1) $\mathbb{N}\to W(k)$ ($1\mapsto p$), (2) $\mathbb{N}\to W(k)$ ($1\mapsto 0$) and just (3) $W(k)$ with trivial log structure. If we want to realize crystalline cohomology of $X$ over these bases by a version of the de Rham-Witt complex, we need a relative version. I know that Hyodo and Kato defined a log version of the de Rham-Witt complex and even showed an isomorphism between cohomology over (1) and (2) after $\otimes \mathbb{Q}$, but again their de Rham-Witt complex is absolute (I forgot whether it computes (1) or (2)).
