First, you seem to be misunderstanding what orbital stability means.
The notion of orbital stability should be considered as in contrast against asymptotic stability. The latter notion is the stability that one usually expects around, say, hyperbolic fixed points of a finite dimensional dynamical system, where a stable fixed point is one where perturbations eventually decay away and the state of the system returns to the fixed point. In many physical settings, however, one cannot expect asymptotic stability, because the fixed point may be such that there are small perturbations in orbit around it. The simplest example is considering the two dimensional dynamical system generated by the vector field $x\partial_y - y\partial_x$. The origin is a fixed point, but small perturbations of the origin do not return to the initial data.
One can then generalize this to non-equilibrium points (such as the traveling KdV solitons), by asking whether initial data which are small perturbations of a particular solution (such as the solitons), will remain for all time small perturbations of the reference solution. Notice that for linear systems orbital stability is trivial for any solution. For nonlinear systems generically there can be positive feedback which would cause perturbations to diverge from the reference solution at large times; and hence orbital stability is not trivial for general nonlinear systems.
For the KdV system, around a soliton, that orbital stability is the right notion (as opposed to asymptotic stability) can be justified via the symmetry of the equation. (The symmetry of the equation itself is not used to "derive orbital stability".) The reason is that given a soliton solution $u(t,x) = \phi_c(x - ct)$, if you consider the initial data given by $v(0,x) = \phi_c(x + \epsilon)$, the corresponding solution will be $v(t,x) = \phi_c(x + \epsilon - ct)$. For small $\epsilon$ you have that $v(0,x) = u(0,x) + \epsilon \partial_x u(0,x) + O(\epsilon^2)$ and hence $v(0,x)$ is a small perturbation of $u(0,x)$. Asymptotic stability must fail for the KdV system because $\| v(t,x) - u(t,x)\|$ is constant in time for any reasonable norm, and so the perturbed solution never returns to the original soliton. And therefore the best one can hope for is orbital stability.
Second, one can also alternatively try to quotient out the influence of the symmetries if one wants better control of the solution at large times. For example, given the discussion above, one may ask whether the translation symmetry is the only mechanism preventing asymptotic stability from holding. That is, is it in fact possible that if we factor in this effect of the symmetry, the perturbed solutions can always be regarded as approaching some other soliton? One way to formulate this is, for example, to say: for a solution $v(t,x)$ evolving from an initial data that is a perturbation of my reference soliton solution $u(t,x)$, I am not going to ask whether $v(t,x)$ eventually approaches $u(t,x)$. Instead, I am going to look at the set of all solutions of the form $u(t,x + \xi)$ for some $\xi\in \mathbb{R}$, and ask whether $v$ eventually approaches one of these solutions. This is the setting where one now considers instead of a singular solution, a manifold of solutions (the soliton manifold). One can then ask whether a perturbed solutions $v(t,x)$ can be optimally decomposed as a (time-dependent) element of the soliton manifold plus a small (perhaps decaying) perturbation. And this gets us to modulation theory.
To conclude, however, the point of view of the soliton manifold has sometimes some additional (coincidental) advantages. Returning to the original KdV equation, one notes that $v(x,t) = \phi_{c+\epsilon}(x - (c+\epsilon) t)$ is also a solution. For sufficiently small $\epsilon$ this $v$ is again a small perturbation (in terms of initial data) from that of $u(x,t) = \phi_c(x - ct)$. However, evidently at large times, the differing speeds will make $v$ and $u$ very different solutions. If we consider the point of view of the soliton manifold relative to space translations this happen to be neatly resolved: while $v$ is not going to be close to $u$ at late times, one can easily check that $v(x,t)$ will be in fact close to $u(x - \epsilon t, t)$, a time-dependent translation of the original soliton solution. This means that if we relax ourselves to study orbital stability about the soliton manifold, then we can also account for initial data perturbations that changes the soliton velocity. (This is not to say that we shouldn't consider the manifold of solutions corresponding to all translations of all different velocity solitons! Maybe this will give us better information!)
To give a more pedestrian example: time translation is another symmetry of the KdV equation. However a time-translated soliton can be also identified with a spatially translated soliton. So trivially the soliton manifold (defined with respect to spatial translations) also captures non-decay due to time-translation.
In this business the general idea is that if you have some obstruction to decay, you can start by trying to see whether you can embed your background solution to some parametrized family of solutions, and see if your obstruction can be characterized as arising from a change in the parameter. One easy way to obtain a parametrized family of solutions is to exploit the symmetries inherent in the equations.
