`Feedforward2Dof` cannot host a dynamic feedforward because `Evaluate()` is const

[gabrielsantosphilips]

Summary

Feedforward2Dof is the natural home for a two-degree-of-freedom structure, but its feedforward interface is memoryless by construction:

template<typename T>
class Feedforward
{
public:
    virtual ~Feedforward() = default;
    virtual T Evaluate(T reference) const = 0;   // const
};

Evaluate being const, with no Reset() on the interface, rules out any feedforward that carries state. The canonical 2-DOF design — a reference pre-filter F(z) shaping tracking independently of the feedback that sets disturbance stiffness — needs exactly that.

Feedforward2Dof::Reset() reinforces the assumption:

template<typename T>
void Feedforward2Dof<T>::Reset()
{
    feedback.Reset();
}

The feedforward is never reset, because it is presumed to have nothing to reset.

Why it matters

The two structures are not interchangeable:

• What the class implements: u = clamp(F(r) + C(r - y)), with F a static map. Good for inverse-plant or gravity/friction feedforward.
• What a 2-DOF controller usually means: u = C(F(z) r - y), with F(z) dynamic. This is what decouples command tracking from load stiffness, and it is the form described in the standard references (Astrom & Hagglund, Advanced PID Control).

A caller wanting the second form has three options today, all bad: mutable state behind a const method, a const_cast, or bypassing Feedforward2Dof altogether.

Suggested direction

1. Drop const from Evaluate.
2. Add virtual void Reset() to Feedforward (defaulted to a no-op so existing static implementations are unaffected), and call it from Feedforward2Dof::Reset().

Both are source-compatible for memoryless feedforwards, which appear to be the only ones in the tree today.

If the intent is that Feedforward2Dof only ever models the static-feedforward structure, then the naming and documentation should say so, since "2-DOF" reads as the pre-filter form to most control engineers.

Context

Hit while implementing a two-degree-of-freedom speed controller. Ended up composing filters::passive::ExponentialMovingAverage as F(z) with an incremental PI directly, rather than through Feedforward2Dof, because the pre-filter needs to advance its state on every Compute.