Storing and loading feedback policies

[wxmerkt]

• A FeedbackMotionSolver only returns the control trajectory as a solution, while the gains remain transient in that instance of the solver.
• This makes loading and replaying control policies impossible.
• The gains are currently stored by the individual solvers, and not in the central FeedbackMotionSolver. DDP/iLQG only have K/k and L/l respectively. iLQR has extra gains.

We need a mechanism to store and load a full policy with solution and gains. Perhaps that requires a data struct (u, K, k, x_ref). Let's have a discussion what's best to proceed. At the bare minimum, we need to be able to access K, k, and x_ref.

[VladimirIvan]

Agreed. Variable type return values are long overdue.

I'm hoping for a lightweight flexible solution that the solvers can populate with different data as required. Maybe based on maps of boost::any and python dictionaries in the wrapper. Possibly with a default conversion to an Eigen::MatrixXd/numpy.array of the state trajectory.

[wxmerkt]

I'd be happy to go with a struct that can be retrieved manually from the solver, i.e., that isn't the solution returned. Otherwise your suggestion sounds great, but is part of a larger overhaul.

[wxmerkt]

We've manually exposed the required quantities from the individual solvers. While a better solution with variable type returns would be nice, it's not a priority.