Closed ddliu365 closed 4 years ago
We use only the Jacobian of the dynamics. It is equivalent to an iterative LQR algorithm.
Despite that, we prefer to think that our solver it is a DDP algorithm with Gauss-Newton assumption.
Crocoddyl has different solvers. One of our novel solver is called feasibility-driven DDP which it is a remarkable improved version of the DDP algorithm.
Additionally, our DDP algorithm is not exactly the classical implementation. We make a relinearization of the Value function to be able to accept the state trajectory as warm-start.
For more details see the Crocoddyl paper described in Readme file.
I will close this issue.
Feel free to ask further questions in other issues.
It seems this solver only used the first derivatives of the dynamics instead of second derivatives which should be included in DDP.