Accelerating Second-Order Differential Dynamic Programming for Rigid-Body Systems

This letter presents a method to reduce the computational demands of including second-order dynamics sensitivity information into Differential Dynamic Programming (DDP) trajectory optimization algorithm. An approach DDP is developed where all necessary derivatives are computed with same complexity as in iterative Linear Quadratic Regulator (iLQR). Compared linearized models used iLQR, more accurately represents locally, but it not often since tensorial and expensive compute. work shows how avoid need for computing derivative tensor by instead leveraging reverse-mode accumulation compute key vector-tensor product directly. We also show structure can be further accelerate these computations rigid-body systems. Benchmarks this multi-link manipulators that benefits included without sacrificing evaluation time, done fewer iterations than iLQR.