AdaptiveNLP: a Framework for Efficient Online Adaptability in NLP Structures for Optimal Control Problems

Direct methods that transcribe an Optimal Control Problem (OCP) to a Nonlinear Program (NLP) have proven effective to solve OCPs. Flexibility in this transcription that can adapt online to a changing environment by adding or removing constraints or changing the discretization of the dynamics can benefit many applications such as motion planning in dynamic environments. This work presents AdaptiveNLP, a software framework that efficiently constructs NLP functions based on pre-computed derivative information and provides functionalities to modify the NLP problem structure with low overhead. This adaptability enables the user to discard constraints known to be inactive which reduces computation times. In Model Predictive Control (MPC), it also allows tailoring a specific MPC iteration's NLP to the environment at that time instance. An MPC example and an adaptive gridding example show the effective reduction of total computation time and the ability to refine the time-grid of an NLP to produce a sparse but highly accurate solution with little overhead, respectively.