D2C — Decoupled Data-Based Approach for Learning to Control Nonlinear Dynamical Systems

Published in IEEE Transactions on Automatic Control, 2022.

Overview

Most data-driven control methods require large amounts of data because they entangle the processes of learning the system dynamics and designing the controller. D2C breaks this coupling by separating these two phases:

  1. Data collection phase — A small amount of exploratory data is collected around a nominal trajectory using open-loop perturbations.
  2. Controller synthesis phase — A local linear model is fit to the data, and a feedback controller is synthesized analytically using this model.

This decoupling leads to significant improvements in sample efficiency while maintaining strong performance guarantees on nonlinear systems.

Key Results

IEEE Paper arXiv Preprint