Smart Predict-Then-Control: Control-Aware Surrogate Refinement for System Identification

TL;DR

This paper presents SPC, a control-aware model refinement method that improves system identification by optimizing surrogate objectives, leading to significant performance gains in control tasks like quadrotor trajectory tracking.

Contribution

The paper introduces SPC, a novel control-aware surrogate refinement technique for model-based control, with theoretical guarantees and practical improvements demonstrated on quadrotor tasks.

Findings

SPC reduces quadrotor tracking RMSE by 70%.

SPC decreases closed loop cost by 42%.

The surrogate function is proven smooth with convergence guarantees.

Abstract

This paper introduces Smart Predict Then Control (SPC), a control aware refinement procedure for model based control. SPC refines a prediction oriented model by optimizing a surrogate objective that evaluates candidate models through the control actions they induce. For a fixed surrogate variant under unconstrained control, we establish the smoothness of the surrogate, projected gradient convergence at a sublinear rate of order one over K, and a bias decomposition that yields a conditional transfer diagnostic. On a wind disturbed quadrotor trajectory tracking task, Updated SPC reduces tracking RMSE by 70 percent and closed loop cost by 42 percent relative to the nominal baseline.