Data Attribution for Model/Learning Based Control (Extended Abstract)

Jiachen Li, Shihao Li, Jiamin Xu, Soovadeep Bakshi, Dongmei Chen

Keywords: Data attribution, Influence functions, Control

Abstract: Ensuring safety of learning-based robotic controllers requires understanding which training data influences performance, yet identifying influential trajectories through exhaustive retraining is computationally prohibitive. We introduce a framework using influence functions to efficiently approximate the impact of individual training trajectories on learned dynamics and control performance. We formulate IF1 to estimate effects on model accuracy and IF2 to quantify impacts on LQR control cost—a proxy for tracking error and stability. Empirical validation demonstrates strong correlations between influence predictions and ground truth.

Submission Number: 18