Inner Product Regularization for Imitation Learning

Abstract

Inner product regularization (InneR) encourages imitation policies to respond consistently to the same perturbation at nearby states. It groups states with aligned policy-induced inner products and regularizes their average so that the policies share directions of insensitivity. The method supports deterministic and stochastic policies, improving robustness to irrelevant state dimensions and distribution shifts. Experiments include simulated systems and diffusion policies for real-world robot manipulation.

Publication
Under review at the International Journal of Robotics Research (IJRR)
Tianyu Qiu
Tianyu Qiu
Ph.D. Student @ UT Austin