Learning Control by Iterative Inversion

TL;DR

This paper introduces iterative inversion, a novel algorithm for learning inverse functions without input-output pairs, applied to control tasks using only demonstrations and supervised learning, achieving scalable and effective imitation.

Contribution

The paper presents iterative inversion for inverse learning without pairs, and applies it to control, demonstrating scalable imitation of diverse behaviors without rewards.

Findings

Successfully learned control policies using only demonstrations.

Achieved non-trivial continuous control on multiple tasks.

Outperformed reward-based methods in behavior imitation.

Abstract

We propose $\textit{iterative inversion}$ -- an algorithm for learning an inverse function without input-output pairs, but only with samples from the desired output distribution and access to the forward function. The key challenge is a $\textit{distribution shift}$ between the desired outputs and the outputs of an initial random guess, and we prove that iterative inversion can steer the learning correctly, under rather strict conditions on the function. We apply iterative inversion to learn control. Our input is a set of demonstrations of desired behavior, given as video embeddings of trajectories (without actions), and our method iteratively learns to imitate trajectories generated by the current policy, perturbed by random exploration noise. Our approach does not require rewards, and only employs supervised learning, which can be easily scaled to use state-of-the-art trajectory embedding techniques and policy representations. Indeed, with a VQ-VAE embedding, and a transformer-based policy, we demonstrate non-trivial continuous control on several tasks. Further, we report an improved performance on imitating diverse behaviors compared to reward based methods.