Adversarial Augmentation for Task-Parameterized Underwater Skill Learning via Digital Twins*

Abstract

Learning from Demonstration (LfD) provides an efficient approach to acquiring diverse underwater skills, with task-parameterized learning enhancing the generalization of policies. However, collecting comprehensive underwater demonstrations across various conditions remains a significant challenge. In this work, we propose an adversarial trajectory augmentation method for Task Parameterized Hidden Semi-Markov Models (TP-HSMM) based on digital twins, inspired by adversarial example generation. Our method aims to improve the performance of motion policies by utilizing adversarial trajectory generation and retraining, leveraging low-cost demonstrations from digital twins. We evaluate the proposed adversarial trajectory augmentation method on two datasets. Comparative experiments demonstrate its effectiveness in reducing trajectory generation errors in new scenarios. Finally, we validate the method through an underwater humanoid plugging experiment, showing that it achieves similar performance to the baseline with fewer demonstrations.