This dissertation presents a self-supervised continual learning framework designed to address out-of-distribution problem encountered by imitation learning-based robotic control policies. Previous studies have attempted to enhance robustness through d...
This dissertation presents a self-supervised continual learning framework designed to address out-of-distribution problem encountered by imitation learning-based robotic control policies. Previous studies have attempted to enhance robustness through dataset augmentation using synthetic data, or to improve adaptability via continual learning that assumes prior access to labeled data from new environments. However, synthetic data fail to fully capture the complex dynamics of the real world, and acquiring ground-truth labels for unseen environments in advance is practically infeasible. To overcome these limitations, we propose a self-supervised continual learning approach tailored for robotic environments under out-of-distribution conditions. In Chapter 3, to prevent catastrophic forgetting during online continual learning, we propose a method that extracts a high-diversity coreset from an offline dataset while preserving essential offline knowledge, serving as memory for continual adaptation. In Chapter 4, we demonstrate that a real-world robot can resolve out-of-distribution issues and adapt to online environments through self-supervised continual learning applied to a robot traversability estimation model. The most critical factor in online continual learning is accurately estimating online self-supervision. In other words, the offline and online stages should not be treated as separate phases but rather viewed as a single unified learning process. From this perspective, in Chapter 5, we introduce an integrated offline–online learning framework: the offline phase enhances robustness using real supplementary demonstrations, whereas the online phase achieves adaptability by autonomously generating self-supervision signals in online environments. Together, the proposed methods in this dissertation improve both the robustness and adaptability of learning-based robotic control policies, contributing to the development of more realistic and autonomous robot learning systems.