Imitation learning (IL) is an important class of modern robot learning algorithm. It aims to enable agents to learn tasks by imitating expert demonstrations. While promising, IL. still faces several challenges that limit its application. This thesis addresses three key problems in imitation learning: leveraging cross-domain expert data, achieving high performance and sample efficiency simultaneously, and ensuring generalization capability.

The first problem tackled is about learning data source: how to effectively leverage cross-domain expert dataset. Traditionally, IL relies on expert demonstrations within the same domain as the target task. However, in real-world scenarios, expert demonstrations may come from different domains. This thesis proposes a novel technique based on invariant...[ Read more ]

Imitation learning (IL) is an important class of modern robot learning algorithm. It aims to enable agents to learn tasks by imitating expert demonstrations. While promising, IL. still faces several challenges that limit its application. This thesis addresses three key problems in imitation learning: leveraging cross-domain expert data, achieving high performance and sample efficiency simultaneously, and ensuring generalization capability.

The first problem tackled is about learning data source: how to effectively leverage cross-domain expert dataset. Traditionally, IL relies on expert demonstrations within the same domain as the target task. However, in real-world scenarios, expert demonstrations may come from different domains. This thesis proposes a novel technique based on invariant representation to transfer knowledge from cross-domain expert data.

The second problem addressed is the performance v.s. sample efficiency dilemma during the learning process. Conventional IL methods often either struggle to match the performance of the expert or require large amounts of online interactions for effective learning. To overcome these limitations, this thesis presents a novel algorithmic framework to unify existing L algorithms, allowing agents to learn from a smaller number of online interactions while achieving competitive results.

The third problem focused on is learning outcome: the challenge of generalization in IL. Agents trained through IL tend to be overly specialized and struggle to generalize to unseen scenarios. This thesis focuses on imitation learning in object-centric scenarios, and introduce techniques that enhance the spatial generalization capabilities of imitation learning algorithms.

Throughout this thesis, extensive experiments and evaluations are conducted on various benchmarks to demonstrate the effectiveness of the proposed methods. By addressing these critical problems, this thesis contributes to advancing the field of imitation learning and provides valuable insights for developing more versatile and efficient learning algorithms.