This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Robust Offline Imitation Learning Through State-level Trajectory Stitching
Created by
Haebom
Author
Shuze Wang, Yunpeng Mei, Hongjie Cao, Yetian Yuan, Gang Wang, Jian Sun, Jie Chen
Outline
This paper presents an offline IL method that leverages low-quality, unlabeled data to address the lack of high-quality expert data and covariate shift in imitation learning (IL). Specifically, we introduce a state-based search framework that connects state-action pairs from incomplete demonstration data, generating diverse and information-rich training paths. Experimental results demonstrate that the proposed method significantly improves both generalization and performance on standard IL benchmarks and real-world robotic tasks.
Takeaways, Limitations
•
Takeaways:
◦
We present a novel method to improve the data efficiency of imitation learning by leveraging low-quality data.
◦
State-based retrieval framework enables the generation of more diverse and information-rich training data.
◦
Proving practicality by verifying performance improvements in actual robotic tasks.
•
Limitations:
◦
The performance improvements of the proposed method may be limited to specific benchmarks and robotic tasks.
◦
State-based search frameworks can be computationally expensive.
◦
Generalization performance evaluations on various types of low-quality data may be lacking.
