Learning Robust Skills for Tightly Coordinated Arms in Contact-Rich Tasks

Multi-arm coordination has been receiving increasing attention in industry and our daily life. However, empowering arms to skillfully manipulate a common object within contact-rich tasks remains a daunting challenge. In this letter, we introduce a novel learning-based framework for arms to manipulate the object skillfully while maintaining tight coordination, which not only avoids dynamic modeling for keeping tight coordination but also eliminates the necessity to pre-program task-specific skills for handling complex contacts. To the best of our knowledge, our study is the first to concentrate on tightly coordinated arms in contact-rich tasks. Specifically, we improve and utilize value-decomposed multi-agent policy gradient methods to simultaneously learn tight coordination skills and task-specific skills. Contrasting with previous work on multi-arm skill learning, our approach efficiently learns skills without integrating human skills. Furthermore, we evaluate our approach through two challenging dual-arm assembly tasks and the performance exhibits promising robustness and generalizability.