Abstract
Long-horizon contact-rich bimanual manipulation presents a significantchallenge, requiring complex coordination involving a mixture of parallelexecution and sequential collaboration between arms. In this paper, weintroduce a hierarchical framework that frames this challenge as an integratedskill planning & scheduling problem, going beyond purely sequentialdecision-making to support simultaneous skill invocation. Our approach is builtupon a library of single-arm and bimanual primitive skills, each trained usingReinforcement Learning (RL) in GPU-accelerated simulation. We then train aTransformer-based planner on a dataset of skill compositions to act as ahigh-level scheduler, simultaneously predicting the discrete schedule of skillsas well as their continuous parameters. We demonstrate that our method achieveshigher success rates on complex, contact-rich tasks than end-to-end RLapproaches and produces more efficient, coordinated behaviors than traditionalsequential-only planners.