Qiyuan Qiao

Reverse to Advance: Teleoperation-Cost Effective Hard Policy Learning from Reversed Easy Tasks

Qiyuan Qiao, Ge Yuan, Can Wang, Dong Xu

High-quality teleoperation datasets are costly to collect, particularly for hard tasks. We observe that many tasks exhibit directional asymmetry: completing the forward hard task is difficult, whereas reversing it by relaxing or disrupting the environment is comparatively easy. T…Read more