Consistent relabeling strategy for dense action periods

Develop a consistent offline action relabeling strategy for low-level actions during dense periods—i.e., contact-rich manipulation segments—in the HYDRA hybrid action framework, so that action labels in these segments can be standardized without harming distributional alignment between training and deployment.

Background

HYDRA improves action consistency by relabeling actions during sparse (free-space) periods using a waypoint controller, which simplifies behavioral cloning and reduces distribution shift. However, while this relabeling is well-defined for sparse segments, the authors note that an analogous, reliable approach for dense (contact-rich) segments—where precise low-level actions are critical—has not been established.

Establishing a consistent, generally applicable relabeling method for dense periods would further reduce action variability in these critical phases and could enhance policy performance without collecting new expert data.