Event

EquiContact: Hierarchical Equivariant Framework Improves Spatial Generalization in Contact-Rich Manipulation

1 reports · 1 sources

Joohwan Seo and co-authors released a preprint proposing EquiContact, a hierarchical SE(3) vision-to-force equivariant framework for learning contact-rich manipulation policies that can generalize across task configuration spaces. The framework consists of a high-level visual planner, Diff-EDF, and a low-level compliant visuomotor policy, G-CompACT. G-CompACT uses local observations from wrist RGB images, end-effector pose feedback given as geometrically consistent error vectors (GCEV), and force/torque readings, and outputs actions defined in the end-effector frame. The authors claim that after the robot learns a contact-rich task such as peg insertion, it can perform similar tasks at different positions and orientations without retraining the low-level policy. The authors report that real-world peg insertion experiments achieved near-perfect success rates under spatially out-of-distribution configurations; there are also screw-driving and surface-wiping experiments used to evaluate the transfer of this localized policy design beyond insertion tasks.