FoldBack: Self-Correcting Masked Generative Policy for Long-Horizon Garment Folding
arXiv cs.AI · 8 Oct 2026, 03:28 AEST
FoldBack is a self-correcting masked generative policy designed for long-horizon garment folding. It improves upon existing policies by enabling the detection and repair of failed interactions without the need for recovery demonstrations or retraining.
Key facts
- We present FoldBack, a self-correcting masked generative policy for long-horizon garment folding.
- FoldBack aligns refinement and grasp verification with pick-and-place events, returns the robot to a retryable pre-grasp configuration while preserving successful grasps, and selectively regenerates the failed segment and selected future actions while avoiding previous failed grasp locations.
- To our knowledge, FoldBack is the first editable full-trajectory policy to unify these decisions, enabling failed interactions to be detected, undone, and repaired before execution continues, without recovery demonstrations or base-policy retraining.
- Across 33 real garments from six categories, FoldBack achieves 75.2% final folding success and 0.837 final-mask IoU, versus 45.7% and 0.689 for the strongest prior baseline.
Read aloud
FoldBack is a self-correcting masked generative policy designed for long-horizon garment folding. It improves upon existing policies by enabling the detection and repair of failed interactions without the need for recovery demonstrations or retraining. We present FoldBack, a self-correcting masked generative policy for long-horizon garment folding. FoldBack aligns refinement and grasp verification with pick-and-place events, returns the robot to a retryable pre-grasp configuration while preserving successful grasps, and selectively regenerates the failed segment and selected future actions while avoiding previous failed grasp locations. To our knowledge, FoldBack is the first editable full-trajectory policy to unify these decisions, enabling failed interactions to be detected, undone, and repaired before execution continues, without recovery demonstrations or base-policy retraining. Across thirty three real garments from six categories, FoldBack achieves seventy five point two percent final folding success and zero point eight three seven final-mask IoU, versus forty five point seven percent and zero point six eight nine for the strongest prior baseline.
Sources in this cluster
- FoldBack: Self-Correcting Masked Generative Policy for Long-Horizon Garment Folding
arXiv cs.AI · 8 Oct 2026, 03:28 AEST