FoldBack:长视野(horizon)衣物折叠的自校正掩码生成式策略
FoldBack: Self-Correcting Masked Generative Policy for Long-Horizon Garment Folding
Lipeng Zhuang · Shiyu Fan · Yingdong Ru · Zhuo He · Florent P. Audonnet · Paul Henderson · et al.
中文摘要
FoldBack 是一种面向长视野衣物折叠任务的自校正掩码生成式策略(masked generative policy)。现有长轨迹策略在抓取失误或滑动后,即使衣物未到达目标构型仍可能继续执行。FoldBack 的恢复机制围绕推理时的三项决策组织:何时进行细化并验证、如何回滚、以及在何处以何种方式重试。它将细化与抓取验证对齐到 pick-and-place 事件,把机器人返回到可重试的预抓取构型同时保留已成功的抓取,并选择性地重新生成失败片段及部分后续动作,规避此前的失败抓取位置。据作者所知,FoldBack 是首个统一上述决策的可编辑全轨迹策略,能够在执行继续前检测、撤销并修复失败交互,且无需恢复演示或基策略重训练。在来自 6 个类别的 33 件真实衣物上,FoldBack 取得 75.2% 的最终折叠成功率与 0.837 的最终掩码 IoU,最优先前基线分别为 45.7% 和 0.689。
关键要点
- 01问题:现有长轨迹策略在抓取失误或滑动后仍继续执行,无法处理衣物未到位的情况
- 02方法:围绕"何时细化验证—如何回滚—何处重试"三个推理决策,结合 pick-and-place 事件做抓取验证与可编辑回滚
- 03结果:在 6 类共 33 件真实衣物上,折叠成功率 75.2%,最终掩码 IoU 0.837,显著优于最强基线 45.7%/0.689
- 04局限:无需恢复演示或基策略重训练,但评估限于 33 件衣物的小规模真实实验,未见更大规模或多样化物体上的验证
解读
尚无解读。
原始英文摘要
arXiv:2610.10462v1 Announce Type: new Abstract: We present FoldBack, a self-correcting masked generative policy for long-horizon garment folding. Existing long-trajectory policies may continue after a missed or slipped grasp even when the garment has not reached the intended configuration. We structure FoldBack's recovery mechanisms around three inference-time decisions: when to refine and verify, how to roll back, and where and how to retry. 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.