UR-VC is an offline, training-free method for correcting normalized-time progress labels in robot demonstrations. It retrieves similar states across different episodes and aggregates their time-derived labels, allowing the estimate to reflect local regressions caused by slips, failed grasps, or partial undoing. The authors evaluate it on real bimanual cloth flatten-and-fold data, a long-horizon deformable-object task, and use the corrected signal to create advantage labels for advantage-conditioned VLA training. The abstract reports a positive trend in real-robot success under matched settings, but provides no numerical results.
No heat snapshots are available in the last 24 hours.