What's this about?
CARE learns likely post-failure states from failed rollouts, synthesises corrections and uses 3D monitoring to re-attempt only the affected stage.
What is an atomic correction?
It is a small local repair—regrasp or realign—rather than restarting the whole task. Preserving progress can improve efficiency but requires reliable stage detection.
Why does this matter?
Author-reported preprint; 11 HF upvotes.
- Failure data comes from policy behaviour rather than random perturbations.
- Results cover simulation and real dual-arm tasks.
- Safety under unseen failures remains an open deployment question.