The research question addresses how to enable Vision-Language-Action (VLA) models to recover from failures during deployment using limited corrective data.

The proposed approach, OopsieRecover, utilizes abduction to infer the intent behind a few corrective demonstrations and encodes this knowledge into dense spatial representations and conditional costs to guide the policy towards more desirable actions.

The authors report experiments across six tasks with varying object poses, scenes, and related tasks, showing that a small number of corrective demonstrations can help VLA models handle a wide range of new situations without needing additional corrective data for each scenario.