RESample: Data Augmentation Framework for Robotic Manipulation
RESample is a coverage-guided data augmentation framework designed to address failure recovery in Vision-Language-Action (VLA) models for robotic manipulation. VLA models, trained on large-scale imitation learning datasets, often fail when execution deviates from standard demonstrations due to distributional shift. RESample actively supplements demonstration datasets with failure modes that may occur in the real world. It trains a conservative coverage function to identify failure cases within the actual data distribution, guiding augmentation to improve policy robustness. The framework was introduced in a paper on arXiv (ID: 2510.17640v4) and is relevant to the fields of AI and robotics.
Key facts
- RESample is a coverage-guided data augmentation framework.
- It addresses failure recovery in Vision-Language-Action (VLA) models.
- VLA models are trained on large-scale imitation learning datasets.
- These datasets mostly contain successful trajectories, lacking corrective supervision.
- Execution deviations lead to distributional shift and policy failures.
- RESample trains a conservative coverage function to identify failure cases.
- The framework supplements demonstration datasets with failure modes.
- The paper is available on arXiv with ID 2510.17640v4.
Entities
Institutions
- arXiv