VOMMI: Collecting and Leveraging Portable Demonstrations for Mobile Manipulation
This study addresses the scarcity of mobile manipulation data and the difficulty of acquiring low-cost motion supervision by proposing VOMMI, a framework that efficiently converts portable RGB demonstrations into post-training signals for Vision-Language-Action (VLA) models. Without requiring robot calibration, the method employs R2-VO for offline trajectory reconstruction and integrates sparse geometric anchors with causal local motion tokens to enable online visual-motion conditioning. Furthermore, an action-group residual adapter is introduced to enhance policy expressiveness. Experimental results demonstrate that, compared with baselines, VOMMI reduces velocity error by 18.2% and trajectory error by 24.6%, while improving task success rate by 8.3 percentage points over OpenPI.