Boosting Robotic Manipulation Generalization with Minimal Costly Data

Understand

The growing adoption of Vision-Language-Action (VLA) models in embodied AI intensifies the demand for diverse manipulation demonstrations.

• However, high costs associated with data collection often result in insufficient data coverage across all scenarios, which limits the performance of the models.
• It is observed that the spatial reasoning phase (SRP) in large workspace dominates the failure cases.
• Fortunately, this data can be collected with low cost, underscoring the potential of leveraging inexpensive data to improve model performance.