DataPlatter tackles a critical challenge in robotic manipulation by optimizing data collection for Vision-Language-Action models to improve performance while minimizing costs.

• Identifies spatial reasoning in large workspaces as the primary source of failures
• Strategically collects low-cost data for these challenging scenarios
• Improves generalization capabilities of robots across varied tasks
• Provides a cost-effective approach to training more capable robotic systems

This research offers significant value for engineering and manufacturing by reducing the data acquisition bottleneck in robotic system development, enabling more adaptable and capable automation solutions for complex manipulation tasks.

DataPlatter: Boosting Robotic Manipulation Generalization with Minimal Costly Data