Preprint Open access
Imperfection for Precision: Upcycling Imperfect Data for High-Precision Robotic Manipulation
Training vision-language-action (VLA) models for high-precision manipulation typically requires task-specific, high-quality data (e.g., teleoperation), which is slow and expensive to collect. To reduce this burden without compromising manipulation precision, we propose $\varepsilon$4P (Imperfection for Precision), a si …