SIEVE: Structure-Aware Data Selection for VLA Imitation Learning

ZGCA · hf · 2026-07-08

SIEVE is a structure-aware data selection method for imitation learning in Vision-Language-Action (VLA) models. The core idea is to identify reusable visuo-motor primitives and their transition interfaces within datasets to boost policy learning efficiency. By using structured data selection logic, this method reduces training data redundancy and improves the generalization capabilities of robotic policies.

Original post →

More from Embodied

Embodied channel →