InSight: Self-Guided Skill Acquisition via Steerable VLAs

Summary

Vision-language-action (VLA) models can learn manipulation skills from demonstrations, but their capabilities are bounded by the skills in the training data.