Reproduce “Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment” on OpenResearch

About this paper

Jacky Kwok, Xilun Zhang, Mengdi Xu, Yuejiang Liu, Azalia Mirhoseini, Chelsea Finn, Marco Pavone

CoVer-VLA helps a vision-language-action robot select actions for user instructions under wording variation. It generates scene-grounded rephrasings and multiple action chunks for each, then uses a learned verifier to score them against the original instruction, observation, and recent action history. On seven challenging SIMPLER tasks, it averages 57.0% ID and 61.0% OOD success, versus 44.0% and 48.7% for instruction augmentation; gains are not universal.

Key insights

• Scaling verification at test time can be a more effective strategy for improving Vision-Language-Action alignment and robustness than solely scaling policy learning.
• A hybrid sampling approach, combining instruction rephrasing with repeated action sampling, significantly reduces action error more efficiently than scaling either dimension independently.
• Contrastive learning with rephrase augmentation enables robust semantic alignment assessment without requiring explicit negative or failure examples, addressing a key challenge in robotics data.

Summary by alphaXiv.