Reproduce “Unified Visuomotor Targets: Supervising VLAs Beyond Physical Actions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhenyang Feng, Unnat Jain

Researchers at UC Irvine developed Unified Visuomotor Targets (UVT), a latent representation blending robot actions and visual dynamics from Latent Action Models, which serves as a refined supervision signal for Vision-Language-Action (VLA) models. This strategy improves VLA training efficiency, enhances robustness, and boosts performance on manipulation tasks by up to 16.2% in early training stages, without modifying existing VLA architectures.

Key insights

  • Supervising VLA policies with a unified latent target that integrates both motor control and visual dynamics information is more effective than direct low-level action supervision.
  • A joint encoding of action and dynamics within a single latent space (UVT) yields significantly larger performance gains and better final performance compared to using auxiliary dynamics prediction heads.
  • The UVT space creates a more structured, task-dependent representation for robot actions, implicitly filtering out low-level actuator noise and leading to smoother, more stable decoded actions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.