Reproduce “VLA-$Δ$: Learning Physical State Transition for Vision-Language-Action Models” 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
Qi Lu, Jiarun Li, Ce Zhang, Miao Pan, Yi Pan, Xu Wang, Wenqi Zhang, Xuhong Zhang
VLA-Δ trains a vision-language-action model to represent changes in task progress, object motion, spatial relations, and contact, then uses that representation to generate actions. Its transition targets supervise training but are not decoded at deployment. It reports 98.40% on LIBERO and a 5.5-percentage-point advantage over π₀.₅ across five real-robot tasks, without leading every task.
Key insights
- Structured state descriptions, localized visual motion, and natural-language descriptions of change provide complementary transition supervision.
- Transition learning can improve action performance even when downstream fine-tuning and inference do not use the auxiliary prediction objectives.
- Robustness gains vary by perturbation: VLA-∆ leads on LIBERO-Plus language variation but not on background or layout changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.