VLA-Δ: Zhejiang University Adds Physical State Transition Supervision to VLA Models
OmniAI-ZJU · hf · 2026-10-16
The OmniAI group of Zhejiang University's ACES Lab proposes VLA-Δ, addressing the weakness of VLA models that map observations directly to demonstrated actions without supervising intermediate physical state changes, making them prone to overfitting observation-action correspondences.
The framework uses physical changes between current and future states as explicit learning targets, building a large-scale pretraining dataset with physical state-transition annotations covering task-relevant semantics and visual motion changes, giving more direct supervision for continuous action generation. At inference the model generates actions directly from current observations with no added inference cost. VLA-Δ is competitive with SOTA on LIBERO and RoboTwin and improves real-world transfer.
More from Embodied
- ELSFP pain points: >3dB insertion loss, costly PMF, harsh packaging requirements — jwt0625 · 2026-10-16
- Agility's Digit 5 pairs Nyobolt fast-charging battery: 90 min work, 9 min charge — xmercury_one · 2026-10-16
- PINT: Shared Predictive Intent Tokens Unify Reasoning, World Modeling, and Robot Actions — Xinyu Guo · 2026-10-16
- TouchScale releases 500 hours of unified egocentric vision-tactile data for scaling study — jiqizhixin · 2026-10-16
- Ghost raises $11M for Core, a personal AI computer that runs locally in your home — nick_linck · 2026-10-16
- Robot shows up at Costco to fetch the famous $1.50 hotdog — majidmanzarpour · 2026-10-16