ZJU's VLA-Δ trains robots to predict physical state changes, hits 98.4% on LIBERO
量子位 · wechat · 2026-11-15
Zhejiang University's OmniAI team presents VLA-Δ, which makes physical state changes (task progress, spatial displacement, contact transitions) explicit pretraining targets via Physical State Difference Queries supervised by structured transition slots, dense optical flow, and future-state difference text. Built on a 2B vision-language backbone and 484 hours / 24M frames of DROID, MolmoAct and RH20T demos, it achieves 98.4% average success on LIBERO and 68.0% across five real dual-arm task categories, beating π₀.₅ by 5.5 points. Paper, code and project page are open.
More from Embodied
- Bindu Reddy makes the bull case for a SpaceX–Tesla merger: robots and space data centers — bindureddy · 2026-11-15
- Scale CEO Alexandr Wang teases a 'superintelligent' pip-boy device — alexandr_wang · 2026-11-15
- TIER IV and Komatsu to put autonomous dump trucks into service by 2027 in Japan and North America — 4310sy · 2026-11-15
- Optimus robot reportedly flew on Starship, returned and was caught by the tower — xiaohu · 2026-11-15
- China flies 10-ton all-electric CE-25A passenger plane with 19 seats and autonomous flight mode — saibharadwaj · 2026-11-15
- OOMWOO: open-source DIY robot vacuum hits 13.1k GitHub stars with Pi, LiDAR and ROS2 — Technovangelist · 2026-11-15