SimVLA enables zero-shot robot learning from simulation

A team led by Kyoungin Baik released SimVLA, a framework that trains vision‑language‑action (VLA) models entirely on simulated robot data. The system first learns from two large simulated datasets—SimAction, covering 35 mobile‑manipulation tasks, and SimVQA, which adds detailed visual‑language supervision. After additional post‑training on simulated rollouts, SimVLA was tested on real‑world tasks such as restocking shelves, pouring liquids, and cleaning. In these tests it performed better than policies trained on 50 real‑world demonstrations, showing that simulation can replace costly real‑world data collection.