OpenVLA

The 7B open vision-language-action model (CoRL 2024, 6.6K★) that outperformed RT-2-X and became the most-forked open VLA of 2024-25 — Stanford-led with Berkeley co-leads (Levine, Pertsch) plus TRI and Google DeepMind co-authors, trained on 970K Open X-Embodiment trajectories. Its successor line (π0 / openpi) ships under Physical Intelligence and is excluded per the spin-out rule.