RT-2-X

RT-2-X is Google DeepMind's vision-language-action model trained on the Open X-Embodiment dataset, released October 2023 alongside the dataset and the smaller RT-1-X. The project pooled data from 22 robot embodiments across 20 institutions: over 1 million episodes covering 500+ skills and 150,000 tasks, released as the largest open real-robot dataset of its kind. RT-2-X builds on RT-2's web-scale vision-language pretraining, giving it emergent semantic understanding: it can interpret abstract commands and transfer web knowledge into robot actions it was never explicitly trained for.

DeepMind reported RT-2-X achieved 3x performance improvement over models trained per-robot, demonstrating that cross-embodiment training generalizes better than narrow specialists. Both the dataset and the trained models were open-sourced, seeding years of follow-on robotics research. Exact release date is era-level (October 2023) and flagged as such.