TL;DR A single end-to-end trained model that learns to map robot observations to actions while benefiting from large-scale pretraining on web language and vision-language data. Appeared in surveys 2026-03-30-cross-embodiment-manipulation