HuRo: 630K robotized human-video episodes lift VLA completion from 51.5% to 80.3%

RLWRLD · hf · 2026-09-22

RLWRLD's HuRo pipeline converts heterogeneous human videos into robot-aligned observations and action trajectories, building a dataset of 630K robotized episodes and 142M frames from five human-video sources for VLA pretraining. Across four real-world manipulation tasks, more robotized pretraining data raises overall completion from 51.5% to 80.3% and OOD completion under spatial/visual shifts from 34.9% to 72.2%. Ablations show visual robotization improves OOD robustness and end-to-end retargeted-action pretraining beats visual-only transfer.

Original post →

More from Embodied

Embodied channel →