DyPES-VLA: cross-embodiment robot manipulation hits 98% on LIBERO

Junfeng Li · hf · 2026-08-07

DyPES-VLA learns shared dynamics priors and embodiment-specific control for cross-embodiment manipulation. It trains VLM with future prediction, and uses embodiment-specific MoE action head to output native actions without manual alignment. Achieves SOTA: 98.0% on LIBERO, 59.25% on RoboCasa-GR1, 89.02% on RoboTwin 2.0.

Original post →

More from Embodied

Embodied channel →