Ropedia W0 trains robot world models on 10,000 hours of human video, lifting LIBERO-Plus by 6.76 points

jiqizhixin · x · 2026-11-09

A preprint from Ropedia, Nanyang Technological University and collaborators presents W0, a video world model spanning locomotion and manipulation:

Training data: combines 10,000 hours of human video with 3,000 hours of robot video; camera trajectories and projected skeletons provide a shared motion-conditioning representation across different bodies, linking an initial observation and text to predicted future egocentric video.
Results: initializing the FastWAM video backbone with the 5B W0 model raises reported LIBERO-Plus success from 50.06% to 56.82% — a 6.76 percentage-point gain with the action head and training recipe unchanged.
Design: embodiment-native kinematics via forward kinematics, Plücker-ray and skeleton maps as conditioning; the video world model and the downstream action system play distinct roles, validated in transfer experiments.

More from Embodied