Ropedia W0 preprint reports real-time, minute-long robot world simulations

Researchers report that Ropedia W0, a model that predicts the consequences of human and robot actions, supports real-time, minute-long simulated sequences after distillation, enabling live VR interaction. The preprint combines locomotion and manipulation across different human and robot body configurations. The team assembled 10k hours of egocentric human data and 3k hours of robot data. Its shared representation describes movement through camera poses and manipulation through projected skeletons. Evaluations span five cross-embodiment benchmarks and held-out human video; the researchers report that increasing human pre-training data improves both human and robot results. Constructing robot skeletons requires extensive manual keypoint annotation and correction.