ModAR reaches 75% on bimanual robot tasks

CMU’s ModAR world-action model reached a 75% observed average success rate versus 72% for pretrained Flex-π while using roughly 20× fewer training FLOPs and no pretraining, according to the authors’ arXiv paper. It outperformed baselines on three real-world bimanual tasks, but that small, author-reported evaluation does not establish general robot-policy transfer. Operators should file it beside NEURA’s architecture-to-delivery gap and demand task-level hardware evidence before buying a world-model claim.