GigaWorld-1 builds world models optimized for evaluating robot policies

Summary

GigaWorld-1 is a systematic study of world models for robotic policy evaluation using WMBench, which analyzes seven video world models and over 324,000 policy rollouts paired with real robot executions. The research shows that long-horizon action-faithful rollout consistency matters more than short-term visual realism, and releases code, models, and datasets for embodied foundation models.