The new ACT-2 robotics model, introduced as a preview, is the first to combine broad generalization capabilities with high reliability, allowing a single fine-tuning example to teach new behaviors that generalize. It achieved a 99% success rate in zero-shot tasks within real, unseen home environments.