The ACT-2 Preview is presented as the first robotics model to combine broad generalization with high reliability, capable of learning new behaviors from a single fine-tuning example. This model achieves a 99% success rate in zero-shot scenarios within real, unseen home environments.