Figure hits 56% in 30 unseen homes

Figure says Helix 2.5 completed whole-body household tasks in 56% of zero-shot trials across 30 Bay Area homes, with no data collection, fine-tuning, or adaptation in the evaluation homes or on their objects, according to Figure’s evaluation report. Its from-scratch comparison scored 9%, but “zero-shot” still covers three behaviors specified with training data elsewhere and the result is company-reported. Robotics operators should file this as evidence that broad human-behavior pretraining can transfer, not as proof of reliable household autonomy; compare the archive’s physical-AI capital thesis before budgeting deployment.