@terrymillerThe Bay Area saw 30 unseen homes where Figure tested Helix 2.5, an AI model for humanoid robots in unknown residences, without collecting new data or retraining. The bots executed tasks like organizing toys, folding towels, and preparing beds across those unfamiliar settings. Results show a 56% zero-shot success rate for Helix 2.5 versus 9% for a model lacking Index pretraining. Outcomes indicate that training on extensive human behavior data enhances generalization to new settings.

@terrymillerThe Bay Area saw 30 unseen homes where Figure tested Helix 2.5, an AI model for humanoid robots in unknown residences, without collecting new data or retraining. The bots executed tasks like organizing toys, folding towels, and preparing beds across those unfamiliar settings. Results show a 56% zero-shot success rate for Helix 2.5 versus 9% for a model lacking Index pretraining. Outcomes indicate that training on extensive human behavior data enhances generalization to new settings.