Figure Unveils Helix 2.5 With Zero-Shot Humanoid Generalization Across 30 Homes
aiinsider ·
Figure introduced Helix 2.5, a neural network intended to let its humanoid robot perform whole-body household tasks in homes it has not seen, without collecting local data or adapting the model there. The company tested the same checkpoint in 30 Bay Area homes on living-room tidying, towel folding and bed making, with success requiring completion of every step and with evaluation objects absent from task-specific training data. Figure reported that Index pretraining increased zero-shot success from 9% for a from-scratch model to 56% under an otherwise identical comparison. It also said Helix 2.5 used half as much task-specific data while extending behavior specification from one environment to 30 unseen homes, and that the robot could self-correct during longer tasks. In a separate scaling test, robot-action prediction improved predictably as Index human-behavior data doubled, which Figure described as a human-to-robot transfer scaling law for a humanoid. Index reportedly generates about 35 minutes of human-experience data per second, while Figure has committed $3.5 billion in computing resources to training Helix.
aiinsider