Humanoid robot company Figure has released Helix 2.5, a new-generation neural network that enables robots to autonomously complete household chores in previously unseen environments without local data collection or fine-tuning. During testing across 30 unfamiliar homes in the San Francisco Bay Area, robots successfully performed tidying, towel folding, and bed-making tasks using zero-shot learning. Pretrained on Figure's human behavior dataset Index, Helix 2.5 achieved a 56% zero-shot full-task success rate in blind tests, up from 9% for models trained from scratch. The system requires only half the task-specific data compared to representative Helix 02 tasks. Figure's Index dataset currently ingests approximately 35 minutes of human behavior data per second, supported by a $3.5 billion compute commitment for Helix training.