Figure introduced Helix 2.5, its most advanced neural network. It enables zero-shot whole-body autonomy across 30 real homes. One foundation model powers three behaviors: tidying, folding towels, and making beds. Index pretraining alone raised zero-shot success from 9% to 56%. Helix 2.5 matched prior success rates using half the task-specific data. It also demonstrated a human-to-humanoid transfer scaling law. Index now generates 35 minutes of new human experience every second. Figure committed $3.5B in compute to train Helix. The system is the first to achieve zero-shot whole-body generalization on a humanoid at this scale.
What changed
Robots learned each new environment one at a time with environment-specific data.
What it unlocks
A humanoid robot can enter an unseen home and perform whole-body tasks without prior data collection or fine-tuning.
- 30 unseen homes
- 9% to 56% zero-shot success
- 2x less task-specific data
- 35 minutes of new human experience per second
Sources