Figure AI - Helix 2.5 Robot: Zero-Shot Home Generalization

Helix 2.5 is a neural-policy humanoid system pretrained on Index, a massive dataset of human behavior, and evaluated for zero-shot whole-body autonomy across 30 unseen Bay Area homes. A single foundation model was fine-tuned into three long-horizon behaviors - tidying living rooms (picking 13-15 scattered toys into a basket), folding towels, and making beds (placing pillows and smoothing the comforter) - that require integrated perception, locomotion, bimanual manipulation, and active perception. Evaluation used no data from the homes or objects, a single fixed checkpoint per task, and strict pass/fail criteria that granted success only for fully completed tasks; no environment-specific adaptation or further tuning was allowed.

Index pretraining was the dominant factor driving transfer: otherwise-identical policies trained from scratch achieved 9% zero-shot success, while Index-initialized policies reached 56%. Helix 2.5 matched prior Helix 02 success using half the task-specific data and displayed robust whole-body self-correction (repositioning, stance changes, moving around obstacles). Pretraining scale followed a precise human-to-humanoid transfer law: doubling Index data predictably reduced downstream action-prediction loss and allowed forecasting the largest run’s loss to four decimal places (forecast error 0.54% of observed variation). Figure positions this as first evidence that broad human pretraining can enable generalizable humanoid behavior and is committing major data and compute to scale further.