Figure Helix 2.5

2026-09-17 · Publisher : Figure AI · Domain : Robotics

Figure unveils Helix 2.5, pretrained on its Index human behavior dataset: the robot tidies living rooms, folds towels and makes beds in 30 unseen homes, with zero data collected on site.

### Background On September 17, 2026, Figure AI unveiled Helix 2.5, which it calls the most advanced neural network it has built. The question: can a humanoid walk into a home it has never seen and get to work right away, with its whole body, on its own? ### Specifications - **Pretraining**: on Index, the global-scale human behavior dataset Figure introduced in August 2026. - **Behaviors**: a single base model adapted into three whole-body behaviors: tidying a living room, folding towels, making a bed. - **Test**: 30 San Francisco Bay Area homes, with no data collection, fine-tuning or adaptation on site. Figure calls it the first demonstration of zero-shot whole-body generalization at this scope. - **Index contribution**: holding task data, architecture and evaluation fixed, Index pretraining lifts zero-shot success from 9% to 56% (results published by the company). - **Efficiency**: half the task-specific data of a representative Helix 02 behavior, for a 30x broader scope. - **Scaling law**: doubling pretraining data improves action prediction smoothly enough to forecast the largest run's loss to four decimal places. ### Why it matters Figure shows that whole-body intelligence can be learned from human experience and transferred to new places, instead of being rebuilt for each site. The company acknowledges general humanoid robotics is not solved. ### Key takeaway Helix 2.5 makes large-scale human video the fuel for humanoid generalization.

Tags : VLA, Humanoïdes, Zéro-shot, Figure AI, Données humaines