Figure Unveils Helix 2.5 Humanoid AI with Zero-Shot 30-Home Generalization

Figure has introduced Helix 2.5, its most advanced neural network to date, designed to enable humanoid robots to operate in unfamiliar environments without prior environment data.

The foundation model demonstrated zero-shot whole-body autonomy across 30 previously unseen Bay Area homes. The humanoid executed three long-horizon behaviors: tidying living rooms, folding towels, and making beds.

Pretrained on Index, Figure's global-scale dataset of human behavior, Helix 2.5 adapted to these distinct tasks spanning locomotion, rigid and deformable manipulation, bimanual coordination, and active perception.

In controlled blind evaluations, Index pretraining alone increased zero-shot task success from 9% to 56% compared to a baseline model trained from scratch on identical task data.

Figure stated:

"The point is not that general humanoid robotics is solved. But Helix 2.5 is the first evidence that whole-body intelligence can be learned from human experience and transferred to new scenarios, rather than rebuilt each time."

The company also reported that Helix 2.5 matched the performance of its prior Helix 02 policy while using half as much adaptation data.

Furthermore, Figure demonstrated the first measured human-to-robot transfer scaling law on a humanoid platform. Pretraining data scaling predictably improved downstream next-action prediction, allowing researchers to forecast test loss to four decimal places prior to training.

Index currently generates approximately 35 minutes of human experience every second. Backed by a $3.5 billion compute commitment, Figure aims to continue scaling its physical AI architectures globally.