Helix 2.5 enables Figure humanoid robots to perform three full body behaviors in previously unseen rental environments without site specific training data. The model reached a 56% zero shot success rate across 30 Bay Area homes, an increase from the 8% success rate found in versions without pretraining on Figure's Index data engine. The system allows robots to enter a new location and immediately begin work using whole body coordination.

The Index system currently produces 50 minutes of training data per second and serves more than 100,000 active weekly users. Figure has committed $3.5 billion of compute for the next iteration of the Helix model to further scale the pretraining of robot behaviors in the real world.