Figure says its Helix 2.5 control model carried out three household behaviors in 30 homes it had not seen during training, without collecting data or updating the model at those locations. In the company's blind comparison, pretraining on human-behavior videos raised end-to-end success from 9 percent to 56 percent over an otherwise matched model trained from scratch.
The test asks whether a humanoid can transfer one learned behavior across new layouts and objects. Figure adapted the same Index-pretrained foundation model to tidy living rooms, fold towels and make beds using task-specific robot data collected elsewhere. Each task then used a fixed checkpoint across all homes; no evaluation rollout was used to choose or modify it.
Index is Figure's pipeline for collecting videos of people performing physical activities. Helix 2.5 was pretrained from random initialization on that dataset, then fine-tuned for the three robot behaviors. To isolate pretraining, Figure held the model architecture, optimization, hyperparameters and task-specific data constant while comparing random and Index-pretrained starting weights.
Success required completing the whole assigned behavior. The robot had to place all 13 to 15 toys in a basket, fold and basket all towels, or position both pillows and both comforter corners at the top of a bed. The homes' own beds, couches and work surfaces were used, and Figure says the evaluation objects did not appear in the task-specific data. A safety intervention or timeout counted as failure.
Figure also trained four models on nested Index subsets spanning an eightfold data range. More pretraining data reduced the error in predicting robot actions excluded from training. That result concerns action prediction, not household task success. Separately, Figure says Helix 2.5 matched a previous location-specific Helix 02 behavior with half as much adaptation data.
The evidence is still a company-run evaluation of three tasks in 30 Bay Area homes. Figure has not released the model, training dataset, trial-level results, confidence intervals or an independent replication, and it does not report how often safety interventions occurred.
If those results hold, the practical change is that a robot behavior could be specified once and deployed across varied homes without gathering fresh demonstrations at each address. The current test supports that direction, while its 56 percent end-to-end success also shows that zero-shot household autonomy remains far from routine reliability.