By Dan Smigrod, Managing Editor, Humanoid Home News
Source: Figure
Context: Manufacturer research release describing zero-shot whole-body evaluations in 30 Bay Area homes.
Published: September 17, 2026
Figure says its Helix 2.5 robotics model completed whole-body household-task evaluations in 30 homes it had not previously seen. The company tested three behaviors: tidying a living room, folding towels and making a bed.
The important claim is not simply that a humanoid performed familiar chores. Figure says the same underlying model entered unfamiliar homes without collecting data, fine-tuning or adapting to those rooms or the objects used in the evaluations.
Index data is central to the result
Helix 2.5 was pretrained using Figure’s Index dataset of human activity. The company then adapted the model to the three evaluation behaviors using task-specific data collected elsewhere. None of the 30 evaluation homes or their test objects appeared in that task-specific data, according to Figure.
In Figure’s comparison, the Index-pretrained policy succeeded on 56% of zero-shot trials, while a policy trained from scratch succeeded on 9%. Figure defined success strictly: the robot had to finish the entire assigned task, such as putting every toy in a basket, folding and placing every towel, or positioning the pillows and comforter to complete the bed-making task.
The company also says the newer system matched the success rate of a representative Helix 02 behavior while using half as much task-specific data. Unlike that earlier behavior, Helix 2.5 was evaluated across unfamiliar homes rather than only where its task data had been collected.
Household variety remains the hard problem
Homes are difficult environments for robots because furniture, layouts, lighting and everyday objects vary from one residence to another. Tasks such as bed making also combine walking, perception, two-handed manipulation and recovery from mistakes over several minutes.
Figure reports that Helix 2.5 sometimes repositioned its body or moved around a bed to correct an incomplete action. That ability matters for future household robots, but the release does not establish consumer readiness or independent real-world reliability. The demonstrations and measurements were produced by Figure, and a 56% strict success rate still leaves substantial room for improvement.
Figure itself says general humanoid robotics is not solved. The results are better read as evidence that broad pretraining may reduce how much site-specific data a robot needs before attempting chores in a new home.
Humanoid Home News takeaway: Helix 2.5 shifts the household-robot question from whether a humanoid can learn one chore in one room to whether the same learned behavior can transfer across many homes. Figure’s 30-home test is meaningful manufacturer evidence, but the reported success rate and lack of independent validation show why supervised trials still matter.
