Figure on Sept. 17, 2026, introduced Helix 2.5, the AI that runs its humanoid robot, and said the robot carried out three long household chores in 30 Bay Area homes it had never seen, with no data collected in any of them. Whole-body work in homes usually fails when the furniture, layout and objects change. Figure is arguing that pretraining on large amounts of human behavior can carry skills into unfamiliar rooms without extra training on site. Pretraining is the first, broad stage of training, when a model learns from a huge collection of text, images or video.
Figure calls the result zero-shot, but that does not mean the robot learned the chores from nothing. Figure says the homes and the objects the robot handled were new to it, while the three chores themselves were still taught through fine-tuning, extra training that adapts an already trained model to a specific job, using data collected elsewhere. Each chore ran on one fixed version of the AI, which was not adjusted in the test homes. The chores were tidying a living room, folding towels and making a bed.
In Figure's comparison, with the chore data, design, training and testing held the same, a version trained from scratch finished whole chores 9% of the time, while the version pretrained on Index, Figure's collection of human-behavior data, finished them 56% of the time, more than six times as often. Figure says a chore counted only if the robot finished all of it, with no partial credit, and that a safety intervention counted as a failure. It also says Index is generating about 35 minutes of new human experience every second, and that it has committed $3.5 billion of computing power to training Helix.
Figure also says that across an 8-fold increase in Index data, its AI got steadily better at predicting a robot's next action on examples it had not trained on, so steadily that the company could forecast its largest run from smaller ones. That measures how well the AI predicts the next move, not how often the robot finishes chores.
By one news report's count of the trials Figure disclosed, the robot succeeded in 237 of 420 attempts. It made the bed in 94 of 140 tries, about 67%, folded towels in 87 of 140, or 62%, and tidied the living room in 56 of 140, or 40%. The report classed the results as internal testing, not real-world deployment. A 56% success rate still means the robot failed on roughly 44% of attempts by Figure's own standard. Another report called Helix 2.5 a research milestone rather than a consumer robot, and noted that the 56% versus 9% comparison comes from Figure and has not yet been peer-reviewed.
It is not yet known whether outside labs can match the 30-home results, how often safety stops would come up in longer runs without supervision, or whether better next-move prediction will turn into more finished chores in messier homes.