Figure posts 4 straight hours of Helix 2.5 in 30 homes
Brett Adcock introduced the unedited footage by warning it was really boring, and paired it with fresh upload numbers from Figure's Index app.
Figure has released four straight hours of its humanoid running Helix 2.5 in 30 homes with no additional training on those houses. The footage follows this week's Helix 2.5 write-up and is the long-form version rather than a cut demo. Brett Adcock posted it with a line calling it really boring, which is roughly the point.
Olivia Lee posted the same four hours as a scaling argument, writing that Figure has the robots, the data and the compute and that it is time to scale up. Adcock replied "Great work Olivia".
The data flywheel numbers
Adcock separately posted this week's figures from the Index app, which pays people to film themselves doing everyday tasks from their own point of view. He listed 53 minutes of real-world data uploaded every second, 2.4 million video uploads in the week, and 115,000 weekly active users. He calls it the data flywheel powering Helix and argues no other robot form factor can learn from human video at the same scale, because the robot shares our body shape.
A summary by @itsolelehmann, which Adcock quoted, spells out the jump. The robot learned how humans move and handle objects from those videos, then got extra training on three specific jobs. Put in 30 homes it had never seen, with towels, beds and toys it had never seen, it made beds, folded towels and tidied toys into a basket. Without the human-video pretraining the same robot finished the job 9% of the time. With it, 56%.
@itsolelehmann also says the more human video goes in, the better the robot gets at predicting the next move, along a curve you can predict. That predictability is the part Figure and Lee are selling, more than any single chore.
Shown versus claimed
What is genuinely new today is the unedited footage and the updated upload counts, not a new capability claim. Figure says the runs are zero-shot in unseen homes. Nobody outside Figure has independently reproduced that, and the number Figure's own material points to is 56%, so the robot still fails close to half the time. As @itsolelehmann puts it, nobody is getting a housekeeper this year.
One caveat on the app numbers. Adcock lists 115,000 weekly active users, while @itsolelehmann's summary says about 90,000 people a week are uploading. Both are in the same range, and neither has been checked by anyone outside Figure.
If you want to watch something really boring: here's 4 hours of our humanoid doing zero-shot work across 30 rental homes
Our humanoid robots are learning from the largest installed dataset in the world: humans operating in the real world
it still fails almost half the time, so nobody is getting a housekeeper this year