Figure's robot did chores in 30 homes it had never seen
Figure tested Helix 2.5 in 30 homes with no prior data. The robot tidied rooms, folded towels and made beds, succeeding 56% of the time.
What was tested
On September 17, humanoid robot company Figure introduced Helix 2.5, a new neural network. It was tested in 30 Bay Area homes where no data had ever been collected, with no extra training, fine-tuning or adaptation to the homes or objects.
The result
One foundation model produced three long tasks: tidying living rooms, folding towels and making beds. Figure says pretraining on Index, its dataset of human behaviour, raised zero-shot success in unseen homes from 9% to 56% against an otherwise identical model trained from scratch.
The data race
Figure says Index now generates about 35 minutes of new human experience every second and that it has committed $3.5 billion of compute with Nscale to train Helix. A 56% rate means the robot finishes the job in an unfamiliar home roughly every other try. Figure calls it the first demonstration at this scope on a humanoid.