Helix 2.5 30-Home Generalization

In this talk

1. 0:00

   Testing Helix in an unfamiliar home

   Figure introduces Helix 2.5 as a system designed to work in homes it has not encountered before. The demonstration includes tidying a living room, making a bed, and folding towels, using objects absent from the robot’s prior experience.

2. 2:20

   Results across 30 homes

   Figure says it deployed Helix 2.5 in 30 Bay Area homes without collecting data in those homes and reported success in each. Corey describes the system as a single model handling room-scale tasks involving perception, movement, and manipulation of both rigid and soft items.

3. 3:49

   Human experience as training data

   Corey explains that Figure created INDEX to collect human experience for robot training, since robotics lacks internet-scale datasets. He reports more than 90,000 weekly contributors and 35 minutes of human experience uploaded each second. In a controlled comparison, strong generalization appeared only in the model whose pre-training included INDEX.

4. 4:54

   Evidence of scaling and plans to expand

   Figure says robot action prediction improved as the dataset was repeatedly doubled, and that the final validation loss could be forecast to four decimal places before the training run. The company cautions that robot learning is not solved, while committing $3.5 billion in compute to Helix.