264,000 downloads in 108 countries, more than 44,000 weekly contributors and 30 minutes of video ingested every second: the dataset must feed the Helix system.
After four months of catimini testing to verify that it was possible to collect physical data directly from humans and on a large scale, Figure renamed its Index application and published it on the App Store and Google Play. The company presents it as the largest truly usable training data set for robotics.
The principle is simple: a candidate contributor receives a recording device, filmed in subjective view of the approved tasks — folding linen, passing the serpillar, tidying a kitchen — and is paid at the minute taken. Figure claims 264,000 downloads across 108 countries, more than 44,000 active users per week, and an intake rate of about 30 minutes of video per second.
The diversity displayed counts as much as the volume: for 1,000 hours collected, Index lists 373 separate tasks, 1,146 different objects handled and 116 environments. The set must be used to drive Helix, the vision-language-action system that controls the brand's humanoids.
Read the full article on the original source
Go to the source →