Project Go-Big

Figure's plan to pretrain its humanoid robot using massive amounts of first-person human video.

Project Go-Big is a data initiative the U.S. humanoid robot company Figure AI announced on September 18, 2025, aimed at building “internet-scale” humanoid pretraining data for its VLA (vision-language-action) model, Helix. Figure partnered with the asset management firm Brookfield, which provides access to more than 100,000 residential units along with large amounts of office and logistics space, used to collect first-person video of people doing everyday things in real homes. Figure says Helix learned to navigate a cluttered home from language instructions — going straight from images and language to chassis velocity commands — using human video alone, and describes this as the first time a humanoid robot has learned this capability end-to-end purely from human video; navigation and manipulation were also merged into the same Helix network. It represents the strategy of substituting human video for part of the teleoperation data a humanoid would otherwise need.

ExampleA user says “walk over to the kitchen table,” and Helix outputs walking-speed commands directly from the camera feed; the training data behind this capability came entirely from first-person human video, with no robot demonstrations involved.

Also called

Go-Big

Related

Figure Helix · Figure AI · Human Video Data · Egocentric Video · Zero-shot · Vision-Language-Action Model

Sources

Project Go-Big: Internet-Scale Humanoid Pretraining（Figure 官方） (Chinese)

As of

2025-09