Figure Project Go-Big vs EgoScale

Project Go-Big is a humanoid pretraining data-collection initiative that Figure announced in a company blog post on September 18, 2025. Figure says it is building what it calls the world’s largest and most diverse humanoid pretraining dataset for Helix, accelerated by a partnership with Brookfield. EgoScale is a research paper from NVIDIA’s GEAR lab (arXiv 2602.16710) that trains a VLA on 20,854 hours of action-labeled egocentric human video and reports a log-linear scaling law.

Both treat egocentric human video as a way to pretrain robots. They differ in source type, in the task they report on, and in how much detail they publish. Neither is a Firsthand product, and Firsthand has no relationship with Figure, Brookfield or NVIDIA. For the longer single-source explainers, see /guides/what-is-figure-project-go-big and /guides/what-is-nvidia-egoscale.

Figure states that Project Go-Big is accelerated by a partnership with Brookfield, which it describes as owning over 100,000 residential units worldwide, alongside 500 million square feet of commercial office space and 160 million square feet of logistics space. It says data collection has already begun in Brookfield environments and will continue to scale.

The post reports one result. Figure trained Helix on 100% egocentric human video, collected passively as people do behaviors in real Brookfield homes, with no robot demonstrations, and Helix now responds to commands such as "Walk to the kitchen table" by generating closed-loop navigation control from pixels. Figure says that, to its knowledge, this is the first time a humanoid has learned end-to-end from images and language to low-level SE(2) velocity commands using only human video. It also says a single Helix network now outputs both manipulation and navigation commands.

The post does not state an hour count, an episode count, an action-label method, a licence or a benchmark number, so this page does not either.

EgoScale pretrains on 20,854 hours of egocentric human video, more than 20 times larger than prior efforts according to the paper. The majority is in-the-wild egocentric recording across household, industrial, retail and educational settings. Action labels come from off-the-shelf SLAM and hand-pose estimation, which the paper says are noisy because collection is unconstrained, while the scale and diversity of the data still provide effective supervision.

The recipe has two stages. Large-scale human pretraining comes first. A lightweight mid-training stage then uses about 50 hours of human data and 4 hours of robot data across 344 tabletop tasks, with human demonstrations collected using the same camera configuration as the robot, matched viewpoints and calibrated intrinsics. The paper reports a 54% average success-rate improvement over a no-pretraining baseline on a 22-DoF dexterous hand.

No. Figure’s reported result is a navigation capability learned from human video with no robot data. EgoScale’s reported result is dexterous manipulation, reached with human pretraining plus a small amount of aligned robot data. Different embodiments, different tasks, different evaluation setups and different levels of published detail mean the 54% figure and the navigation result answer different questions.

Read them as two independent pieces of evidence for the same broad idea: human egocentric video carries transferable signal for robots. Neither says how much data your robot needs.

EgoScale states a log-linear relationship between human data scale and validation loss, and says that loss correlates strongly with real-robot performance. The paper is explicit that it does not extrapolate beyond the range it measured.

Project Go-Big states an intent to scale collection across Brookfield environments and frames diversity of real-world spaces as the goal. It does not publish a scaling curve. For how published scaling claims differ from announcements, see /compare/egoscale and /guides/why-dataset-diversity-matters-in-robot-learning.

EgoScale’s structure is the clearer template: broad human-video pretraining first, then a small aligned mid-training dataset closely matched to the target robot. That mirrors the stage split in /guides/robot-pretraining-data. Figure’s post adds that, for navigation, human video alone produced a transfer result, but it presents that as an initial learning result rather than a general recipe.

For the wider trade-off between human video and robot-collected data, see /compare/human-video-vs-robot-data-for-pretraining.

Project Go-Big is Figure’s own initiative for Figure’s own model, built on a property partnership specific to Figure and Brookfield. EgoScale’s corpus is a fixed research mixture. A team that needs data scoped to its own robot, tasks and environments has to source it separately: a public corpus, a custom collection, or both.

A custom collection starts with a written skill spec covering the scenes, tasks, objects and people a corpus is missing, proven with a fast, on-demand Spec Pilot before volume scales. Capture is head-worn, bimanual and whole-body for humanoid pretraining, delivered under a perpetual, buyer-owned commercial licence with a documented consent chain, in RLDS, LeRobot, WebDataset, HDF5, zarr and Rerun .rrd. See /custom-collection/humanoid-pretraining-data and /custom-collection/pretraining-data-collection.

Ask how action labels were produced, whether the viewpoint matches your robot, whether sensor streams are synchronized, how coverage was planned against a spec, and whether the licence permits commercial training with documented consent. Neither Project Go-Big nor EgoScale answers those for outside buyers, which is why the checklist in /guides/robot-pretraining-data-diversity-plan starts from your robot’s coverage gaps rather than from either announced scale.