Figure Project Go-Big vs HumanScale

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. HumanScale is a research paper on arXiv, titled “HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining”.

Both put egocentric human video at the centre of pretraining. The difference is the evidence. Project Go-Big reports one navigation result and no comparison. HumanScale is built as a comparison, human video against teleoperated real-robot data. Neither is a Firsthand product, and Firsthand has no relationship with Figure, Brookfield or the HumanScale authors. For the single-source explainers, see /guides/what-is-figure-project-go-big and /guides/what-is-humanscale.

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.

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.

The abstract on the paper’s arXiv page frames the problem this way: teleoperated real-robot trajectories remain the dominant pretraining source for embodied foundation models because of their precise action supervision and embodiment alignment, but their scalability is limited by collection cost, acquisition difficulty and low behavioral and environmental diversity. Egocentric human video is a scalable, lower-cost and more diverse alternative, and the paper says its effectiveness against teleoperated data was underexplored.

To test it, the authors run a systematic study comparing the two as pretraining sources under fixed post-training and validation protocols. They report that egocentric data, when processed through a carefully designed filtering and labeling pipeline, can lead to superior performance. With the same amount of pretraining data, models pretrained on egocentric data achieved a 24% lower validation loss on real-robot action prediction, and 52.5% and 90% higher success rates on in-distribution and out-of-distribution real-robot task execution.

Other pages on this site describe HumanScale at the level of its research question. The figures here come from the abstract on the paper’s arXiv page, which is the official source. This page does not state a corpus size, capture rig, filtering method or licence for HumanScale, because the abstract does not. The paper itself is the place to read those.

Both lean toward human video, with very different strength of evidence. Project Go-Big’s result shows a navigation capability can be learned with no robot data at all, but it runs no robot-data arm, so it cannot say which is better. HumanScale is the source that tests the question directly, and in its reported setup human video came out ahead at equal pretraining data.

HumanScale does not argue for dropping robot data entirely. Its stated paradigm is to pretrain on egocentric human video to learn diverse world representations, then adapt with a small amount of labeled real-robot data for action-space alignment. That matches the trade-off described in /compare/human-video-vs-robot-data-for-pretraining: human video scales on a worn rig and a participant, while robot data aligns a policy to one embodiment.

No. Figure’s result is zero-shot, language-commanded navigation of homes from human video alone, output as low-level SE(2) velocity commands. HumanScale’s results are validation loss on real-robot action prediction and success rates on real-robot task execution, in a study built around a controlled comparison. Different tasks, different measurements and different evaluation setups mean the navigation result and the paper’s percentages answer different questions.

What they share is direction, not size. Reading Project Go-Big as confirming HumanScale’s percentages, or HumanScale as confirming Figure’s navigation claim, goes beyond what either source states.

The abstract attributes the result to egocentric data “when processed through a carefully designed filtering and labeling pipeline”. In other words, the reported advantage is a property of processed human video, not of raw footage. Project Go-Big’s post does not state how its video is turned into training signal.

For a buyer, that makes label quality part of the dataset, not an add-on. This site documents what action supervision for human video looks like, wrist-level pose plus articulated hand pose, in /guides/hand-pose-and-wrist-action-labels-for-pretraining. What a given program ships is set in its skill spec.

They do not, because neither source states an hour count in the material this page uses. Project Go-Big states property access, which describes how many real spaces a partnership opens up, not how much recorded data exists. The HumanScale abstract states that both arms used the same amount of pretraining data, which is a design condition rather than a total.

The shared point is diversity. Figure frames diversity of real-world spaces as its goal, and HumanScale cites low behavioral and environmental diversity as a limit of teleoperated data. A pretraining plan should start from your robot’s coverage gaps rather than a headline number. See /guides/why-dataset-diversity-matters-in-robot-learning.

Project Go-Big is Figure’s own initiative for Figure’s own model, built on a property partnership specific to Figure and Brookfield. HumanScale is a study, and neither source describes a way for outside teams to obtain its data. A team that wants data scoped to its own robot, tasks and environments has to source it separately.

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 clips are filtered and labeled, 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. Also ask how much robot data alignment will need for your embodiment. Neither source answers those for outside buyers. See /guides/robot-pretraining-data-diversity-plan.