# Figure’s Project Go-Big made human-video provenance part of its Helix pretraining claim

Figure announced Project Go-Big on September 18, 2025, pairing a humanoid pretraining-data initiative with a vendor claim of zero-shot human-video-to-robot navigation.

Canonical: https://rangoon.ai/insights/archive-figure-project-go-big-september-2025/
Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/)
Published: 2026-10-24
Updated: 2026-10-24
Event date: 2025-09-18
Tags: [Figure](https://rangoon.ai/insights/tags/figure/), [Helix](https://rangoon.ai/insights/tags/helix/), [humanoid robotics](https://rangoon.ai/insights/tags/humanoid-robotics/), [vision-language-action](https://rangoon.ai/insights/tags/vision-language-action/)
Source note: Historical reporting: Figure announced Project Go-Big on September 18, 2025, describing a humanoid pretraining-data initiative and a claim that Helix could translate egocentric human video into robot navigation without robot-specific training for that result. Figure’s September 17 Brookfield partnership announcement is immediate primary context for the stated data-collection relationship. Data scale, transfer performance, “first” language and navigation outcomes are Figure’s claims, tied to its collection method and system; neither source establishes general rights to human footage, a public dataset, a public Helix release or independently verified deployment performance.

Image: Human-video-to-robot training boundary. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied.

## Key takeaways
- Project Go-Big was a data and pretraining announcement, not a release of a public humanoid model or dataset.
- Human-video-to-robot transfer depends on collection provenance, task alignment and the vendor’s evaluation conditions.
- Figure’s zero-shot navigation and scale statements are supplier claims, not independent proof of broad household autonomy.

## The September event paired a data-collection plan with a transfer claim [source 1](https://www.figure.ai/news/project-go-big)

On September 18, 2025, Figure announced Project Go-Big, which it described as an internet-scale humanoid-pretraining effort, alongside a claim of zero-shot transfer from egocentric human video to robot navigation. The company said Helix was trained on human video for the reported navigation result and framed that work around natural-language instructions in cluttered spaces. This is a meaningful AI-robotics event because it connects a training-data strategy to a specific model behavior. It does not make human video interchangeable with robot demonstrations, establish that every human recording transfers to a robot action, or show that the resulting policy will behave correctly in an arbitrary home. The event is a vendor statement about its data and learning approach.


## The adjacent Brookfield announcement defined a collection relationship, not a public corpus [source 1](https://www.figure.ai/news/project-go-big) [source 2](https://www.figure.ai/news/figure-announces-strategic-partnership-with-brookfield)

One day earlier, Figure announced a partnership with Brookfield that it said would support real-world humanoid-pretraining data collection and AI infrastructure for Helix. Figure described human video capture in Brookfield environments as a route to data about navigation and manipulation. This source is important context because it identifies a commercial relationship and the kind of collection environment the company associated with Project Go-Big. It does not publish a dataset, a license, participant-consent terms, retention policy or a right for other teams to train on the footage. Data provenance is more than a headline count: a responsible technical record needs the collection scope, permissions, exclusions, metadata, labeling route and controls over how recordings enter training.


## Human-video transfer remained a model and evaluation boundary [source 1](https://www.figure.ai/news/project-go-big)

Figure said its reported navigation result used egocentric human video rather than robot-specific data and called the result zero-shot human-to-robot transfer. The phrase describes the company’s experimental setup and outcome, not a general rule that a robot can infer safe actions from any video. A humanoid has different sensors, control loops, physical limits and failure consequences from a person wearing a camera. Transfer also depends on what the model sees, how instructions map to actions, whether a scene matches training distribution and how the deployment handles ambiguity. The September release supports attributing Figure’s stated method and result to Figure. It does not independently establish its “first” characterization, the completeness of its video coverage or a performance guarantee in unseen buildings.


## Data scale increased the need for traceable selection and evaluation [source 1](https://www.figure.ai/news/project-go-big) [source 2](https://www.figure.ai/news/figure-announces-strategic-partnership-with-brookfield)

A larger pretraining initiative can broaden visual and behavioral diversity, but its practical value comes from how a team governs the path from recording to model update. Teams need to distinguish passive recordings from demonstrations, retain source and consent boundaries, inspect whether labels express intended behavior, and hold out environments for evaluation. They also need separate criteria for a navigation demonstration and a physical deployment near people or property. Figure’s announcements provide useful evidence that the company intended to link human-centric data collection with Helix training and commercial settings. They do not establish that scale itself solves distribution shift, reduces a deployment’s operational risk, or makes every data subject and downstream use appropriately governed. The public record is a starting point for scrutiny, not a completed assurance case.


## Sources
1. [Figure: Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer](https://www.figure.ai/news/project-go-big)
2. [Figure: Figure Announces Strategic Partnership with Brookfield](https://www.figure.ai/news/figure-announces-strategic-partnership-with-brookfield)

## Related reading
- [Figure’s Helix split language grounding from a 200 Hz humanoid control loop](https://rangoon.ai/insights/archive-figure-helix-february-2025/)
- [Physical Intelligence’s π*0.6 used deployment experience and corrections as VLA training inputs](https://rangoon.ai/insights/archive-physical-intelligence-pistar06-recap-november-2025/)
- [π0.5 tested a single VLA on home-cleaning tasks outside its training scenes](https://rangoon.ai/insights/archive-physical-intelligence-pi05-april-2025/)

## More information

- [Documentation index](https://rangoon.ai/llms.txt)
- [AI access and policies](https://rangoon.ai/ai/)
- [Source repository](https://github.com/hypler-dev/rangoon)