What Is Figure's Project Go-Big?

Project Go-Big is Figure’s large-scale humanoid pretraining data-collection initiative, announced in a company blog post on September 18, 2025. Figure describes it as an effort to build what the company calls the world’s largest and most diverse humanoid pretraining dataset, intended to feed Helix, Figure’s Vision-Language-Action (VLA) model for generalist humanoid control.

Every Project Go-Big fact on this page traces to Figure’s own announcement at figure.ai/news/project-go-big. No other coverage, estimate, or third-party report of Project Go-Big is used as a source anywhere on this site, and no hour count, throughput figure, or customer name is attributed to Firsthand based on the Project Go-Big announcement. Firsthand has no relationship with Figure or Brookfield; this page is an independent explainer of Figure’s public statements.

The announcement frames the project around a problem this site’s other pretraining guides describe the same way: robotics has no large-scale equivalent of ImageNet for vision, Wikipedia for language, or YouTube for generative video — no existing internet-scale corpus of robot behavior to pretrain on. Figure’s stated bet is that egocentric human video can substitute for that missing corpus, because a humanoid’s perspective and kinematics mirror a human’s closely enough that skills learned from watching people can transfer directly.

Figure states that Project Go-Big is accelerated by a partnership with Brookfield Asset Management, under which Figure collects egocentric human video inside real Brookfield properties. Figure’s announcement describes Brookfield’s asset base as including more than 100,000 residential units worldwide, alongside roughly 500 million square feet of commercial office space and about 160 million square feet of logistics space, and states that data collection has already begun in Brookfield environments with plans to keep scaling.

Figure does not publish an hour count, episode count, or completion date for Project Go-Big in this announcement, so this page does not state one. The scale language Figure uses is about property access — the number and variety of real-world spaces the Brookfield partnership opens up — not a measured dataset size.

Figure’s announcement argues that humanoid robots have a structural advantage other robot form factors lack: a humanoid’s perspective and kinematics mirror a human’s, which Figure says makes it possible to transfer knowledge directly from everyday human video rather than requiring costly robot demonstrations, hand-coded programs, or staged environments. That argument mirrors the viewpoint-matching rationale this site gives for humanoid pretraining data generally: a humanoid perceives the world from a head-mounted camera, so egocentric human capture recorded from that same vantage point matches the robot’s observation space far more closely than third-person video or single-arm teleoperation.

Figure’s stated comparison to other pretraining breakthroughs — ImageNet, Wikipedia, YouTube — frames egocentric human video as robotics’ closest available substitute for an internet-scale corpus, since no equivalent library of first-person robot-task video exists to pretrain on directly.

Figure’s announcement reports one concrete result tied to Project Go-Big data: training Helix on 100% egocentric human video, collected passively in real Brookfield homes, with no robot demonstrations at all, produced a navigation capability Figure calls a zero-shot human-to-robot transfer. Figure states that Helix can now respond to conversational commands such as "walk to the kitchen table" or "go water the plants," generating closed-loop navigation control from pixels and language through cluttered home environments — and that this is, to Figure’s knowledge, the first time a humanoid robot has learned this end-to-end, from images and language to low-level velocity commands, using only human video with no robot-specific data or training.

Figure also states that a single Helix network now outputs both high-rate dexterous manipulation and navigation commands, which the announcement frames as eliminating the need for separate, task-specific or data-source-specific systems. Figure notes that Helix had previously focused on upper-body manipulation tasks — laundry folding, dishwasher loading, package reorientation — and that the navigation result is a new capability layered onto that same unified model.

This page states only the capability Figure’s announcement itself reports — zero-shot navigation transfer from 100% human video, with no robot data. It does not extend that claim to manipulation tasks, does not state a success rate or benchmark number, and does not claim Project Go-Big data was used for anything beyond what Figure’s post describes.

Project Go-Big joins a set of public pretraining references this site covers elsewhere — EgoScale, GEN-0, and GR00T N1 — each making a different kind of egocentric-pretraining claim from a different source type. Project Go-Big is a company blog announcement, like GEN-0, rather than a peer-reviewed paper like EgoScale or GR00T N1, so it states a partnership, a data-collection approach, and one reported capability, without publishing the benchmark tables or training methodology a research paper would include.

What distinguishes Project Go-Big’s reported result is embodiment and task: it is specifically a humanoid navigation result, learned end-to-end from human video with zero robot-specific training, rather than a manipulation scaling law. See /guides/what-is-gen-0 and /guides/what-is-nvidia-egoscale for how those two other announcements and papers frame their own scale and results.

Project Go-Big is further public evidence for a claim this site’s humanoid pretraining pages already make: egocentric human video, captured from a head-worn viewpoint that matches a humanoid’s own perspective, is a viable pretraining source for humanoid foundation models, and real-world property access and task diversity matter as much as raw scale. Figure’s own framing leans on partnership-driven access to varied real environments, not a single large but narrow capture, which lines up with the diversity-first planning this site recommends for any humanoid pretraining collection.

It is also a reminder that Project Go-Big is Figure’s own initiative, built for Figure’s own Helix model, through a property partnership that is specific to Figure and Brookfield. A team that does not have that kind of partnership, and needs humanoid pretraining data scoped to its own robot’s embodiment, task list, and environments, has to source that data itself — through a public corpus, a custom collection, or some combination of the two. See /custom-collection/humanoid-pretraining-data for how this site describes bimanual, whole-body egocentric capture for humanoid pretraining, planned for diversity across homes, workplaces, tasks, and people rather than matched to any one announced initiative’s numbers.

A custom humanoid pretraining collection starts with a written skill spec covering the scenes, tasks, objects and people the model needs to cover, captured from the same head-worn, bimanual, whole-body viewpoint a humanoid perceives from, followed by a fast, on-demand pilot batch scored against that spec before scaling to full volume. See /custom-collection/pretraining-data-collection for how that process runs end to end, and /guides/robot-pretraining-data for a broader introduction to how pretraining data differs from fine-tuning or post-training data.

That is the same due-diligence sequence this site recommends for any pretraining data source, whether the reference point being planned against is a research paper like EgoScale, an announced training run like GEN-0, or a partnership-driven initiative like Project Go-Big. None of them settle the question of exactly how much data any specific humanoid needs — each is evidence that egocentric human video helps, from a different team and a different evaluation setup — so a pretraining plan should start from the target robot’s own coverage gaps, not from matching any single announced project’s access or scale.