Figure Project Go-Big vs GR00T N1

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. GR00T N1 is NVIDIA’s “open foundation model for humanoid robots”, described in arXiv 2503.14734 and released through the github.com/NVIDIA/Isaac-GR00T repository.

Both put human video in the training picture for humanoids. The difference is what each source shows. Project Go-Big reports one navigation result and no dataset size. GR00T N1 documents a model and a layered data recipe. Neither is a Firsthand product, and Firsthand has no relationship with Figure, Brookfield or NVIDIA. Firsthand data was not used to train either. For the single-source explainers, see /guides/what-is-figure-project-go-big and /guides/gr00t-n1-data-pyramid-explained.

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.

GR00T N1 is described as an open foundation model for humanoid robots and as a vision-language-action model with a dual-system design: a vision-language module interprets the scene and instruction, and a diffusion transformer generates motor actions. The abstract says it is trained “with a heterogeneous mixture of real-robot trajectories, human videos, and synthetically generated datasets”.

The repository README now describes GR00T N1.7, a later release that the README says adds 20K hours of EgoScale human video to pretraining. This page uses the original paper for N1 and labels anything taken from the README as N1.7. The paper’s own scope note says the model “focuses primarily on short-horizon tabletop manipulation tasks”.

They sit at different points in a pipeline. Project Go-Big is about getting human video at scale in real homes for Helix. GR00T N1 is about how a model combines layers once data exists. The GR00T N1 data pyramid puts web data and egocentric human video at the wide base, synthetic data in the middle and real-robot trajectories at the narrow top, as laid out in /guides/gr00t-n1-data-pyramid-explained.

Project Go-Big, as described, belongs to the base of that picture: human video collected without a robot. Figure does not say whether or how it plans to blend that video with other layers, so this page does not either. How the layers are mixed in practice is covered in /guides/building-a-robot-pretraining-data-mixture.

GR00T N1 states that its human videos have no explicit action labels. It trains a VQ-VAE on pairs of frames to produce a latent-action label, and for neural trajectories also uses a trained inverse dynamics model to infer pseudo-actions. For human videos the paper uses “learned latent actions as flow-matching targets”. Terms are defined in /glossary/latent-actions and /glossary/inverse-dynamics-model.

Project Go-Big’s post says Helix produces low-level SE(2) velocity commands from pixels, but it does not describe how the human video becomes a training signal. A buyer should not assume the same latent-action method, or any other, for Figure’s data.

No. Figure’s result is zero-shot, language-commanded navigation of homes from human video alone. GR00T N1’s results are against baselines: at 100 demonstrations per task it averages 45.0% against 33.4% for Diffusion Policy across three simulation benchmarks, and on the GR-1 humanoid full-data success averages 76.8% against 46.4%. Different tasks, different measurements and different evaluation setups mean the two answer different questions.

Neither result isolates human video on its own against robot data. Project Go-Big runs no robot-data arm, and the GR00T N1 sections reviewed include no run that removes human video, so its figures do not measure that layer alone. Reading either as proof of what human video replaces goes beyond what the source states.

They do not compare directly. GR00T N1 states 88 hours of in-house teleoperation expanded to 827 hours of neural trajectories, and 780,000 simulation trajectories, with human-video hours per dataset in its Table 7. Project Go-Big states property access, which describes how many real spaces a partnership opens up, not how much recorded data exists. Figure says only that collection has begun and will continue to scale.

The shared point is diversity. Figure frames diversity of real-world spaces as its goal, and GR00T N1 draws on seven human-video datasets. 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, and the post describes no way for outside teams to obtain its data. The Isaac-GR00T repository lists the Apache License 2.0, and the paper states the GR00T-N1-2B checkpoint, training data and simulation benchmarks are publicly available. The human videos it draws on come from other datasets, each with its own terms, which the paper does not restate.

A team that plans to train a commercial model on either needs to confirm terms directly with the publisher, in writing, before use. A fixed public corpus also cannot be reshaped afterwards to your target tasks, embodiment or coverage gaps.

A team that wants human data scoped to its own robot, tasks and environments has to source it separately from 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. The pilot process is described in /custom-collection/egocentric-data-collection-pilot.

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 which layer of your data mixture the data fills, how clips are labeled and whether action signals are measured or inferred, 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 source answers those for outside buyers. See /guides/robot-pretraining-data-diversity-plan.