EgoScale vs GR00T N1: Training Data Compared

EgoScale is a research study from NVIDIA’s GEAR lab (arXiv 2602.16710, with a project page at research.nvidia.com/labs/gear/egoscale) that asks a narrow question: how far does pretraining on egocentric human video scale for robot dexterous manipulation? It pretrains on 20,854 hours of action-labeled first-person human video, then aligns the result to a robot with a much smaller mid-training stage of roughly 50 hours of human video and about 4 hours of robot data across 344 tabletop tasks.

Its headline findings are a log-linear scaling law between human-video pretraining scale and validation loss, which the paper reports correlating with downstream success, and a 54% average task-success improvement on a 22-degree-of-freedom dexterous hand compared with no pretraining. For the single-paper explainer see /guides/what-is-nvidia-egoscale; for how EgoScale compares to Firsthand-style capture, see /compare/egoscale. This page only compares its training data with GR00T N1’s.

GR00T N1 is an open vision-language-action foundation model for generalist humanoid robots, described in NVIDIA’s paper arXiv 2503.14734. The paper describes a dual-system architecture: a vision-language module that interprets the scene and the instruction, and a diffusion-transformer module that generates motor actions, trained jointly end-to-end. The model is reported across multiple robot embodiments in simulation and deployed on the Fourier GR-1 humanoid for language-conditioned bimanual manipulation.

For training data, the paper is best known for its data pyramid: a wide base of web data and human video, a middle layer of synthetic data, and a narrow top layer of real robot trajectories. The layer-by-layer explainer lives at /guides/gr00t-n1-data-pyramid-explained. Here the pyramid matters only as the comparison point to EgoScale’s single-source recipe.

The core difference is breadth of source type. EgoScale’s pretraining corpus is one kind of data — egocentric human video — at large scale, with real robot data entering only in the small mid-training stage. The study’s bet is that a big, diverse, action-labeled human-video corpus can carry most of the learning before any robot is involved.

GR00T N1 spreads its training across heterogeneous sources and arranges them by how specific they are to the target robot. Web data and human video supply general visual and manipulation priors at the widest volume. Synthetic data supplies robot-specific dynamics and edge cases more cheaply than teleoperation. Real-robot trajectories supply the only data exactly matched to the target embodiment, at the smallest volume.

Put simply: in EgoScale, human video is the whole pretraining story; in GR00T N1, human video is one layer of a broader plan. A data buyer following EgoScale’s recipe is mostly sourcing human video. A buyer following GR00T N1’s recipe is sourcing human video for the base and also needs a synthetic pipeline and real-robot collection for the layers above it.

Both papers face the same obstacle: human video records no robot joint angles or gripper commands, because no robot was involved. They solve it in different ways.

EgoScale estimates actions after capture. It runs off-the-shelf SLAM to recover camera trajectory and scene geometry, and hand-pose estimation to recover the hands, then uses those estimates as action labels. That scales to tens of thousands of hours, but the labels are estimates derived from the video rather than measurements taken on a synchronized rig at capture time.

GR00T N1 learns actions from the video. It uses latent actions — a learned, compressed representation of what changed between frames — and inverse-dynamics pseudo-labels that infer the action most likely to explain an observed state change. That lets unlabeled human video contribute action-relevant signal without an explicit label on every frame. See /glossary/latent-actions and /glossary/inverse-dynamics-model.

For a buyer, both approaches reward the same thing: clean, consistent, well-calibrated footage. Estimated labels are only as good as the video they are estimated from, and pseudo-labels inherit the same dependence. Footage that already carries calibrated 3D hand pose and action segments gives either recipe a more direct supervisory signal than labels inferred later.

EgoScale uses a two-stage recipe. Stage one is large-scale pretraining on human video. Stage two is a small, aligned mid-training stage — around 50 hours of human video paired with about 4 hours of matched robot data across 344 tabletop tasks — that closes the embodiment gap. The heavy lifting is in stage one; stage two is short and tightly scoped.

GR00T N1 pretrains on its mixed pyramid and then post-trains for specific embodiments and tasks. The layers do not map one-to-one onto stages, but the general pattern is that pretraining leans on the base and middle layers, and post-training narrows toward real-robot data for the target skills. See /guides/pretraining-vs-post-training-robot-data for how this site separates the three stages.

Both shapes are wide at the bottom and narrow at the top. EgoScale achieves that shape with one source plus a small robot slice; GR00T N1 achieves it with several source types stacked by specificity. For a team building its own mixture, /guides/building-a-robot-pretraining-data-mixture covers how to choose proportions.

If your plan resembles EgoScale, your biggest procurement question is human video: how much, how diverse, and how it is labeled. The study’s scaling result suggests that more diverse human video keeps helping, so coverage across scenes, tasks, objects and people matters as much as total hours. You will still need a small aligned slice of robot data for your own embodiment.

If your plan resembles GR00T N1, human video is your base layer but not your whole budget. You also need a synthetic data pipeline for the middle layer and real-robot trajectories for the top. The base layer is usually the cheapest to scale and the easiest to commission against a written spec before a robot or simulator is ready, which makes it a natural first purchase.

In both cases, three questions decide whether a human-video source is usable: how were action labels produced, does coverage match your target task distribution, and is the licence written for commercial training? Neither paper answers the third question for your model; see /guides/how-to-license-training-data-for-commercial-models.

Yes, for the human-video part of each recipe. A custom collection supplies head-mounted, first-person human demonstration video captured to a buyer’s written skill spec — the scenes, tasks, objects and people a corpus is missing �� with calibrated, synchronized capture and labels such as 3D hand pose, action segments and visibility flags. That fits EgoScale-style pretraining directly and fits the base layer of a GR00T N1-style pyramid.

It does not replace the synthetic middle layer or real robot trajectories. What it can do is scope human demonstrations to the same tasks a robot team plans to teleoperate, so the human and robot slices of a mid-training or post-training stage line up. A Spec Pilot — a small first batch that proves the spec — comes before volume scales, and data ships on a buyer-owned commercial licence in RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd. See /custom-collection/pretraining-data-collection for the engagement model and /guides/robot-pretraining-data for how the stages divide the work.

Lean toward an EgoScale-style plan if your target is dexterous manipulation, you want most of your learning to come from scalable human video, and you can collect a small aligned robot slice for your own hand or arm. Your main investment is a large, diverse, well-labeled human-video corpus.

Lean toward a GR00T N1-style plan if you are training a generalist or humanoid policy across embodiments, you already run or plan to run simulation, and you can afford real-robot collection for the top layer. Your investment is spread across layers, with human video as the scalable base.

Many teams borrow from both: GR00T N1’s layered structure for the overall mixture, and EgoScale’s emphasis on large, action-labeled human video for the base. Whichever you choose, write a diversity plan before you set an hour target — /guides/robot-pretraining-data-diversity-plan is a starting framework — and treat both papers’ numbers as reference points rather than targets.