HumanScale vs GR00T N1: Human Data in Robot Pretraining

HumanScale and GR00T N1 both put human video into robot pretraining, but they answer different questions. HumanScale asks whether egocentric human video can match or beat teleoperated real-robot data as a pretraining source. GR00T N1 asks how to train one humanoid foundation model on a layered mixture that includes human video, real-robot trajectories and synthetic data.

Neither is a Firsthand product. Firsthand has no relationship with NVIDIA or the HumanScale authors, and Firsthand data was not used to train GR00T N1 or in the HumanScale study. For the single-source explainers, see /guides/what-is-humanscale and /guides/gr00t-n1-data-pyramid-explained.

HumanScale is a research paper on arXiv, titled “HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining”. Its abstract 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 whose effectiveness the paper says was underexplored.

The authors run a systematic study under fixed post-training and validation protocols. With the same amount of pretraining data, models pretrained on egocentric data, processed through a carefully designed filtering and labeling pipeline, 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. The figures come from the abstract on the paper’s arXiv page, the official source.

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. It is 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”.

GR00T N1 does not collect its own human video. The paper sources “diverse human egocentric video data from open datasets”, drawing on seven: Ego4D, Ego-Exo4D, Assembly-101, EPIC-KITCHENS, HOI4D, HoloAssist and RH20T-Human. Human video sits at the wide base of the GR00T N1 data pyramid alongside web data, as laid out in /guides/gr00t-n1-data-pyramid-explained.

HumanScale’s abstract names egocentric human video processed through a filtering and labeling pipeline, but does not name the source corpus, the capture rig or the filtering method. The paper is the place to read those. This page does not state them, because the abstract does not.

In GR00T N1, robot data and human video are layers of one mixture. Real-robot trajectories sit at the narrow top of the pyramid, including 88 hours of in-house teleoperation, and the paper expands that teleoperation to 827 hours of neural trajectories, with 780,000 simulation trajectories as another layer. The mixture is the point, so the paper does not frame human video and robot data as rivals.

HumanScale frames them as rivals first, then as partners. It compares the two as pretraining sources at equal data, and in its reported setup human video came out ahead. Its recommended paradigm is still not to drop robot data: 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. The general trade-off is described in /compare/human-video-vs-robot-data-for-pretraining.

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.

HumanScale attributes its result to egocentric data “when processed through a carefully designed filtering and labeling pipeline”. The reported advantage is a property of processed human video, not raw footage, which makes label quality part of the dataset. This site documents what action supervision for human video looks like in /guides/hand-pose-and-wrist-action-labels-for-pretraining.

GR00T N1 states scale in hours and trajectories: human-video hours per dataset in its Table 7, 88 hours of in-house teleoperation, 827 hours of neural trajectories and 780,000 simulation trajectories. The HumanScale abstract states that both arms used the same amount of pretraining data. That is a design condition, not a total, and the abstract gives no hour count.

So the two do not compare on size. The shared point is diversity: HumanScale cites low behavioral and environmental diversity as a limit of teleoperated data, 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.

No. HumanScale’s results come from a controlled comparison: validation loss on real-robot action prediction and success rates on real-robot task execution, human video against teleoperated data at equal pretraining data. 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 percentages answer different questions. Reading GR00T N1 as confirming HumanScale’s figures, or the reverse, goes beyond what either source states.

For HumanScale, the abstract does not state a corpus size, an hour count, a capture rig, a filtering method, the tasks, a licence or a delivery format. For GR00T N1, the human video is not its own capture, so capture rig and consent terms belong to the seven source datasets, and the sections reviewed include no run that removes human video.

Neither source describes how an outside team would obtain data scoped to its own robot. Both are self-reported research, and neither replaces your own evaluation on your embodiment.

HumanScale is a study, and its abstract states no licence or availability path. 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. Ask which layer of your mixture the data fills, how clips are labeled and whether action signals are measured or inferred, whether streams are synchronized, and how much robot data alignment will need for your embodiment. See /custom-collection/pretraining-data-collection and /guides/robot-pretraining-data.