EgoVerse vs GR00T N1: Human Data in Robot Foundation Models

EgoVerse is described in arXiv 2604.07607, “EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World”. The abstract calls it “a collaborative platform for human data–driven robot learning that unifies data collection, processing, and access under a shared framework”. The current release includes “1,362 hours (80k episodes) of human demonstrations spanning 1,965 tasks, 240 scenes, and 2,087 unique demonstrators”.

The paper pairs the dataset with a study: “a large-scale study of human-to-robot transfer with experiments replicated across multiple labs, tasks, and robot embodiments under shared protocols”. See /guides/what-is-egoverse for the single-paper explainer.

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. See /guides/gr00t-n1-data-pyramid-explained for the pyramid in detail.

No. EgoVerse is mainly a data resource and an experiment about how human data transfers to robot policies. GR00T N1 is a trained model whose paper describes a data strategy. They answer different questions: what a human-demonstration corpus contains and how well it transfers, versus how a foundation model blends human video with other layers.

The sources link only lightly. EgoVerse’s related-work section cites GR00T N1 among foundation models that build on pseudo-labelled web video. The GR00T N1 paper, from March 2025, does not mention EgoVerse. Neither source describes any shared data. This page leads with the two-way comparison; for the other pairings see /compare/egoscale-vs-gr00t-n1, /compare/egoscale-vs-egoverse and /compare/egoverse-vs-humanscale.

EgoVerse collects its own human data, in two parts. EgoVerse-A follows a controlled, standardised protocol on Project Aria glasses (Gen 1), a head-worn device the paper describes as 75 g. EgoVerse-I comes from industry partners using custom head-mounted rigs; the paper says “most systems use stereo fisheye RGB cameras to achieve accurate hand pose estimation”, with depth streams where available and inertial sensing to reconstruct camera motion. A phone-based pipeline uses an iPhone on a head strap at 1080p and 30 FPS.

GR00T N1 does not collect its own egocentric video. Its human-video layer comes from seven existing datasets, which the paper says “do not include explicit action labels”. Capture hardware therefore varies by dataset and the paper does not restate it.

EgoVerse states its scale in one place: 1,362 hours, 80k episodes, 1,965 tasks, 240 scenes and 2,087 demonstrators. Its appendix splits that into EgoVerse-A (75 hours, 6 flagship tasks) and EgoVerse-I (two industry partners, 1,035 and 250 hours). The paper stresses “bounded diversity”: tasks feasible for bimanual mobile manipulators, with natural variation across scenes, objects and demonstrators.

GR00T N1 does not summarise its human-video hours in the text reviewed; the paper points to Table 7 for per-dataset statistics. Its other stated numbers describe the rest of the pyramid: 88 hours of in-house teleoperation expanded to 827 hours of neural trajectories, and 780,000 simulation trajectories described as equivalent to 6,500 hours of human demonstration data. The pyramid is organised by quantity, with the most data at the human-video base.

EgoVerse attaches pose signals to every episode. Per frame it estimates 3D hand pose using 21 keypoints per hand, paired with a calibrated 6-DoF head pose from visual–inertial SLAM. In the transfer study these “serve as proxies for human end-effector motion, enabling cotraining across embodiments”. EgoVerse-A adds per-episode task descriptions, scene identifiers, primary manipulated objects and demonstrator metadata. EgoVerse-I adds 1–2 second language descriptions, active-hand indicators and static-versus-mobile manipulation flags.

GR00T N1 has no action signal in its human video, so it manufactures one. 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”. The repository adds that N1.7 uses a relative end-effector action space shared across robot and human embodiments.

EgoVerse uses the simplest recipe: co-training. A shared encoder and flow-matching action decoder are trained on the union of human and robot data, with a conditional flow-matching loss for each. Robot data is collected by teleoperation, about 150–300 demonstrations per task, and the study varies how much human data is added. The paper states plainly that its study “mainly focused on human-and-robot co-training” and leaves pre-train-then-fine-tune to future work. No synthetic data is described.

GR00T N1 uses a layered recipe. Large quantities of web data and human video form the base, synthetic data (simulation and neural trajectories) the middle, and real-robot data the top. All layers are co-trained in pretraining, then the model is post-trained on a single embodiment, with neural trajectories sampled 1:1 against real trajectories in the low-data case. The pyramid exists, per the paper, to address “data islands” across robot embodiments.

See /compare/gen-0-vs-gr00t-n1 and /glossary/data-pyramid for the broader strategy.

EgoVerse reports three findings across labs, tasks and robots. Co-training with human data “leads to clear and reproducible performance improvements”, by up to 30% in the main experiments. The benefit of scaling depends on aligned human–robot data that share task semantics and scene context, and different forms of diversity contribute unevenly. One task on one robot got worse.

GR00T N1 reports results against baselines. At 100 demonstrations per task it averages 45.0% against 33.4% for Diffusion Policy across three simulation benchmarks. On the GR-1 humanoid, full-data success averages 76.8% against 46.4%, and the paper says 10% of the data gets within 3.8 points of the baseline trained on everything. It also reports that the pretrained checkpoint reached 76.6% and 73.3% on two generalisation tasks.

Neither result isolates the same thing. EgoVerse tests human data directly. GR00T N1 tests a whole recipe; the sections reviewed include no run that removes human video, so its figures do not measure that layer alone.

The Isaac-GR00T repository lists the Apache License 2.0, and the README says GR00T N1.7 is “fully commercially licensable under Apache 2.0”. The paper states that 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.

EgoVerse states no data licence in the paper text reviewed. It describes EgoDB, with “controlled data access” and a way to sync filtered subsets to local machines, but it does not state public release terms. A team planning a product on EgoVerse data should ask the authors and get commercial-use terms in writing.

EgoVerse leaves open pretraining at scale and open-ended generalisation. The authors say broader algorithmic work such as pretrain-and-fine-tune is future work, and that the diversity experiments use offline metrics, so further robot rollouts are needed to confirm they transfer. Its licence and public availability are not stated in the text reviewed.

GR00T N1 leaves open the contribution of human video, the human-data licence chain and long-horizon work. The paper says the model “focuses primarily on short-horizon tabletop manipulation tasks” and names long-horizon loco-manipulation as future work. It also notes that synthetic generation still struggles with diversity and physical realism. Its action labels for human video are inferred rather than measured.

Custom collection is not a substitute for either source. It fills coverage that a fixed corpus and a fixed pyramid cannot. EgoVerse reports that scaling depends on aligned human–robot data. A program might use either source as an input, then commission a fast, on-demand pilot batch scoped to the tasks, scenes and failure cases its own model is missing.

Firsthand collects custom egocentric video to a buyer’s skill spec, tests the spec on a small Spec Pilot batch before scaling, and delivers in RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd. The default licence is buyer-owned, perpetual and non-exclusive, confirmed in writing before capture. See /guides/robot-pretraining-data for how pretraining and post-training divide the work, and /custom-collection/pretraining-data-collection to request a custom pretraining dataset.