EgoScale vs EgoVerse: Egocentric Pretraining Data Compared

EgoScale is a research program from NVIDIA’s GEAR lab documented in a paper at arXiv 2602.16710 and on NVIDIA’s own GEAR lab research page at research.nvidia.com/labs/gear/egoscale/. It studies whether pretraining a robot foundation model on large-scale egocentric human video improves downstream manipulation performance, and how that improvement scales with data volume. It is not a downloadable dataset release documented on this site — it is a published research recipe built around a 20,854-hour egocentric video corpus.

See /guides/what-is-nvidia-egoscale for the full single-source explainer, and /compare/egoscale for a side-by-side against a custom, commercially licensed pretraining collection.

EgoVerse is described in the paper "EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World" at arXiv 2604.07607. The paper reports roughly 1,362 hours of egocentric human demonstrations across approximately 80,000 episodes, spanning 1,965 tasks, 240 scenes and 2,087 unique demonstrators, contributed by a consortium of academic labs and industry partners.

The dataset splits into EgoVerse-A, collected under a standardized protocol across participating academic labs using Project Aria Gen 1 glasses for controlled study, and EgoVerse-I, sourced from industry partners using custom wearable rigs and a phone-based pipeline with denser, language-based annotations. See /guides/what-is-egoverse for the full single-dataset explainer.

EgoScale is primarily a research recipe: a two-stage training methodology (broad human pretraining, then a small aligned human-robot mid-training stage) built around a 20,854-hour corpus, studied for what it reveals about scaling behavior rather than released as a standalone downloadable dataset documented on this site. EgoVerse is primarily a dataset: a published collection of roughly 1,362 hours of annotated human demonstrations, intended to be used directly by other researchers.

That distinction matters for how each is used. EgoScale’s value to a buyer is mainly as a citable reference point for how egocentric-video scale affects pretraining loss and downstream task success. EgoVerse’s value is mainly as a source of annotated human-demonstration footage with documented hand-pose and language labels, subject to its own unstated license.

EgoScale’s paper describes its action labels as derived from off-the-shelf SLAM (for camera trajectory and scene geometry) and hand-pose estimation, applied to aggregated egocentric video rather than captured live on a purpose-built, hardware-synchronized rig.

EgoVerse’s paper documents 3D hand pose using 21 keypoints per hand, paired with a calibrated 6-DoF head pose from visual-inertial SLAM, captured across three different rigs: Project Aria Gen 1 glasses, custom industry wearables, and a phone-based pipeline. EgoVerse-I episodes add fine-grained (1–2 second) language descriptions, active-hand indicators and static-versus-mobile manipulation flags.

EgoScale reports a log-linear scaling law between human pretraining data volume and validation loss, and ties that trend to a 54% downstream success-rate improvement on a 22-degree-of-freedom dexterous hand versus no pretraining — a quantitative, measured result tied to a specific evaluation setup.

EgoVerse’s paper does not publish the same kind of scaling-law curve in the material cited on this site. Its stated finding is about transfer rather than scale: co-training benefits from human data depend on how closely that data aligns with the target robot’s own task semantics — a caution that applies to EgoScale’s corpus as much as EgoVerse’s.

Neither source states a published open commercial license in the material cited on this site. EgoScale’s paper and GEAR lab page do not document a standalone open dataset license for the 20,854-hour corpus — any release terms for the underlying data would need confirming directly with NVIDIA before commercial use. EgoVerse’s arXiv paper does not state a license anywhere in the text this site cites.

For an early-stage robotics team, an unconfirmed or unstated license is a real blocker for a funded product, not a detail to defer. A custom collection closes that gap directly: Firsthand’s default is a buyer-owned, perpetual, non-exclusive commercial license, confirmed in writing before capture begins. See /guides/buyer-owned-training-data for why that distinction matters.

EgoScale is most useful as a citable, quantitative reference point for how egocentric-video scale affects pretraining loss and downstream manipulation success — evidence for the general principle that scale helps, measured on one embodiment and task set. EgoVerse is most useful as a reference for task, scene and demonstrator diversity, and for its documented hand-pose and language annotation style, while its lack of a stated license remains an open question for any commercial use.

Neither substitutes for a program’s own coverage plan. Both point toward the same conclusion this site makes throughout its pretraining guides: write a diversity spec across scenes, tasks, objects and people before setting an hour target, using /guides/robot-pretraining-data-diversity-plan as a starting framework.

Custom collection is not a competitor to EgoScale or EgoVerse — it is how a program fills the specific coverage gap that neither a research recipe’s fixed corpus nor a published dataset’s own scope can close for a particular buyer’s tasks and environments. A program might treat EgoScale’s scaling-law result as a benchmark for how scale affects loss, use EgoVerse as a reference for task and demonstrator diversity, and then commission a custom, commercially licensed batch scoped to the exact scenes, tasks and failure cases its own model is missing.

Firsthand delivers custom batches in RLDS, LeRobot, WebDataset, HDF5, zarr and Rerun .rrd, so a custom collection mixes cleanly alongside data referenced from either source at any pipeline stage. See /guides/robot-pretraining-data for how pretraining, mid-training and post-training divide that work, and /custom-collection/pretraining-data-collection for the full engagement model.