Egocentric Datasets for Robot Pretraining Compared

EgoScale, EgoVerse and Egocentric-10K are the three public egocentric video corpora referenced across this site’s compare and guide pages as pretraining sources for robot foundation models. Each has a single official source this site cites: EgoScale from its arXiv paper (2602.16710) and NVIDIA’s GEAR research page, EgoVerse from its arXiv paper (2604.07607), and Egocentric-10K from its Hugging Face dataset card (huggingface.co/datasets/builddotai/Egocentric-10K).

See /guides/what-is-nvidia-egoscale, /guides/what-is-egoverse and /guides/what-is-egocentric-10k for a single-dataset explainer of each, sourced strictly from that dataset’s own official page or paper.

EgoScale’s paper reports 20,854 hours of action-labeled egocentric human video. EgoVerse’s paper reports roughly 1,362 hours across approximately 80,000 episodes. Egocentric-10K’s Hugging Face dataset card reports roughly 10,000 hours. All three sit well below GEN-0’s reported 270,000+ hours, stated on Generalist AI’s official blog.

Hour count alone does not indicate task or environment coverage — a large corpus can still miss the specific skill or setting a model needs. See the comparison table above for how scale, label source and license differ across the three, and /guides/robot-data-scaling-laws for how this site frames the relationship between corpus scale and downstream model performance more generally.

EgoScale’s action labels are derived from off-the-shelf SLAM and hand-pose estimation applied to aggregated egocentric video — a post-hoc estimation pipeline, not a purpose-built capture rig. Its paper does not document hardware-triggered sync or a universal metric depth stream.

EgoVerse’s paper states its hour and episode counts but does not publish capture or label methodology beyond that, and this site does not infer one on its behalf. Egocentric-10K’s dataset card similarly states its hour count and Apache 2.0 license without publishing a capture or label methodology. Where a source does not state a fact, this page says so rather than estimating it.

Egocentric-10K is the clearest case: its dataset card states an Apache 2.0 license, which permits commercial use under that license’s terms. EgoScale’s paper does not state a commercial license for its underlying data release, so confirm the specific release’s current terms before commercial use. EgoVerse’s paper does not state a license at all.

See /compare/public-dataset-licenses for the fuller license comparison across public egocentric and robotics datasets referenced on this site, and always check each dataset’s current official page directly — license terms can change between releases.

Across all three corpora, none is captured to a named task or environment spec, none states a hardware-synchronized multi-sensor capture rig, and none publishes per-participant consent documentation. EgoScale’s hand pose is estimated post-hoc rather than captured live, and EgoVerse and Egocentric-10K do not state a hand-pose methodology at all.

See the second table above for how each of those gaps maps to what a custom pretraining collection adds — a written skill spec, a buyer-owned commercial license, synchronized depth/IMU/hand-pose streams, and documented per-participant consent.

Commission a custom collection once a program needs a specific task or environment none of the three corpora covers, a commercial license a paper or dataset card does not clearly state, or synchronized depth, IMU and 3D hand-pose streams none of them documents. See /custom-collection/pretraining-data-collection for how a written spec becomes a delivered batch, typically starting with a fast, on-demand Spec Pilot before scaling to full volume.

Most programs do not treat this as either-or: EgoScale, EgoVerse or Egocentric-10K is often used for early experimentation and broad pretraining, then a custom batch fills the specific task, environment or license gap the public corpus leaves. See /compare/public-vs-custom-pretraining-datasets for the fuller version of that decision, and /guides/robot-pretraining-data for how pretraining data fits alongside mid-training and post-training in a full pipeline.