Egocentric-10K vs EgoScale for Robot Pretraining

Egocentric-10K is a published dataset hosted on Hugging Face by builddotai at huggingface.co/datasets/builddotai/Egocentric-10K. Its dataset card states roughly 10,000 total hours of egocentric video, 1.08 billion frames and 192,900 video clips with a median clip length of 180 seconds, collected exclusively inside real factories across 85 factory sites.

Each clip is captured with a single device the card calls Build AI Gen 1 — a monocular, head-mounted camera recording 1080p video at 30 fps with a 128° horizontal by 67° vertical field of view and no audio track. The card reports a mean of 4.68 hours of footage per worker, structured as WebDataset-format TAR shards, each paired with a per-worker intrinsics.json file carrying calibrated fisheye camera parameters. The card states the dataset is licensed under Apache 2.0. See /guides/what-is-egocentric-10k for the full single-dataset explainer.

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. 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.

Egocentric-10K is primarily a dataset: a published collection of roughly 10,000 hours of factory-floor footage with per-worker camera intrinsics, intended to be used directly by other researchers. 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.

That distinction matters for how each is used. Egocentric-10K’s value to a buyer is mainly as a large, permissively licensed source of factory-manipulation footage. EgoScale’s value is mainly as a citable reference point for how egocentric-video scale affects pretraining loss and downstream task success.

Egocentric-10K standardizes on one device — Build AI Gen 1, a monocular head-mounted camera — for every worker, which keeps resolution, frame rate and field of view identical across the whole corpus. Its card documents calibrated per-worker camera intrinsics but no action, hand-pose or language labels.

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.

On raw hours, EgoScale is the larger corpus at 20,854 hours against Egocentric-10K’s roughly 10,000 hours — about twice as much footage. Egocentric-10K’s card does not report a scaling-law study; it is a description of scale and capture, not a pretraining-loss curve.

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 that Egocentric-10K’s card does not attempt to replicate.

Egocentric-10K’s Hugging Face dataset card states an Apache 2.0 license — a widely used permissive license that allows use, modification and redistribution, including for commercial purposes, subject to the license’s own terms.

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. For an early-stage robotics team, that 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.

Egocentric-10K’s gap is domain breadth and annotation depth. Its card states the footage is collected exclusively inside real factories, so a team pretraining for a household, retail, office or outdoor task has no coverage of that domain in this corpus, and no action or hand-pose labels ship with it. EgoScale’s gap is license certainty and capture precision: its release terms are not a published open license on this site, and its action labels come from post-hoc estimation rather than a purpose-built synchronized rig.

Both sources share one structural limitation for robot learning specifically: neither is robot execution data. EgoScale’s own mid-training stage exists precisely because broad human pretraining alone does not align to a target embodiment — a finding that applies equally to Egocentric-10K’s factory footage.

Custom collection is not a competitor to Egocentric-10K or EgoScale — it is how a program fills the specific coverage gap that neither a published dataset’s own scope nor a research recipe’s fixed corpus can close for a particular buyer’s tasks and environments. A program might pretrain broadly on Egocentric-10K’s factory-floor footage, treat EgoScale’s scaling-law result as a benchmark for how scale affects loss, 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.