EgoScale vs GEN-0: Two Approaches to Robot Pretraining Data

EgoScale and GEN-0 are different kinds of sources making different kinds of claims. EgoScale is a peer-reviewed-style research paper from NVIDIA’s GEAR lab (arXiv 2602.16710) that documents a measured log-linear scaling law between egocentric human-video pretraining scale and validation loss, plus a reported downstream result on a real robot. GEN-0 is a company blog announcement from Generalist AI, published November 4, 2025, stating that its robot foundation model trained on more than 270,000 hours of data, with an explicit emphasis on data quality and diversity over raw volume.

Neither is a Firsthand product. Both are treated on this site as public reference points for the same broader argument this site makes throughout its pretraining guides: composition and diversity shape pretraining outcomes alongside — not instead of — scale.

EgoScale pretrains on 20,854 hours of action-labeled egocentric human video, with action labels produced from off-the-shelf SLAM and hand-pose estimation applied after capture rather than measured on a purpose-built synchronized rig. See /compare/egoscale for the full technical breakdown.

GEN-0’s announcement states its model trained on more than 270,000 hours of data. Generalist AI’s official write-up does not break down the source composition of those hours in the material cited on this site, so this page does not attribute a specific data type, capture method or label source to GEN-0 beyond what the announcement itself states. See /guides/what-is-gen-0 for the full explainer.

EgoScale: 20,854 hours of egocentric human video for pretraining, plus a much smaller mid-training stage of roughly 50 hours of human video and about 4 hours of robot data across 344 tabletop tasks.

GEN-0: more than 270,000 hours, as stated in Generalist AI’s official blog post. That figure is more than an order of magnitude larger than EgoScale’s pretraining corpus, but the two numbers are not measuring the same thing in the same level of published detail — EgoScale’s figure comes with a documented scaling-law methodology, while GEN-0’s comes from a company announcement without the same published depth.

EgoScale’s action labels are produced from off-the-shelf SLAM (for camera trajectory and scene geometry) and hand-pose estimation applied post-hoc to the raw egocentric video — an approach that scales to tens of thousands of hours but is estimated rather than measured directly on a synchronized capture rig.

GEN-0’s announcement does not document its labeling or action-representation approach in the material cited on this site, so this page states that as not stated rather than inferring a method.

EgoScale makes a quantitative claim: a log-linear scaling law between egocentric human-video pretraining scale and validation loss, which the paper reports correlating with real downstream robot task performance — measured as a 54% average success-rate improvement on a 22-degree-of-freedom dexterous hand versus no pretraining.

GEN-0 makes a qualitative claim: Generalist AI states directly that data quality and diversity matter more than sheer volume while scaling to 270,000+ hours. That is a stated priority, not a published scaling curve or a benchmark score, and this page does not attribute a quantitative result to GEN-0 beyond what the announcement itself says.

EgoScale’s recipe is specific on this point: a large human-video pretraining stage, followed by a small aligned mid-training stage using real robot data (about 4 hours across 344 tasks) to close the embodiment gap. See /compare/human-video-vs-robot-data-for-pretraining for the full comparison of that trade-off.

GEN-0’s announcement does not specify a human-video-versus-robot-data split for its 270,000+ hours in the material cited on this site, so this page does not attribute that composition to GEN-0. What both sources agree on, in their own framing, is that composition and diversity matter alongside scale — see /guides/why-dataset-diversity-matters-in-robot-learning.

A program that wants a citable, quantitative reference point for how egocentric-video scale affects pretraining loss and downstream success has more to work with in EgoScale’s published methodology and evaluation. A program that wants evidence that a very large model has been trained with data quality and diversity as an explicit priority has GEN-0’s public statement, though without the same level of published detail.

Neither source 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.

Treat both EgoScale’s hour figure and GEN-0’s hour figure as reference points for the range of published and announced pretraining scale, not as a target to match. The right size for a specific pretraining, mid-training or fine-tuning order depends on the target model’s architecture and task distribution, and how much of that distribution is already covered — not on matching the most recent public number.

A custom pretraining collection starts with a written skill spec covering the scenes, tasks, objects and people a corpus is missing, quoted before capture begins — often starting with a Spec Pilot, a small first batch that proves the spec before scaling volume. See /custom-collection/pretraining-data-collection for the full engagement model.

Custom collection is not a competitor to EgoScale or GEN-0 — it is how a program fills the specific coverage gap neither a research paper’s fixed corpus nor a company’s announced model can close for a particular buyer’s tasks and environments. A program might pretrain broadly using public reference points like EgoScale as a benchmark for scaling behavior, note GEN-0’s public framing on quality and 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 collected or referenced at any pipeline stage. See /guides/robot-pretraining-data for how pretraining, mid-training and post-training divide that work.