EgoScale vs HumanScale: Human Video for Robot Pretraining

EgoScale and HumanScale are both published research papers studying the same underlying question from independent labs: does pretraining a robot policy on large-scale egocentric human video improve downstream manipulation performance, and how does that improvement scale with data volume? EgoScale comes from NVIDIA’s GEAR lab (arXiv 2602.16710), with an accompanying official research page. HumanScale is a separate publication (arXiv 2606.20521) studying the same broad research question.

Neither is a Firsthand product, and neither is a commercial dataset release. Both are treated on this site as public reference points for an argument this site makes throughout its pretraining guides: human-video scale helps robot pretraining, and that pattern is now supported by more than one independent research line rather than a single lab’s isolated result.

The clearest difference is how much each paper is documented on this site. EgoScale is the source this site cites in the most technical depth — a specific corpus size, a documented label-production method, a two-stage pretraining-then-mid-training recipe, an evaluation platform and a reported benchmark result. HumanScale is referenced consistently at the level of its research question and its role alongside EgoScale, because this site does not extract or restate HumanScale’s specific hour counts, architecture details or benchmark figures.

That asymmetry is a sourcing choice, not a claim that HumanScale’s research is less substantial. A reader who needs HumanScale’s full technical detail should consult arxiv.org/abs/2606.20521 directly. What this page can state with confidence is the structural fact: two independently published papers are studying the same scaling question, which is itself evidence the question is an active area of robotics research.

EgoScale pretrains on 20,854 hours of action-labeled egocentric human video, then runs a much smaller aligned mid-training stage using roughly 50 hours of human video plus about 4 hours of robot data across 344 tabletop tasks. That split — a large human-video pretraining corpus followed by a small, closely matched robot-aligned stage — is documented in detail at /compare/egoscale.

HumanScale studies human-video pretraining for robot policies, the same broad category of data as EgoScale’s corpus, EgoVerse’s roughly 1,362-hour dataset, and Egocentric-10K’s roughly 10,000-hour dataset. This site does not state a specific hour count, capture setup or label format for HumanScale’s own corpus, because that specific figure is not stated elsewhere on this site. Where this comparison needs concrete numbers, it uses EgoScale’s published figures and notes explicitly where HumanScale’s own numbers are not asserted.

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. That approach scales to tens of thousands of hours but is estimated rather than measured directly on a hardware-synchronized capture rig — a tradeoff this site documents in detail at /compare/egoscale.

HumanScale’s label or action-representation methodology is not documented in the material this site cites, so this page states that as not stated rather than inferring a method. A team that needs to know exactly how HumanScale represents actions should read the paper at arxiv.org/abs/2606.20521 directly.

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 at all.

HumanScale’s reported contribution, as documented on this site, is framed at the level of its research question and its role alongside EgoScale: a separate published research line specifically studying the scaling behavior of human-video pretraining for robot policies. That two independently published papers, from what this site describes as different labs and different evaluation setups, converge on the broader conclusion that human-video scale matters is itself part of the evidence — it suggests the pattern is not an artifact of one paper’s specific corpus or benchmark choice.

EgoScale’s documented limits: its action labels come from post-hoc estimation rather than a synchronized rig, so label precision is bounded by what SLAM and hand-pose estimation can recover under occlusion, fast motion or unusual lighting. Its scaling law describes what happened with its own corpus and evaluation setup — evidence that scale helped in that specific study, not a guarantee every corpus scales identically regardless of diversity or label quality. And commercial usage terms for any EgoScale-related release should be confirmed directly against its published license before relying on it for a commercial model.

HumanScale’s limit, as far as this site can state it, is different in kind: because this site does not extract its specific dataset scale, capture methodology or benchmark results, it cannot state how HumanScale’s exact numbers compare with EgoScale’s, or whether its corpus carries action labels of comparable precision. More broadly, the same structural limitation applies to both: a published research dataset is captured for its own study’s purpose, with its own scene, task and demographic distribution, and is not reshaped after the fact to a specific buyer’s target task set, embodiment or coverage gaps.

Together, EgoScale and HumanScale are evidence that planning a pretraining pipeline around human-video scale — rather than assuming robot-only teleoperated data is the only credible path to a working policy — is a pattern supported by more than one independent research line. Neither paper substitutes for a buyer’s own coverage plan: EgoScale’s 20,854-hour figure and its reported 54% result describe what happened with its own evaluation setup, and HumanScale’s specific figures are not published on this site at all.

The practical takeaway this site repeats throughout its pretraining guides applies here too: 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. A provider’s ability to show how its collection design maps to that spec — not a quoted hour count alone — is what distinguishes a pretraining order that fills a real coverage gap from one that simply adds volume.

A custom pretraining collection starts with a written skill spec naming the scenes, tasks, objects and people a corpus is missing — informed by what both EgoScale and HumanScale point toward (human-video scale helps pretraining) without copying either paper’s own scene mix, task distribution or label format, which were built for that paper’s own study rather than a specific buyer’s model. The spec is quoted before capture begins, often starting with a Spec Pilot, a small first batch scored against the spec before scaling to full volume.

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 — whether a program is pretraining broadly using EgoScale and HumanScale as public reference points for scaling behavior, or aligning a broad prior to one target robot with a small mid-training batch, the same structure EgoScale’s own recipe uses. See /custom-collection/pretraining-data-collection for the full engagement model and /guides/robot-pretraining-data for how pretraining, mid-training and post-training divide that work.