EgoVerse vs HumanScale: Human Data for Robot Learning

EgoVerse is described in “EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World” (arXiv 2604.07607). Its abstract states that the current release includes 1,362 hours (80k episodes) of human demonstrations spanning 1,965 tasks, 240 scenes and 2,087 unique demonstrators, with standardized formats, manipulation-relevant annotations and tooling for downstream learning.

The paper pairs that dataset with a study: “experiments replicated across multiple labs, tasks, and robot embodiments under shared protocols.” It has two parts. EgoVerse-A is collected under standardized protocols across academic labs; EgoVerse-I is sourced from industry partners collecting data in the wild. See /guides/what-is-egoverse for the single-dataset explainer.

HumanScale is described in “HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining” (arXiv 2606.20521). It asks how egocentric human video compares with teleoperated real-robot data as a pretraining source, under fixed post-training and validation protocols.

Its abstract states the conclusion as a paradigm: “pretrain on egocentric human video to learn diverse world representations, then adapt with a small amount of labeled real-robot data for action-space alignment.” The paper says its results are encouraging but preliminary. See /guides/what-is-humanscale for the single-paper explainer.

EgoVerse documents its own capture. EgoVerse-A uses Project Aria glasses (Gen 1) as the standard device; industry partners contribute EgoVerse-I with custom-built rigs; and a phone-based system uses an iPhone on a head strap recording 1080p at 30 FPS. Each frame carries 3D hand pose with 21 keypoints per hand and a 6-DoF head pose. EgoVerse-I adds fine-grained (1–2 s) language descriptions, active-hand indicators and static-versus-mobile manipulation flags.

HumanScale does not describe a capture rig of its own. It curates a 5,000-hour subset from the egocentric portion of HumanNet, selected for diversity in objects, skills and environments. Because that footage has no robot actions, the paper estimates end-effector poses and gripper states from retargeted hand-pose signals and uses them as pseudo-action labels. The first source is a purpose-built human-demonstration collection; the second is a curated slice of a larger pool.

The headline scales differ in kind, not just size. EgoVerse reports 1,362 hours as a dataset release. HumanScale reports 5,000 hours as the size of the subset it trains on, and says the pool it draws from, HumanNet, curates one million hours of human activity, over 800,000 of them egocentric. The two hour counts measure different things, so they are not a like-for-like ranking.

On diversity, EgoVerse isolates it experimentally. Its controlled subset uses 16 demonstrators and 16 scenes, and it reports that scene diversity improves generalization to unseen scenes, with the strongest gains under limited data, while the marginal benefit of additional demonstrators decreases as scene coverage grows. HumanScale instead attributes its scaling behavior to curated diversity, noting that its subsets are curated for broad coverage rather than testing diversity axes one at a time.

EgoVerse studies co-training. A behavior-cloning loss is computed on the aggregated human and robot dataset, with per-step losses on mini-batches of both. Robot data is roughly 150–300 demonstrations per task. Its extended discussion names pretraining on heterogeneous human data followed by fine-tuning as a direction for future work.

HumanScale studies pretraining followed by post-training. It pretrains the same world-action model on 5,000 hours of egocentric video or 5,000 hours of multi-embodiment robot data, then post-trains both on the same 1,500 AgiBot World trajectories. Robot data appears at two points: as the comparison arm in pretraining, and as the post-training set that aligns the action space. See /glossary/co-training for the definition of the first recipe.

EgoVerse reports that co-training with EgoVerse-A improves performance in both in-domain and out-of-domain settings by up to 30%. It also reports a performance decrease on the bag-grocery task for one robot, which it attributes to that robot’s embodiment limits. Its scaling finding is conditional: neither 8 hours of diverse EgoVerse-A data nor domain-aligned human data alone drove significant gains, but including just 2 hours of domain-aligned data facilitated transfer from 2 hours of diverse data.

HumanScale reports that validation loss falls monotonically as egocentric pretraining grows from 100 to 5,000 hours, fitted by a log-linear law. On seen tasks, robot pretraining stays close to egocentric pretraining; on unseen tasks, robot-pretraining loss stays near 0.025 while egocentric pretraining reaches 0.0204, roughly 20% lower. In real-robot rollouts the egocentric-pretrained model reaches 92.5% in-distribution and 90.0% out-of-distribution success, against 40.0% and 0.0% for the no-pretraining baseline.

EgoVerse states that its study mainly covered co-training and that its scene and demonstrator experiments use offline metrics, so additional robot rollouts are needed to confirm those diversity effects on a robot. It leaves pretrain-then-fine-tune as future work.

HumanScale states that its training is limited to 5,000 hours by the availability of real-robot data, that its evaluations use a world-action model on a Wan 2.2 backbone, and that it is still evaluating vision-language-action models across more embodiments. Neither paper states a data license in the text cited here, and neither reports a single human-to-robot hour ratio a team could copy.

Read together, the papers point at the same planning question from two sides. EgoVerse suggests a team should plan a small aligned human set alongside its robot demonstrations, because that set is what let diverse human data help. HumanScale suggests the diversity of the pretraining human video matters, and that robot data is the aligning step rather than the bulk. Neither paper turns this into a recommended ratio, and this page does not either.

For how a team sizes the human share, see /guides/human-vs-robot-data-ratio-for-pretraining, and for the comparison of the two data types see /compare/human-video-vs-robot-data-for-pretraining.

Custom collection is not a substitute for either study. It is how a team gets human egocentric data scoped to its own tasks, scenes and robot, which neither a fixed dataset nor a curated web slice is built to match. A team might use EgoVerse as a reference for what aligned and diverse human data looks like, use HumanScale as a reference for pretraining on human video, and then commission a fast, on-demand pilot batch to test its own spec before scaling.

Firsthand delivers custom batches in RLDS, LeRobot, WebDataset, HDF5, zarr and Rerun .rrd, with a buyer-owned commercial license by default. See /custom-collection/pretraining-data-collection for the engagement model and /guides/robot-pretraining-data for how pretraining, mid-training and post-training divide the work. Firsthand has no relationship with the authors of either paper.