EgoVerse vs GEN-0: Human Data vs Robot Data at Scale

EgoVerse is described in arXiv 2604.07607, “EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World”. The abstract calls it “a collaborative platform for human data–driven robot learning” that enables “contributions from individual researchers, academic labs, and industry partners”. The current release has “1,362 hours (80k episodes) of human demonstrations spanning 1,965 tasks, 240 scenes, and 2,087 unique demonstrators”.

It has two parts. EgoVerse-A follows standardised protocols mirrored across academic labs. EgoVerse-I is sourced from industry partners “collecting data in the wild”. The paper pairs the dataset with a transfer study replicated across labs and three robot platforms, and starts from a stated premise: “robot data collection remains expensive and difficult to scale”. See /guides/what-is-egoverse for the single-paper explainer.

GEN-0 is announced in a Generalist AI blog post, “GEN-0 / Embodied Foundation Models That Scale with Physical Interaction”, cited as November 2025. The post calls it “a new class of embodied foundation models built for multimodal training directly on high-fidelity raw physical interaction”, pretrained on “our in-house robotics dataset, which includes over 270,000 hours of real-world diverse manipulation data, growing at a rate of 10,000 hours a week”.

It describes the data as manipulation trajectories “collected across diverse activities in 1,000s of homes, warehouses, and workplaces worldwide”, using “1,000s of data collection devices and robots”. The post does not say how much comes from robots rather than other devices, or whether any is human video. It is a company blog post, not a peer-reviewed paper. See /guides/what-is-gen-0.

They differ in cost, in what they record, and in how close they sit to the robot that will use them. EgoVerse states the cost case directly: collecting robot demonstrations “requires physical hardware, expert teleoperation, and controlled setups”, while humans “naturally perform manipulation and loco-manipulation tasks across diverse environments on a daily basis”. Human data also lets researchers focus on diverse experience “while deferring embodiment decisions downstream”.

The price is the embodiment gap. The paper says effective human–robot transfer “remains an open research problem, with unresolved questions around the embodiment gap and scaling behavior”. Data recorded on the target robot has no such gap, but it needs hardware and operators; the GEN-0 post describes custom hardware and infrastructure built for its scale. For the general argument see /compare/human-video-vs-robot-data-for-pretraining.

The first table sets them side by side. Scale is the widest gap: GEN-0 states over 270,000 hours against EgoVerse’s 1,362. Treat the two as different units. EgoVerse counts hours of human demonstrations with per-frame pose annotations. GEN-0 counts hours of manipulation data from its own operations and gives no annotation schema.

Capture is the other contrast. EgoVerse documents hardware for every source: Aria glasses at 75 g, partner head-mounted rigs, and an iPhone on a head strap. GEN-0 states only that it built “custom hardware” for scale; its capture rigs are not described. On annotations, EgoVerse estimates 3D hand pose and head pose per frame, which “serve as proxies for human end-effector motion”. For other pairings see /compare/egoscale-vs-gen-0 and /compare/gen-0-vs-humanscale.

EgoVerse reports that “policy performance generally improves with increased human data, but that effective scaling depends on alignment between human data and robot learning objectives”. A small aligned set anchored the benefit: 2 hours of domain-aligned data enabled transfer from 2 hours of diverse data, a trend that continued to 8 hours. On diversity, it notes that beyond a certain scale “environmental diversity plays a more critical role in generalization than additional data collected within individual scenes”.

GEN-0 states that “more pretraining data and compute consistently (and predictably) improve downstream post-training performance”, with a power-law relationship between pretraining data size and downstream error across 16 task sets. Models post-trained on only 5.6 hours of task-specific data improved with more pretraining, and the best models reached “up to 99% peak performance in certain cases”. Its Science of Pretraining section adds that “data quality and diversity matters more than sheer volume”.

The two agree on diversity, but neither is evidence for the other. EgoVerse’s scaling experiments use 2 and 8 hours of human data, and its diversity results use an offline metric. GEN-0 reports pretraining at far larger scale. EgoVerse’s reference list includes the GEN-0 post, cited in its diversity section; the earlier GEN-0 post does not mention EgoVerse.

Neither source gives a pretraining mixture ratio. EgoVerse’s experiments held a fixed 1:1 human-to-robot batch ratio, which is a setting of that study rather than a finding, and the paper lists pre-train and fine-tune as future work. The GEN-0 post does not describe human video in its dataset. What the sources can support is a set of conditions.

Lean on human egocentric data when robot time is the bottleneck, when you want data not tied to one embodiment, or when you need breadth of scenes and demonstrators quickly. EgoVerse’s results suggest budgeting for scene diversity before adding hours in the same scenes. Lean on robot data when you already operate robots at volume, when your target is one embodiment and its action space, or when you judge by closed-loop success on that robot. In both cases EgoVerse’s alignment finding applies: data matched to your tasks and scenes separated flat scaling from positive scaling.

This is the page’s reading of two sources, not a recommendation either author makes. See /guides/building-a-robot-pretraining-data-mixture for how to build a mixture.

The EgoVerse authors say the study “mainly focused on human-and-robot co-training” and that broader exploration, such as pre-train and fine-tune, is future work. The diversity experiments rely on offline metrics, so robot rollouts are needed to confirm them. In one case, bag-grocery on Robot B, adding human data lowered performance, which the authors hypothesise stems from a strategy mismatch between human and robot demonstrations. Licence and public release terms are not stated in the text reviewed.

The GEN-0 post is an announcement, not a methods paper, and defers further findings to “future posts”. It states no licence or public release, does not break its hours down by robot, device or human source, and gives no annotation schema. It describes no way for outside teams to obtain the corpus.

Custom collection is not a substitute for either source. It covers what a fixed public corpus and a closed in-house one cannot: tasks, scenes and failure cases specific to your own program. EgoVerse finds that aligned data anchors scaling, and GEN-0 says that having “multiple data collection strategies at scale allows us to continually A/B test which data improves pretraining the most”. Both point the same way: treat each data source as something to test.

Firsthand collects custom egocentric video to a buyer’s skill spec, scoped to the buyer’s own target tasks and environments. Many programs start with a Spec Pilot, a small, fast first batch that proves coverage and label quality before scaling volume, with a median 48-hour first validated batch where coverage is deep. Data includes frame-accurate action segments and manipulation-grade 3D hand pose captured on the rig, delivered in RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd.

The default licence is buyer-owned, perpetual and non-exclusive, confirmed in writing before capture. See /custom-collection/pretraining-data-collection and /custom-collection/egocentric-data-collection-pilot. Firsthand has no relationship with Generalist or the EgoVerse authors.

The sources suggest six questions. First, the licence in writing, since neither source states one. Second, the capture hardware and calibration, which EgoVerse documents for every source and GEN-0 does not. Third, the per-frame annotations and how hand and head pose are produced. Fourth, scene and demonstrator diversity rather than hours alone, since both sources point that way. Fifth, alignment to your target tasks and robot. Sixth, a small pilot with a pass bar before volume; see /guides/evaluating-a-robot-pretraining-data-pilot.