Egocentric-10K vs GEN-0 Training Data

Egocentric-10K is a Hugging Face dataset published by Build AI. The card opens: “Egocentric-10K is the largest egocentric dataset. It is the first dataset collected exclusively in real factories.” Its statistics table lists 10,000 total hours, 1.08 billion frames and 192,900 video clips, with a 180-second median clip length and 4.68 mean hours per worker. The card also calls it “state-of-the-art in hand visibility and active manipulation density compared to previous in-the-wild egocentric datasets”; the text reviewed gives no benchmark behind that.

Footage is H.265/MP4 at 1080p and 30 fps, from a monocular head-mounted camera with no audio. It is organised in WebDataset format by factory and worker, with calibrated fisheye intrinsics per worker and a JSON metadata file per clip. The licence line reads “Licensed under the Apache 2.0 License.” Access is gated: “You need to agree to share your contact information to access this dataset.” See /guides/what-is-egocentric-10k for the single-source explainer.

GEN-0 is announced in a Generalist AI blog post cited as November 2025. It calls GEN-0 “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”.

The post describes the data as manipulation trajectories collected in homes, warehouses and workplaces, using “1,000s of data collection devices and robots”. It does not give a split between robots and other devices, and does not say whether any of it is human video. It is a company blog post, not a peer-reviewed paper. See /guides/what-is-gen-0.

They record different things. Egocentric-10K records what a worker sees through a head-mounted camera while doing factory work. The card gives video, camera intrinsics and clip metadata, and states no hand-pose, action or language labels for the main set. GEN-0 trains on manipulation trajectories and measures “next-action prediction error”, so its pretraining data carries robot actions in a way a head-camera video does not.

They also differ in who can use them. Video from a person has no embodiment gap to close if you only want visual and task-structure data, but turning it into a policy input means producing action or pose signals yourself. Robot data is matched to a robot, but needs hardware and operators; the GEN-0 post describes custom hardware, dataloaders and network 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 Egocentric-10K’s 10,000. Treat the units as different. Egocentric-10K counts hours of head-camera footage, with a stated size of 16.4 TB. GEN-0 counts hours of manipulation data and mentions compressing dozens of petabytes, which is not a like-for-like size.

Capture and format are documented for Egocentric-10K down to field of view, frame rate and codec; GEN-0 states only that it built custom hardware. Licence is the sharpest contrast. Egocentric-10K states Apache 2.0 on the card. GEN-0 states no licence, and the post describes no way for outside teams to obtain its dataset. For other pairings see /compare/egoverse-vs-gen-0, /compare/egoscale-vs-gen-0 and /compare/gen-0-vs-humanscale.

The Egocentric-10K card states no scaling experiment and no intended use in the text reviewed. Its claims are about the dataset itself: size, factory setting, and hand visibility. It lists a section for models trained or fine-tuned on it, but the text reviewed names no result. Any scaling behaviour with this dataset would have to be measured by the team using it. See /compare/egocentric-10k-vs-gr00t-n1 for how it sits beside a published training recipe.

GEN-0 is the opposite: the post is about scaling. It 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” when paired with all 550+ hours of post-training data.

The post also warns against reading scale alone: “data quality and diversity matters more than sheer volume”, and “different mixtures of pretraining data” give models different characteristics. That is a statement about GEN-0’s own robot data, not a finding about human video, and it should not be applied to Egocentric-10K as evidence.

Neither source recommends a mixture. What the sources can support is a set of conditions, and this is the page’s reading of them, not a recommendation either publisher makes.

Lean on an open egocentric dataset when you need a licence you can read today, when your targets are factory or industrial manipulation, and when you can supply the labelling or pose estimation the card does not state. It suits early experiments and visual pretraining where you want to start without collecting anything. Check what is not stated first: consent and release terms, and the labels, are not described in the card text reviewed.

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 progress by closed-loop success on that robot. GEN-0’s results come from that setting, and the post describes no outside access to the corpus. In practice a team does not choose one: it tests each source as one component of a mixture and measures the effect on its own tasks. See /guides/building-a-robot-pretraining-data-mixture.

For Egocentric-10K, the card text reviewed does not state the consent or release basis for the footage, any hand-pose or action annotation, an intended use, or a training result. The gate asks for contact information, and the card does not describe what is done with it. One setting, real factories, is a strength for factory tasks and a limit for anything else.

For GEN-0, the post is an announcement, not a methods paper, and defers further findings to “more on these learnings in future posts”. It states no licence or release, does not break its hours down by robot, device or human source, and gives no annotation schema.

Custom collection is not a substitute for either source. It covers what a fixed public dataset and an in-house corpus cannot: tasks, scenes and failure cases specific to your own program, with a licence you hold in writing. GEN-0 says that having “multiple data collection strategies at scale allows us to continually A/B test which data improves pretraining the most”. The same logic applies to a buyer: 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, /custom-collection/egocentric-data-collection-pilot and /provenance. Firsthand has no relationship with Generalist or the Egocentric-10K publisher.

The sources suggest six questions. First, the licence in writing, since GEN-0 states none and a card’s licence line does not settle consent. Second, the capture hardware and calibration, which Egocentric-10K documents and GEN-0 does not. Third, the annotations, and how any hand or action signal is produced. Fourth, scene and site diversity rather than hours alone, since GEN-0 says quality and diversity matter more than volume. 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. Review any licence with counsel; this page is not legal advice.