Egocentric-10K vs HumanScale: Egocentric Data for Robot Pretraining

Egocentric-10K is a dataset hosted on Hugging Face by builddotai at huggingface.co/datasets/builddotai/Egocentric-10K. Its card states: “Egocentric-10K is the largest egocentric dataset. It is the first dataset collected exclusively in real factories.” The statistics table lists 10,000 total hours, 1.08 billion frames, 192,900 video clips, a median clip length of 180.0 seconds and 16.4 TB of storage.

Each clip comes from one device, which the card calls Build AI Gen 1: a monocular, head-mounted camera recording H.265/MP4 at 1080p and 30 fps, with a 128° horizontal and 67° vertical field of view and no audio. See /guides/what-is-egocentric-10k for the single-dataset explainer.

HumanScale is a research paper at arxiv.org/abs/2606.20521, titled “HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining”. The paper states it conducts “a systematic study comparing egocentric human video and teleoperated real-robot trajectories as pretraining data sources for embodied foundation models, under fixed post-training and validation protocols”.

To isolate the pretraining source, the authors curate a 5,000-hour subset from the egocentric portion of HumanNet and compare it with an equal-scale multi-embodiment real-robot corpus. Both pretrain the same autoregressive world-action model. See /guides/what-is-humanscale for the single-paper explainer.

No. Egocentric-10K is an artefact you can load into a training run. HumanScale is evidence about a choice you make before that run: whether egocentric human video is a good pretraining source compared with teleoperated robot data.

The sources do not link the two. The HumanScale text we reviewed does not name Egocentric-10K; its dataset table lists a separate Build AI release, Egocentric-100K, at 100,405 hours. Nothing on the Egocentric-10K card refers to HumanScale.

Egocentric-10K’s capture setup is fully specified: one standardised monocular head-mounted device for every worker, recorded in real factories. Because every clip shares one camera, resolution, frame rate and field of view are constant, and each per-worker intrinsics.json file records calibrated fisheye parameters (an OpenCV fisheye model with four distortion coefficients) at 1920×1080.

HumanScale does not describe a capture rig for its subset. It draws 5,000 hours from the egocentric portion of HumanNet, described as one million hours of human activity with over 800,000 hours of egocentric footage. For a buyer, HumanScale shows what kind of data worked at scale, not which hardware to specify.

On raw hours, Egocentric-10K’s 10,000 hours is twice the 5,000-hour subset HumanScale pretrains on. But the numbers measure different things: one is a corpus you can download, the other is the scale of a controlled experiment that also tests 100 and 1,000 hours.

On diversity the picture reverses. Egocentric-10K is single-domain by design: its card states collection exclusively in real factories. HumanScale selects its subset for diversity and reports that egocentric video shows broader workspace, interaction and scene coverage than real-robot data at matched duration.

Egocentric-10K ships video plus metadata. Each clip has a JSON file with factory_id, worker_id, video_index, duration_sec, width, height, fps, size_bytes and codec, and each worker folder has calibrated intrinsics. The card describes no action, hand-pose or language labels, so a team using it for policy learning would add its own labelling.

HumanScale processes its subset through what the abstract calls “a carefully designed filtering and labeling pipeline”, which the paper describes as filtering, deduplication, privacy review and pseudo-action labelling via hand-pose retargeting. The labels are per-clip end-effector poses and gripper states estimated from retargeted hand-pose signals, placing the human data in the same action space as the robot data.

HumanScale uses human video for pretraining and a small amount of real-robot data afterwards. Its conclusion is to “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”. Concretely, it post-trains each pretrained model on 1,500 AgiBot World trajectories covering 15 tasks, then evaluates on held-out trajectories from those tasks (Seen) and on 25 tasks excluded from post-training (Unseen).

On Seen tasks the paper finds the two sources comparably effective. On Unseen tasks, scaling robot pretraining produced no consistent improvement, while egocentric pretraining kept improving. The abstract reports 24% lower validation loss on real-robot action prediction and 52.5% and 90% higher success rates on in-distribution and out-of-distribution execution. Those percentages equal the point gaps in Table 2 against the Wan2.2 no-pretraining baseline (92.5 vs 40.0, and 90.0 vs 0.0), so check which comparator a quoted figure refers to.

The paper states its limits: one world-action model built on Wan 2.2, training capped at 5,000 hours because of limited robot data, and a vision-language-action study left to future work. Egocentric-10K reports no transfer experiment.

Egocentric-10K’s card states: “Licensed under the Apache 2.0 License.” Access runs through Hugging Face, where the card notes you must agree to share your contact information to download. The licence text itself governs what that allows.

HumanScale states no licence for its data in the text reviewed. The paper says code will be released at a GitHub repository it names, but it states no licence or release terms for the 5,000-hour subset, and the subset’s availability is not stated. Ask the authors directly and get commercial-use terms in writing.

Egocentric-10K leaves open domain breadth, annotation depth and transfer evidence. It covers factory work only, ships no action or hand-pose labels, and reports no experiment showing how models pretrained on it perform on a robot. It is a large, openly licensed source of factory footage, and that is what it claims to be.

HumanScale leaves open reproducibility, licence clarity and generality. The paper does not state how a reader can obtain the exact subset, its evidence comes from one architecture and one post-training set, and its labels are estimates rather than measured ground truth.

Read together, they suggest a sequence rather than a choice. HumanScale offers a published argument that diverse egocentric video is a strong pretraining source and that a small amount of aligned robot data should follow. Egocentric-10K offers a concrete, permissively licensed pool of one kind of egocentric data. Use the first to justify a strategy and the second as one input.

Custom collection is not a substitute for either source. It fills coverage that a published dataset’s scope and a study’s fixed subset cannot. A program might use Egocentric-10K as one input, take HumanScale’s finding as a reason to prioritise diversity, then commission a fast, on-demand pilot batch scoped to the scenes, tasks and failure cases its own model is missing.

Firsthand collects custom egocentric video to a buyer’s skill spec, tests the spec on a small Spec Pilot batch before scaling, and delivers 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 /guides/robot-pretraining-data for how pretraining and post-training divide the work, and /custom-collection/pretraining-data-collection for the engagement model.