Egocentric-10K vs GR00T N1 Training Data

Egocentric-10K is a dataset published by Build AI on Hugging Face. Its card opens: “Egocentric-10K is the largest egocentric dataset. It is the first dataset collected exclusively in real factories.” It reports 10,000 hours, 1.08 billion frames and 192,900 video clips, filmed with a monocular head-mounted camera at 1080p and 30 fps.

The card also claims it is “state-of-the-art in hand visibility and active manipulation density compared to previous in-the-wild egocentric datasets”. That is the publisher’s own description. See /guides/what-is-egocentric-10k for the single-source explainer.

GR00T N1 is NVIDIA’s “open foundation model for humanoid robots”, described in arXiv 2503.14734 and released through the NVIDIA/Isaac-GR00T repository. It is a vision-language-action model: a vision-language module interprets the scene and instruction, and a diffusion transformer generates motor actions. The abstract says it is trained “with a heterogeneous mixture of real-robot trajectories, human videos, and synthetically generated datasets”.

The paper organises that mixture as a data pyramid: web data and human video at the base, synthetic data in the middle, real-robot data at the top. See /guides/gr00t-n1-data-pyramid-explained and /glossary/data-pyramid for the pyramid in detail. The repository README now describes GR00T N1.7, which it says adds 20K hours of EgoScale human video to pretraining; this page uses the original paper for N1 and labels README material as N1.7.

No. Egocentric-10K is raw material: footage plus metadata. GR00T N1 is a trained model, and what its paper offers a data buyer is a description of how mixed data, including human video, is used. They answer different questions: what a large factory-video corpus contains, versus how a foundation model blends human video with other layers.

Neither source names the other. The Egocentric-10K card text reviewed does not mention GR00T N1, and the GR00T N1 paper lists seven human-video datasets that do not include Egocentric-10K. For other pairings see /compare/egocentric-10k-vs-egoscale, /compare/egocentric-10k-vs-egoverse, /compare/egocentric-10k-vs-humanscale and /compare/egoverse-vs-gr00t-n1.

Egocentric-10K is its own capture. The card gives the device as Build AI Gen 1, a monocular head-mounted camera, with a 128° by 67° field of view and no audio. Footage is organised by factory and worker: the folder tree runs from factory_001 to factory_085, with each worker’s clips in WebDataset shards of up to 1 GB.

GR00T N1 does not film its own human video. The paper says it first sources “diverse human egocentric video data from open datasets”, and lists seven, covering cooking, assembly, tool use and other “task-oriented activities performed in natural environments”. Capture hardware therefore varies by dataset and the paper does not restate it. Its robot data is different: teleoperation on a Fourier GR1 using tracked wrist poses and hand-tracking gloves, retargeted to the robot.

Egocentric-10K states its scale in one table: 10,000 hours, 1.08 billion frames, 192,900 clips, a median clip of 180.0 seconds and a mean of 4.68 hours per worker. Its diversity is of one kind: every clip comes from a real factory, across the factory folders in its structure.

GR00T N1 does not summarise its human-video hours in the text reviewed; the paper points to Table 7 for per-dataset statistics. The numbers it does state describe the rest of the pyramid: 88 hours of in-house teleoperation expanded to 827 hours of neural trajectories, 140,000 AgiBot-Alpha trajectories, and 780,000 simulation trajectories described as equivalent to 6,500 hours of human demonstration data. Its human video spans several activity types, but the paper reports no single hours figure to set beside 10,000.

Egocentric-10K ships a JSON metadata file beside each video: factory and worker identifiers, a video index, duration, dimensions, frame rate, file size and codec. Each worker folder also holds an intrinsics.json with calibrated fisheye parameters. The card text reviewed states no action labels, hand pose or language annotations.

GR00T N1 has no action signal in its human video, so it manufactures one. It trains a VQ-VAE on a current frame and a future frame to produce a latent-action label, and the paper uses “learned latent actions as flow-matching targets” for human videos. For neural trajectories it also uses an inverse dynamics model to infer pseudo-actions. The repository adds that N1.7 uses a relative end-effector action space shared across robot and human embodiments. See /glossary/latent-actions and /glossary/inverse-dynamics-model.

GR00T N1 uses a layered recipe. Human video forms the base, supplying what the paper calls broad visual and behavioural priors. Synthetic data forms the middle: simulation trajectories built from a small set of human demonstrations, and neural trajectories generated by fine-tuned video models. Real-robot data completes the top and grounds the model in real execution. All layers are co-trained in pretraining, then the model is post-trained on a single embodiment, with neural trajectories sampled 1:1 against real trajectories in the low-data case.

Egocentric-10K supplies one layer only: human video, with no robot, synthetic or action data attached. A team using it would still need to decide how to label it, how to mix it with robot data and how to evaluate transfer. The card text reviewed reports no such experiment.

The Egocentric-10K card states: “Licensed under the Apache 2.0 License.” Access is gated: the card says users must agree to share their contact information. The card text reviewed does not describe consent or release terms for the workers filmed.

The Isaac-GR00T repository lists the Apache License 2.0, and the README says GR00T N1.7 is “fully commercially licensable under Apache 2.0”. The paper states the GR00T-N1-2B checkpoint, training data and simulation benchmarks are publicly available. The human videos it draws on come from other datasets, each with its own terms, which the paper does not restate.

Egocentric-10K leaves open anything beyond factory work, any measured action or hand-pose signal, and any evidence that it improves a robot policy. The card text reviewed states none of these. It also gives no way to specify new tasks: it is a fixed release.

GR00T N1 leaves open the contribution of human video, the human-data licence chain and long-horizon work. The paper says the model “focuses primarily on short-horizon tabletop manipulation tasks” and names long-horizon loco-manipulation as future work. Its results are reported against baselines, and the sections reviewed include no run that removes human video, so they do not measure that layer alone.

Custom collection is not a substitute for either source. It fills coverage that a fixed corpus and a fixed pyramid cannot. A program might use Egocentric-10K or GR00T N1 as an input, then commission a fast, on-demand pilot batch scoped to the tasks, scenes 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 to request a custom pretraining dataset.