Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

NVIDIA Human-1M-100K

100,000 hours of first-person human work video with hand poses and timed task labels

License

CC BY 4.0

Data: CC BY 4.0, per the Governing Terms in the dataset card and its Hugging Face licence tag; the card says it is ready for commercial and non-commercial use. Attribution: reuse must credit NVIDIA and link the licence. People in the video: CC BY 4.0 does not license privacy or publicity rights, and the card does not describe consent. Hardware prototypes: a separate repository, with software under Apache-2.0 and hardware designs under CC BY 4.0. Models: GR00T 2 is not released, so its terms are not known. Checked 2026-11-10 against README.md and DATASET_CARD.md in the dataset repository and the LICENSE files of the hardware repository.

Overview

Human-1M-100K is the openly released part of Human-1M, NVIDIA's collection of about one million hours of first-person video of people doing physical work, recorded with sensors so that the video comes with head and hand motion and language. The open subset holds 100,000 hours in 3,405,573 episodes, with 406,815 recorded task clusters, 41,288 collectors, 81,557 scenes and nine environment categories, filmed both in real homes and workplaces and in studios. Every episode carries head-camera video, mono or stereo and sometimes with wrist cameras, the head-camera pose, 21 keypoints per hand with confidence, a one- or two-sentence goal (T1) and a sequence of timed actions of 1 to 10 seconds (T2). The data is in the LeRobot v3.0 format, licensed CC BY 4.0 for commercial and non-commercial use, and was published on Hugging Face on 4 November 2026. NVIDIA says Human-1M is training data for its Cosmos and GR00T models, and its GR00T 2 research page names it as the source of that model's human video.

Human-1M is NVIDIA's collection of sensorized human demonstrations: about one million hours of egocentric video of people doing work with their hands, from preparing food and stocking shelves to packing orders, assembling parts and repairing equipment. On 4 November 2026 NVIDIA released 100,000 hours of it as Human-1M-100K on Hugging Face under CC BY 4.0. The release counts 3,405,573 episodes, 406,815 recorded task clusters, 41,288 collectors, 81,557 scenes and nine environment categories. Some of it was recorded in the wild, in real homes and workplaces, and some in studios where specific skills, assembly and lab protocols can be repeated.

An episode is one person doing one piece of useful work toward a measurable goal, usually something that takes 5 to 10 minutes. The video must be at least 30 FPS and 480p at unaltered speed, with one or both hands in view for at least 95% of the episode. The head camera is mono or stereo, and some episodes add a camera on each wrist. Each frame carries the 6-DoF pose of the head camera, 21 keypoints per hand as positions in metres and rotations, and a presence mask and confidence for each hand. Language comes at two levels: T1 is a one- or two-sentence summary of the goal, the main actions and the key objects, and T2 splits the whole episode into consecutive segments of 1 to 10 seconds, each naming which hand does what to which object. Episodes also record camera calibration, a country code, whether the place was public or private, and pseudonymized collector, scene and source identifiers.

NVIDIA describes four quality gates combining human review and automatic checks, covering annotation, task usefulness, diversity and 3D tracking. Its own audit chart covers 25 delivery weeks with 561,000 accepted hours and 461,000 rejected submission hours.

The data is stored in LeRobot v3.0 datasets, organised by vendor, batch and version, with Parquet tables for frames, episodes, tasks, subtasks, environments and scenes and one MP4 file per camera. The field reference documents every column. The full data sits in a public Hugging Face storage bucket, which reported 1,321 TB in 2,561,357 files on 10 November 2026, so most users will take a slice. To look before downloading, the dataset repository has short preview clips and three complete sample episodes, and the Human-1M Dataset Explorer shows video, 3D reconstruction, robot retargeting, T2 labels and a search over task clusters in the browser.

NVIDIA's GR00T 2 research page says the model was pretrained on 402,730 hours of egocentric human video from Human-1M, together with 10,184 hours of handheld-gripper demonstrations and 23,500 hours of robot teleoperation. In NVIDIA's own test, raising the human video from none to all of it lifted task progress on 72 robot tasks from 50% to 70%. GR00T 2 itself is not released yet: the page lists its code as available in December 2026. The dataset card also says Human-1M powers NVIDIA's Cosmos models. Separately, NVIDIA has published the hardware designs for the handheld and wearable devices it built to collect GR00T 2 data, with their software in the GR00T data hardware repository.

Why it matters

Robot learning needs far more examples of skilled hand work than robots can record: teleoperation is slow and costly, and ordinary video of people lacks the hand motion and labels a model can learn actions from. Human-1M-100K offers 100,000 hours of real work with head and hand pose and timed action labels already attached, under CC BY 4.0, which allows commercial use. That makes it the largest open egocentric human-manipulation dataset in this catalogue, next to smaller releases such as Eidon Tracker POV and robot-free manipulation data like HiFi-UMI-2K. It is material for pre-training VLA and world models, for imitation learning research and for motion retargeting from human hands to robot hands, and it arrives in the LeRobot format many robot-learning tools already read. It is also, by NVIDIA's account, the human data behind the next GR00T foundation model, which lets anyone study that data before the model itself is out.

Problem it solves

Teaching robots to use their hands takes large amounts of demonstration data, and collecting it by robot teleoperation is slow, expensive and limited to the tasks a lab can stage. Video of people doing real work is plentiful but has no hand motion or task structure attached, so a model cannot easily learn actions from it. Human-1M-100K provides real human work at scale with the missing parts already added: head and hand pose, per-hand confidence, a goal for each episode and timed action labels, in a documented LeRobot format and under a licence that allows commercial use.

Key takeaways

  • 100,000 hours of first-person human work, the open subset of NVIDIA's one-million-hour Human-1M collection
  • 3,405,573 episodes from 41,288 collectors across 81,557 scenes, 406,815 task clusters and nine environment categories
  • Mono or stereo head video, optional wrist cameras, 6-DoF head-camera pose and 21 keypoints per hand with confidence
  • Two-level language labels: a goal for each episode (T1) and consecutive timed actions of 1 to 10 seconds (T2)
  • LeRobot v3.0 format, with a field reference, preview clips, three complete sample episodes and a browser explorer
  • CC BY 4.0, open for commercial use; NVIDIA names Human-1M as the human video behind GR00T 2

Who it is for

  • Researchers pre-training VLA, world and robot foundation models on human demonstration data
  • Teams working on retargeting human hand motion to robot hands and dexterous manipulation
  • Computer-vision researchers studying egocentric video, hand pose and action segmentation
  • LeRobot users who want large-scale human data in a format their tools already read

Limitations

  • Human demonstrations, not robot actions: using them to train a robot needs retargeting or a model that learns from video
  • Very large: the storage bucket reported 1,321 TB in 2,561,357 files on 10 November 2026, so most users will download a selected part
  • Only 100,000 of the roughly one million hours of Human-1M are open; the rest stays with NVIDIA
  • The bucket folder is named release-staging and was still being updated on 10 November 2026
  • Claims about how much the data improves models are NVIDIA's own, from its GR00T 2 research page; GR00T 2 itself is not released yet
  • Recorded by several collection sources with different camera setups, and some fields, such as the rig or the privacy category, are missing for some sources
  • The card does not say how consent was obtained from the people recorded or whether faces are blurred; identifiers are pseudonymized
  • CC BY 4.0 requires attribution and does not license the privacy or publicity rights of the people in the video

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.