100K-HOUR RELEASE100,000 hoursOpen-source subset
Human-1M is the largest and most diverse high-quality egocentric dataset which contains egocentric video, dense language annotations, head and hand motion focused on real world manipulation. Human-1M spans economically valuable work in the wild and in controlled studios, and powers NVIDIA's flagship Cosmos and GR00T models.
Follow a task from its overall goal to each hand–object interaction, then explore stereo geometry and motion.
Play the stereo video or load 3D to explore both views together.
Inspect human-to-robot motion alignment in the retargeting viewer.
Area: recorded hours
Useful training data needs accurate instructions, clear interactions, reliable motion, and varied experience. Human-1M combines human review and quantitative checks across four quality gates.
Language connects a demonstration to what the person is trying to accomplish.
Clear, active demonstrations expose how skilled people handle objects and tools.
Diversity is checked at the task, scene and collector, and environment levels.
Hand geometry supports motion learning, reconstruction, and robot retargeting.
Our diversity tool groups semantically similar task descriptions, measures task coverage, and flags excessive repetition.
Scene and collector IDs track coverage across people and collection locations, limiting concentration in a single scene or collector.
Environment-level targets balance hours across nine categories, spanning controlled studios, households, and a range of workplaces.
Production task annotation audit history · 25 delivery weeks
Real homes and workplaces expose the variation a robot will encounter: unfamiliar objects, clutter, different layouts, tools, and human strategies.
Standardized workspaces support repeatable skills, targeted task collection, and controlled comparisons across people and sensor configurations.
Controlled environments with standardized setups and recording protocols focused on tabletop manipulation at natural human speed.
Tasks selected to cover target skills, object interactions, and robot task distributions.
Multi-step tasks involving fitting, inserting, fastening, or joining parts for assembly kits.
Curated and assembly tasks with 100% hands in frame, no idle time, and robot-aligned demonstration speed.
A single head-mounted RGB camera, with depth and 3D hand pose inferred from monocular observations.
Synchronized head-mounted cameras provide depth cues for 3D hand tracking and scene reconstruction.
Additional wrist-mounted views capture close-up manipulation and hand–object interactions.
6-DoF head-camera pose and 21 keypoints per hand, with tracking masks and confidence, camera intrinsics and extrinsics, and rig, collector, scene, and environment identifiers.
Explore human tasks across the broad occupation groups in the U.S. Bureau of Labor Statistics Occupational Outlook Handbook.
Human-1M tasks are associated with related work categories using actions and objects in their canonical names or supplied instructions. Each task group appears in at most one category. Ambiguous descriptions remain unassigned.
Broad category reference: BLS Occupational Outlook Handbook. Task associations are Human-1M analysis.
vendor_id, batch_id, and the v0 version. Each batch’s versions.json sits alongside v0/.