nvidia/PhysicalAI-Human-1M-100K

Across AI, a common route to scale is distilling human experience into models: what people write, see, and create. The physical AI analogue is sensorized human data—egocentric video, dense language annotations, and head and hand pose that connect perception, action, and intent.

Human-1M: Scaling Physical AI through Human Experience

Explore the dataset →

1M hours Human-1M collection

100,000 hours Open-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.

Ready for commercial and non-commercial use.

Human-1M, at a glance

Anatomy of an episode

Video, motion, and language come together at the episode level. T1 captures the overall goal; T2 breaks the task into timed actions. Stereo and wrist views connect the scene to head and hand motion.

Bundle flower stems connects synchronized stereo and wrist video, six timed T2 actions, stereo 3D reconstruction, head and hand tracks, and robot motion retargeting in one complete demonstration.

The episode is Human-1M’s atomic unit: a proficient demonstrator performs economically useful physical work toward a discrete, measurable goal, usually achievable within 5–10 minutes. Video requirements are at least 30 FPS / 480p, unaltered speed, visible and in-focus hands and task-relevant objects, and one or more hands in frame for at least 95% of the episode.

T1 is a 1–2 sentence summary of the primary goal, major actions, and key objects. T2 specifies a contiguous sequence of 1–10-second segments covering the episode, bounded by observable action intent. Each segment specifies hand laterality, precise verbs, and unambiguous objects, and is interpretable on its own.

Motion annotations include 6-DoF head-camera pose and 21 keypoints per hand, with tracking masks and confidence. Full episode requirements.

Explore video, 3D reconstruction, retargeting, and T2 annotations in Spaces →

Human-1M diversity

The slice spans 406,815 recorded task clusters, 81,557 scenes, 41,288 collectors, and nine environment categories. In-the-wild collection captures variation in real homes and workplaces; studios make targeted skills repeatable, including curated tasks, assembly, and in-lab protocols.

From preparing food and stocking shelves to packing orders, assembling parts, and repairing equipment, these demonstrations capture economically valuable human activity. The Space’s Everyday index of real human work associates tasks with broad categories from the BLS Occupational Outlook Handbook.

Explore task clusters, search, and the human-work index

Quality at every step

Four quality gates combine human review and quantitative checks: high quality dense annotation with accurate language and timing; high quality tasks with useful, clearly visible interactions; high quality diversity at the task, scene and collector, and environment levels; and high quality 3D tracking checked for alignment, bone-length consistency, and motion smoothness.

Production T1 audit history · 25 delivery weeks. 561K accepted T1 hours · 461K gross rejected submission hours.

Metadata and dataset structure

The release is organized by vendor → batch → version. Each v0/shard_*/ directory contains a LeRobot v3.0 dataset. Video, language, and motion align through frame timestamps and episode metadata.

release-staging/oss_100k_v1/
└── vendors/
    └── {vendor_id}/
        └── {batch_id}/
            ├── versions.json
            └── v0/
                └── shard_00/
                    ├── data/
                    │   └── chunk-000/file-000.parquet
                    ├── meta/
                    │   ├── info.json
                    │   ├── episodes/chunk-000/file-000.parquet
                    │   ├── tasks.parquet
                    │   ├── subtasks.parquet
                    │   ├── environments.parquet
                    │   └── scenes.parquet
                    └── videos/
                        └── observation.images.{camera}/
                            └── chunk-000/file-000.mp4

Component	Contents
Frame data	Timestamps, head-camera and hand pose, masks, confidence, collector/scene IDs, and label indices.
Episode metadata	Episode and rig IDs, anonymous vendor_id, camera intrinsics/extrinsics, and data/video intervals.
Language and context	Task descriptions, timed subtask occurrences, environment identifiers and labels, and scene descriptions.
Video	Monocular or stereo head views, with optional wrist cameras.

Field reference · Machine-readable schema · Download complete sample episodes

Governing Terms: This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0) available at https://creativecommons.org/licenses/by/4.0/legalcode.