HumanPlus-1000

HumanPlus-1000 is a large-scale multimodal human behavior dataset designed for learning and modeling human perception, motion, and interaction in real-world environments.

It captures synchronized egocentric visual observations, full-body motion, hand motion, camera motion, and 3D environment information, with the goal of providing paired human perception–action data for embodied intelligence, human motion modeling, and robotics.

Current Release: This repository currently provides a 100-session preview subset of HumanPlus-1000 for demonstrating the dataset format, modalities, annotations, and research use cases.

To request access to the full HumanPlus-1000 dataset, please contact us at info@humanplus.xyz

---

Overview

HumanPlus-1000 is designed around long-horizon human activities captured in real-world environments.

Unlike datasets that provide only egocentric video or isolated motion sequences, HumanPlus aims to align multiple representations of human behavior within the same recording:

• what the human sees,
• how the human body moves,
• how the hands interact with the environment,
• how the human moves through the 3D world,
• and what task is being performed.

Each recording is organized as a session, containing synchronized sensor data, reconstructed motion, calibration information, and metadata.

---

Preview Release

The current Hugging Face release contains:

100 demo sessions

These sessions are provided as a preview of the HumanPlus-1000 data infrastructure and are intended to help researchers:

• inspect the dataset format,
• develop data loaders,
• test multimodal learning pipelines,
• explore human motion and egocentric perception,
• and reproduce example HumanPlus tasks.

The preview subset does not represent the final scale or complete distribution of HumanPlus-1000.

---

Modalities

Depending on the session, HumanPlus may provide:

Egocentric Vision

• Stereo fisheye video
• Rectified / undistorted egocentric video
• Depth estimation

Human Motion

• Full-body motion
• SMPL / SMPL-H parameters
• Global body translation and orientation
• Hand pose

World Motion

• Camera trajectory
• SLAM trajectory
• Camera pose in the reconstructed world coordinate system

Inertial Sensing

• Wearable IMU measurements
• Synchronized orientation / acceleration signals

Calibration

• Camera intrinsics
• Camera distortion parameters
• Stereo extrinsics
• Camera-to-human calibration
• Coordinate-system transformations

Semantic Metadata

• Session information
• Scene
• Task / activity
• Duration
• Available modalities
• Data quality indicators

---

HDF5 Annotation Structure

The annotation.hdf5 file stores time-synchronized multimodal annotations, including egocentric vision, human motion, hand motion, IMU measurements, SLAM trajectories, and 3D environment information. All modalities are temporally aligned through shared timestamps. Human motion and camera trajectories are further aligned in a common world coordinate system, enabling joint modeling of human behavior and environmental interaction.

annotation.hdf5
│
├── calibration/
│   ├── @transform_convention
│   ├── fisheye_left/
│   │   ├── intrinsics
│   │   ├── distortion
│   │   ├── camera_model
│   │   ├── distortion_model
│   │   └── image_size
│   ├── fisheye_right/
│   │   ├── intrinsics
│   │   ├── distortion
│   │   ├── camera_model
│   │   ├── distortion_model
│   │   └── image_size
│   ├── stereo/
│   │   └── T_right_left
│   ├── rectified_stereo/
│   │   ├── @left_half
│   │   ├── @hamer_camera_side
│   │   ├── R_rect_left
│   │   ├── R_rect_right
│   │   ├── image_size
│   │   └── intrinsics
│   ├── T_head_camera
│   ├── T_mocapworld_slamworld
│   └── body_imu/
│       ├── R_mocapworld_imuworld
│       └── device2bone
│
├── video/
│   ├── fisheye_left_file
│   ├── fisheye_right_file
│   ├── image_size
│   └── timestamp_ns
│
├── slam/
│   ├── @camera_frame
│   ├── T_slamworld_camera
│   └── point_cloud
│
├── depth/
│   ├── @min_depth_m
│   ├── @max_depth_m
│   ├── @invalid_value
│   ├── depth
│   └── intrinsics
│
├── imu/
│   ├── body/
│   │   ├── timestamp_ns
│   │   ├── accel_xyz
│   │   ├── orientation
│   │   ├── acc
│   │   └── rot
│   └── head/
│       ├── timestamp_ns
│       ├── accel_xyz
│       ├── gyro_xyz
│       ├── acc
│       └── rot
│
├── body_motion/
│   ├── T_mocapworld_root
│   ├── smplh_pose
│   ├── body_keypoints
│   └── foot_contact_probability
│
├── hand_motion/
│   ├── left/
│   │   ├── mano_hand_global_orient
│   │   ├── mano_hand_pose
│   │   ├── mano_betas
│   │   ├── joints_3d
│   │   ├── wrist_position_camera
│   │   ├── valid
│   │   └── confidence
│   └── right/
│       ├── mano_hand_global_orient
│       ├── mano_hand_pose
│       ├── mano_betas
│       ├── joints_3d
│       ├── wrist_position_camera
│       ├── valid
│       └── confidence
│
├── synchronization/
│   ├── utc_ns
│   └── video_frame_indices
│
└── behavior_annotation/
    ├── activity_summarization
    ├── motion_narration
    └── atomic_action

Data Organization

A HumanPlus recording is organized around a session.

Example:

HumanPlus-1000/
├── metadata/
│   └── sessions.parquet
│
└── data/
    ├── HP_S000001/
    │   ├── fisheye_left.mp4
    │   ├── fisheye_right.mp4
    │   ├── annotation.hdf5
    │   └── metadata.json
    │
    ├── HP_S000002/
    │   └── ...
    │
    └── ...