Xperience-10M is a large-scale egocentric multimodal dataset of human experience for embodied AI, robotics, world models, and spatial intelligence.

It contains 10 million experiences (interaction) and 10,000 hours of synchronized first-person recordings with six video streams, audio, stereo depth, camera pose, hand mocap, full-body mocap, IMU, and hierarchical language annotations. With 2.88 billion RGB frames, 720 million depth frames, 576 million pose and mocap frames, and ~1 PB of total data, Xperience-10M is, to our knowledge, by far the largest egocentric dataset with structured 3D/4D multimodal annotations.

Xperience-10M is built for training and evaluating models that do not just see the world, but also understand motion, geometry, interaction, and embodied behavior as a unified stream of experience.

It is designed to support research and product development in:

embodied AI
world modeling
robot learning from human experience
egocentric perception
action understanding
multimodal foundation models
human-object interaction
sensor fusion
3D/4D scene and motion understanding
real-to-sim and sim-to-real pipelines

Check out xperience-10m-sample for data sample (coffee making!).

Check out HOMIE-toolkit for sample code to load/visualize Xperience data.

What makes Xperience-10M different

Most existing egocentric datasets provide only a partial view of embodied experience: RGB video, sparse labels, or limited motion signals. Xperience-10M is designed differently.

It treats experience as a multimodal, structured, and temporally grounded signal. Each episode can include:

what the wearer sees
what the wearer hears
how the camera moves through space
how the hands move
how the full body articulates
what the depth geometry looks like
what the IMU measures
what task, subtask, action, interaction, and objects are involved

This makes Xperience-10M especially useful for building systems that learn from real human experience at scale.

Supported Tasks and Use Cases

Xperience-10M can support a broad range of tasks, including but not limited to:

egocentric action recognition
task and subtask prediction
action captioning
temporal action localization
human-object interaction understanding
object grounding and recognition
audio-visual learning
visual-language pretraining
embodied reasoning
stereo and monocular depth estimation
visual odometry and trajectory learning
SLAM and camera pose estimation
hand pose estimation
body motion estimation
multimodal sensor fusion
imitation learning and behavior modeling
policy learning for robotics
world model training

Languages

The language annotations are in English.

Dataset Structure

Xperience-10M is organized as a collection of episodes. Each episode contains synchronized egocentric video files together with a unified annotation.hdf5 file storing annotations, calibration, geometry, motion, inertial signals, and metadata.

Episode Layout

A typical episode folder contains:

episode/
├── fisheye_cam0.mp4  # fisheye camera 0
├── fisheye_cam1.mp4  # fisheye camera 1
├── fisheye_cam2.mp4  # fisheye camera 2
├── fisheye_cam3.mp4  # fisheye camera 3
├── stereo_left.mp4   # rectified stereo left
├── stereo_right.mp4  # rectified stereo right
└── annotation.hdf5   # all annotations and metadata

Modalities

Each episode may include the following modalities:

Four fisheye video streams
Two rectified stereo video streams
Audio aligned with all video streams
Stereo depth
Camera pose / SLAM trajectory
Two-hand motion capture
Full-body motion capture
IMU
Episode metadata
Hierarchical language captions, including:
- task
- subtask
- action
- interaction
- objects

HDF5 Annotation Structure

The annotation.hdf5 file stores synchronized annotations and metadata in the following structure:

annotation.hdf5
├── calibration/
│   ├── cam0/
│   ├── cam1/
│   ├── cam2/
│   ├── cam3/
│   └── cam01/
├── slam/
│   ├── quat_wxyz
│   ├── trans_xyz
│   ├── frame_names
│   └── point_cloud
├── depth/
│   ├── depth
│   ├── confidence
│   ├── scale
│   ├── depth_min
│   └── depth_max
├── hand_mocap/
│   ├── left_joints_3d
│   ├── right_joints_3d
│   ├── left_translation
│   ├── right_translation
│   ├── left_mano_hand_pose
│   ├── right_mano_hand_pose
│   ├── left_mano_hand_global_orient
│   ├── right_mano_hand_global_orient
│   ├── left_mano_hand_betas
│   └── right_mano_hand_betas
├── full_body_mocap/
│   ├── keypoints
│   ├── contacts
│   ├── Ts_world_cpf
│   ├── Ts_world_root
│   ├── body_quats
│   ├── left_hand_quats
│   ├── right_hand_quats
│   ├── betas
│   └── frame_nums
├── imu/
│   ├── device_timestamp_ns
│   ├── accel_xyz
│   ├── gyro_xyz
│   └── keyframe_indices
├── video/
│   ├── device_timestamp
│   ├── frame_number
│   └── length_sec
├── metadata/
└── caption

Annotation Details

Go checkout HOMIE-toolkit for more tech details of our annotations.

Key Statistics

Statistic	Value
Total number of experiences (interactions)	10M
Video with audio	10,000 h
RGB frames	2.88B
Depth frames	720M
Camera poses	576M
Mocap frames	576M
IMU frames	7.2B
Caption sentences	16M
Caption words	200M
Caption vocabulary size	6K
Number of objects	350K
Total storage	~1 PB
Total trajectory length	39,000 km

Uses

Direct Use

Xperience-10M is intended for direct use in:

multimodal pretraining
egocentric perception research
action understanding
motion understanding
3D/4D reconstruction and tracking
action-language grounding
embodied foundation model training
robotics and imitation learning
world model training

Out-of-Scope Use

Xperience-10M is not intended for:

identity recognition
person re-identification
biometric profiling
surveillance applications
inferring sensitive personal attributes
safety-critical deployment without additional validation and safeguards

Limitations

Despite its scale and richness, Xperience-10M still has limitations.

It reflects the environments, devices, and activity distributions represented in the collected data.
Depth, pose, SLAM, and mocap annotations may contain noise or estimation error.
Semantic annotations may not fully capture every relevant contextual factor in an episode.
The scale of the dataset may require substantial storage and compute infrastructure for training.

Social Impact

Xperience-10M can help advance world models, embodied AI, assistive systems, spatial intelligence, and robot learning from real-world human experience.

At the same time, egocentric multimodal data raises important questions around privacy, consent, and downstream misuse. We encourage all users to work with the dataset responsibly and to align usage with privacy protection, human-centered AI principles, and beneficial real-world applications.

Privacy, Ethics, and Consent

Because Xperience-10M contains egocentric recordings of real-world human activity, privacy and consent are central considerations.

All data in Xperience-10M was collected and processed under appropriate consent and review procedures. Personally identifying or sensitive content is handled according to the dataset release policy. Access to some or all portions of the dataset may be controlled to protect participant privacy and support responsible use.

Access

Xperience-10M is released for research and other non-commercial uses under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Because of the scale of the dataset (~1 PB) and the sensitive nature of egocentric multimodal data, access may be provided through controlled distribution channels. Users are expected to follow the dataset usage terms and any accompanying privacy, security, or redistribution requirements released with the dataset.

Before using Xperience-10M, please make sure you understand:

the non-commercial restriction
attribution requirements
any privacy and responsible-use conditions associated with the data
any additional access procedures specified by the dataset maintainers

Citation

@dataset{xperience_10m,
  title={Xperience-10M: A Large-Scale Egocentric Multimodal Dataset with Structured 3D/4D Annotations},
  author={Ropedia},
  year={2026},
  publisher={Hugging Face},
  note={Dataset}
}

Downloads last month: 2,021