ropedia-ai/Ropedia-Xperience-2

Ropedia Xperience 2

A Gold Standard for Scaling Physical AI Environments and Behaviors. 1,500+ scanned scenes and 4,000+ egocentric recordings, registered in one metric frame.

📄 Paper (arXiv:2611.17060) · 🤗 Paper page · 🌐 Project page

First batch online. Xperience 2 is being published in batches. The first batch (126 scenes, about 1.07 TB) is complete and listed in the tables below with thumbnails. Further batches will follow until all 1,500+ scenes are online. Watch the repository for updates.

1,500+	scenes scanned in 3D
596	locations
12	scene categories
4,000+	egocentric recordings
74	tasks
131 h	of first-person video

Capture

Scanned once. A handheld LiDAR scanner captures the whole space in one pass, in 3D and at metric scale.

Multiple tasks, recorded in one scene. Then the operator wears HOMIE Gen 2 (head-mounted, 4 fisheye cameras, IMU, 4-channel audio) and records every task assigned to the scene.

Everything in one scene. Every scan is metric and gravity-aligned, and every recording is registered to its scan.

Scenes

Scans of 596 locations: homes, stores, restaurants, car parks and streets, in 12 categories: living room, bedroom, bathroom, kitchen, multi-room, office, meeting room, store, restaurant, parking lot, outdoor corridor and urban street. Each reconstruction is reviewed and graded: perfect, usable or reuse candidate.

What each scene contains

• 3D scan (handheld LiDAR scanner): a 3D Gaussian splat, a coloured triangle mesh, and the scanner's multi-camera images with masks and COLMAP poses.
• HOMIE Gen 2 walk-through: undistorted keyframes from the 4 head cameras with camera poses, IMU, a mesh and point cloud, and time-segmented scene descriptions.
• HOMIE Gen 2 task recordings: full-body and hand motion capture (SMPL-H style body, MANO hands), the head-mounted device trajectory, IMU, and annotations of the task, its sub-tasks and actions, with objects and hand states.

Benchmarks

Xperience 2 scenes are used in three benchmarks (paper):

• Depth & Geometry: 11 models.
• 3D Object Detection: 7 methods. Annotators labelled the category and 3D box of every object in 20 scenes.
• Spatial Intelligence: 12 models, 500 questions on the same 20 scenes.

Quick start

Browse the scene table (thumbnails, categories, descriptions, tasks) with 🤗 Datasets:

from datasets import load_dataset

scenes = load_dataset("ropedia-ai/Ropedia-Xperience-2", "scenes", split="train")
print(scenes[0]["category"], scenes[0]["tasks"])

Download one scene instead of the whole dataset:

hf download ropedia-ai/Ropedia-Xperience-2 --repo-type dataset \
    --include "scenes/2092913920584933377/*" --local-dir xperience2

Only the motion capture and annotations (skips the large image archives):

hf download ropedia-ai/Ropedia-Xperience-2 --repo-type dataset \
    --include "metadata/*" "scenes/*/mocap/*" --local-dir xperience2

Stream the keyframes of a walk-through (WebDataset tar, one sample per timestamp, up to four cameras):

import webdataset as wds

url = "https://huggingface.co/datasets/ropedia-ai/Ropedia-Xperience-2/resolve/main/scenes/<scene_id>/mapping/<capture_id>/rgb.tar"
for sample in wds.WebDataset(url).decode("pil"):
    print(sample["__key__"], sample["cam0.jpg"].size)   # key = timestamp in ns
    break

Repository layout

metadata/
  scenes.parquet      one row per scene: thumbnail, category, description, tasks, sizes
  captures.parquet    one row per walk-through (mapping) or task recording (mocap): date, duration, frames, task
  segments.parquet    one row per annotated time segment (scene text, sub-task, actions)
  files.parquet       every file of the uploaded batches with its path, size and kind
scenes/<scene_id>/
  preview/thumb.webp  overview render of the scan
  3d_scanning/        handheld LiDAR scan
    gaussian_splat.ply      3D Gaussian splatting, degree-0 SH (x y z, normals, f_dc, opacity, scale, rot)
    mesh.obj                triangle mesh with per-vertex colour
    colmap_sparse.zip       COLMAP model (cameras / images / points3D, txt + bin)
    images_fisheye.zip      camera_<id>/<unix_s>.jpeg from the scanner's cameras
    masks_fisheye.zip       8-bit PNG masks, one per image
  mapping/<capture_id>/   HOMIE Gen 2 walk-through
    rgb.tar                 undistorted pinhole keyframes, cam0..cam3 per timestamp
    poses/cam{0..3}.txt     camera-to-world per keyframe: image tx ty tz qx qy qz qw
    calibration.json        camera intrinsics and the original fisheye calibration
    imu.csv                 ~200 Hz gyro (rad/s) and accelerometer (m/s^2)
    mesh.ply, point_cloud.ply
    caption.json            time-segmented scene descriptions
    metadata.json           recording settings and timing
  mocap/<capture_id>/     HOMIE Gen 2 task recording
    full_body_mocap.hdf5    30 Hz body: root and head poses, joint quaternions, 52 keypoints, contacts, betas
    hand_mocap.hdf5         per-frame hand detections, 2D/3D keypoints, MANO parameters
    hand_mocap_opt.hdf5     stereo-optimised MANO hands
    poses/f_slam.txt        30 Hz device trajectory: t_s tx ty tz qx qy qz qw
    imu.csv, calibration.json (fisheye, Kalibr camchain style)
    caption.json            task, sub-tasks, actions with frame ranges, objects, hand states
    mocap_qc_report.json    motion-capture quality metrics
    metadata.json

scene_id is a stable 19-digit identifier; capture_id is a UUID. A walk-through and a task recording with the same capture_id come from the same HOMIE Gen 2 session.

Conventions

• Time. Unix epoch: keyframe names and IMU in nanoseconds; f_slam.txt in seconds; scanner image names in seconds. Frame numbers in task-recording captions are 30 Hz frame indices starting at 1.
• Poses. Text pose files are camera-to-world with quaternions in x y z w order. COLMAP files follow COLMAP (world-to-camera, w x y z).
• Frames. HOMIE Gen 2 files are given in the SLAM world frame of their session (+Z up). In the files of the first batch, the scan has its own frame; the recording-to-scan registration is not part of these files yet.

Annotations

Captions were generated automatically by a vision-language model from the egocentric video. They can contain wrong object names, counts or actions.

Privacy

Data was recorded in homes and public places. If you find personal information that should not be here, please open a discussion on this repository and we will remove it.

License

CC BY-NC 4.0: free for research and other non-commercial use with attribution. For commercial licensing, please open a discussion on this repository.

Citation

@article{li2026xperience2,
  title   = {Ropedia Xperience 2: A Gold Standard for Scaling Physical AI Environments and Behaviors},
  author  = {Hao Li and
             Fangzhou Hong and
             Zhaoxi Chen and
             Hui En Pang and
             Kairui Hu and
             Mingze Sun and
             Runmao Yao and
             Xinhao Ji and
             Xiyuan Zheng and
             Xuanli Chen and
             Yalun Dai and
             Yap Kim Hui and
             Dingwen Zhang and
             Yuanyuan Gao and
             Ziang Cao and
             Cheng Huang Yean and
             Jackson Peng and
             Rui Duan and
             Yaobo Liu and
             Zetong Wang and
             Ziwei Liu},
  journal = {arXiv preprint arXiv:2611.17060},
  year    = {2026},
  url     = {https://arxiv.org/abs/2611.17060}
}