Xp-Teleop

Long-horizon humanoid manipulation on a Unitree G1 with Inspire FTP hands, recorded from a head-mounted RGB camera and cut into individual actions.

Xp-Teleop is the long-horizon loco-manipulation benchmark introduced in Ropedia W0.

27 episodes · 543 action segments · 96,064 frames · 30 fps · 1280×720

Layout

meta/
  info.json            schema for every field
  takes.jsonl          27 episodes
  segments.jsonl       543 action segments — the index
  tasks.jsonl          task list
  calibration.jsonl    camera intrinsics and extrinsics per episode
takes/<take_id>/segments/<segment_id>/
  video_rgb.mp4        clean RGB, without burned-in annotations
  video_keypoints.mp4  RGB with skeleton/keypoint overlays
  frame_timestamps.csv original per-frame timestamps
  frames.npz           per-frame arrays (see meta/info.json)
  masks.zip            robot body mask, one PNG per frame
  imu.csv              head IMU, ~1.6 kHz
  tactile.jsonl        Inspire tactile, 1062 taxels per hand
  caption.json         the action annotation

Load

import json, numpy as np

segments = [json.loads(l) for l in open("meta/segments.jsonl")]
s = segments[0]
d = np.load(f"{s['path']}/frames.npz", allow_pickle=True)

d["uv"]       # (n, 51, 2) keypoints in pixels
d["xyz_cam"]  # (n, 51, 3) keypoints in metres, camera frame
d["cam_c2w"]  # (n, 4, 4)  camera-to-world
d["q"]        # (n, 35)    joint angles

Clean RGB video

video_rgb.mp4 in every segment is encoded directly from the original JPEG frames (1280×720, 30 fps, H.264, CRF 18), without burned-in annotations; frame n corresponds to row n of frames.npz. frame_timestamps.csv gives the original per-frame timestamps (frame_idx, monotonic_ns, device_ts_usec) in the same order. video_keypoints.mp4 is the same footage with the robot skeleton rendered on top. The indices meta/raw_video_segments.jsonl and meta/raw_video_sources.jsonl list the source take and frame range of every segment.

Notes

• frame_timestamps.csv also carries color_path / depth_path, the relative paths of the original colour JPEG and depth frames in the recording; those raw files are not part of this dataset.
• In pour_and_wash_the_plate, segments c026–c032 have a video_keypoints.mp4 rendered from an earlier version of the skeleton; frames.npz and video_rgb.mp4 are the reference.

License

CC BY 4.0.

Citation

@article{liu2026ropedia,
  title={Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience},
  author={Liu, Tianqi and Zhao, Hanfei and Liang, Qifeng and Pang, Hui En and Xu, Jiarui and Mu, Sayang and Qiu, Penghao and Zhang, YingSheng and Wang, Jun and Cao, Ziang and Sun, Mingze and Cao, Yukang and Hong, Fangzhou and Chen, Zhaoxi and Liu, Ziwei},
  journal={arXiv preprint arXiv:2610.26855},
  year={2026}
}