Accept the terms to access HuRo-Egocentric10K
Access is granted automatically once the form is submitted.
This repository holds two kinds of file under two different licences.
The Parquet tables β hand keypoints, camera parameters and poses, arm masks, language instructions and retargeted ALLEX joint trajectories β are CC BY 4.0. Egocentric-10K is Apache-2.0, and the licences of the models in the pipeline are not held to reach the numbers those models produced; the card explains why.
The arm-removed video is CC BY-NC 4.0, because the frames are produced by ProPainter, whose S-Lab licence is non-commercial.
No unmodified Egocentric-10K frames and no robot-rendered frames are included.
Your username, email address and the fields below are collected solely to administer access to this dataset and to evidence acceptance of these terms. They may be provided in aggregate to the licensors of the material for the purpose of verifying compliance.
Log in or Sign Up to review the conditions and access this dataset content.
HuRo-Egocentric10K
The Egocentric-10K share of the HuRo dataset: 62,678 segments / 8,616,805 frames of factory first-person video across 6,773 clips, 79.8 hours at 30 fps. Each segment holds the frames with the human arms inpainted out, per-frame hand and camera annotation, a language instruction, and the Allex joint trajectory retargeting produced from them.
This is the robot-independent half of HuRo. No robot is rendered into the frames: you retarget to the robot you care about and render it yourself, which is stages 8 to 10. The expensive stages 1 to 7 β hand reconstruction, camera estimation, narration, inpainting β are already done here and never run again.
Produced by the HuRo robotization pipeline. Paper: arXiv:2609.10706 Β· Project page: 3587jjh.github.io/HuRo
What's in this repository
The repository name says which corpus this came from, not what is inside. What is here:
| Human annotation tables | β
clips_chunked/original/annot/ β one Parquet table per segment, robot-independent |
| Arm-removed video | β
clips_chunked/original/video/ β one *_inpainted.mp4 per segment, the frames with the human arms painted out |
| Robot trajectory tables | β
ALLEX β clips_chunked/allex/annot/ |
| Unmodified source frames | β never β download Egocentric-10K yourself, and tools/prepare_segments.py cuts it to these segments |
| Robot-rendered frames | β never β the ALLEX meshes are evaluation-only. Stage 9 renders them on your machine |
Robot trees are keyed by robot name, so a second robot arrives as a sibling of allex/ and |
|
| nothing above it changes. ALLEX is the only one published so far. |
Licences differ by file class β see Licence.
What you can do with this
90.6 GiB: 12.2 GiB human annotations, 72.1 GiB arm-removed video, 6.3 GiB Allex trajectory tables. Every path below runs without touching upstream. 62,621 of the 62,678 segments carry an Allex trajectory (99.9%); the other 57 ship their human annotations alone.
No frames ship unmodified, and no frames ship with a robot in them. Every video file here is a segment with the human arms inpainted out and nothing put back in their place.
- The original pixels are not included β another 80 GB that nothing in the pipeline needs, and that you can rebuild in one command because Egocentric-10K is Apache-2.0 and openly downloadable. See Rebuilding the raw segments.
- The Allex-rendered pixels are not included either.
pipeline/stage9_robot_overlay.pyreproduces them from what is here, andstage10_lerobot_convert.pypacks the result as a LeRobot V2.0 dataset. Both need a GPU; nothing else in the pipeline has to run.
| You want | You need | Stages you run |
|---|---|---|
| Hand and camera annotations over arm-removed video | original/ |
none |
| Allex joint trajectories, to train on states or to check ours | allex/annot/ |
none |
| Robotize onto your own robot | original/ |
8 β 9 β 10 |
| The Allex LeRobot dataset the paper trains on | original/ + allex/annot/ |
9 β 10 |
Row three is the point of this release. Because the arm-removed video ships alongside the annotations, retargeting to a different robot skips the whole annotation half of the pipeline and the inpainting β you go straight to IK and rendering. Adding a robot needs one YAML file and a URDF, no code change. See Robotizing onto your own robot.
Row four needs the Allex meshes, which come from the ALLEX model repo and carry their own licence β see ALLEX below. Nothing in this repository depends on them.
Read this first
- Access is gated. Accept the terms on the dataset page and access is granted automatically. The acceptance is recorded; see Licence for what the terms commit you to.
- Arm-removed video is included; original video and robot renders are not. Egocentric-10K is Apache-2.0, so the derived frames may be redistributed β the inpainted frames ship here. The unaltered source frames do not; see Rebuilding the raw segments. Neither does any frame with a robot composited into it.
- Non-commercial anyway. The source licence permits commercial use; the pipeline does not. HaWoR (CC BY-NC-ND 4.0), ProPainter (S-Lab 1.0 non-commercial) and MANO (research only) touched every frame. Read that as the binding constraint, not the Apache-2.0 tag upstream.
- Gated to mirror upstream. Egocentric-10K asks for contact details; so does this.
- States and actions are pipeline estimates, from monocular hand reconstruction and IK, not motion capture. Pre-training corpus, not an evaluation benchmark.
- Compute your own normalisation statistics β stage 10 leaves
meta/stats.jsonout, because they depend on the corpus and resolution mix you train on.
Contents
clips_chunked/
βββ original/annot/<clip_id>/<start>_<end>.parquet human annotations, stages 2-7
βββ original/video/<clip_id>/<start>_<end>_inpainted.mp4 the segment's frames, arms removed (stage 7)
βββ allex/annot/<clip_id>/<start>_<end>.parquet Allex joint trajectory, stage 8
βββ original/video/<clip_id>.done stage 6 marker
βββ original/annot/<clip_id>.done stage 7 marker
βββ allex/annot/<clip_id>.done stage 8 marker
segments.csv clip_id, start, end, fps, resolution, language
Three .done markers per clip, which is what the pipeline reads: stage 7 checks stage 6's
original/video/<clip>.done, stage 8 is satisfied by original/annot/<clip>.done, and stage 9
requires its robot's stage-8 marker, allex/annot/<clip>.done. There is deliberately no
stage-9 marker β stage 9 skips every clip whose marker already exists
(stage9_robot_overlay.py:368-372), so shipping one would make your own render silently produce
nothing.
allex/annot/ carries the full 37-column schema, including overlay_valid β the per-frame flag
saying the Allex projection was usable in that frame when we rendered it. Use it to drop bad
frames before you spend GPU time rendering.
The clips_chunked/ tree is the same layout the
HuRo-Ego4D repository ships, plus the video that
Ego4D's licence forbids there. Every column is documented under
Columns below, and in
examples/README.md.
Rebuilding things yourself
Cutting your own segments from the source corpus, retargeting onto a different robot, rendering it
into the frames, and packing the result as a LeRobot dataset are all documented in
REBUILDING.md in this repository. None of it is needed to use the dataset as
published.
Columns
One row per video frame, in the 37-column schema shared by every HuRo release. Both tables carry
all 37 columns; they differ in which are filled. Read a row with decode_row from common/io.py
β arm_mask is bit-packed and absent per-hand arrays are stored as NaN lists β or run the repo's
reader:
python examples/read_parquet.py clips_chunked
original/annot/<clip_id>/<start>_<end>.parquet β robot-independent
| Column | Type | Meaning |
|---|---|---|
clip_id, frame_id |
string, int32 | Egocentric-10K clip id; frame index, restarting at 0 per segment |
height, width |
int32 | the frame size the annotations were made at |
intr_model |
string | given β Egocentric-10K supplies its own intrinsics, so no calibration stage ran |
fx, fy, cx, cy, xi |
float64 | the camera model as recorded. xi is 0.0 on every row of all 62,678 segments, so nothing here is rectified |
pinhole_fx, pinhole_fy, pinhole_cx, pinhole_cy |
float64 | the pinhole every geometry column is expressed against. Equal to fx/fy/cx/cy throughout this subset, because xi is 0 |
cam_pose |
float32[4,4] | cam-to-world, OpenCV, metric-scaled and gravity-aligned |
arm_mask |
binary, bit-packed (H,W) bool | the human-arm mask stage 7 inpainted. Filled on every row, and the reason you can skip stages 1β6 |
hands[] |
list of struct | one entry per detected hand β see below |
narr |
struct | think, left, right, bimanual β the VLM's per-segment narration |
language |
string | the merged instruction, one per segment |
n_person_det |
int32 | null in this subset |
hands[] fields: box float32[4] [x1,y1,x2,y2] in the rectified frame, conf, side int8
(0 left, 1 right), side_conf, kpts3d float32[21,3] in MediaPipe order and camera frame,
wrist_rot float32[3] axis-angle β null where stage 5 could not resolve the wrist orientation
β finger_rot float32[15,3] axis-angle, and hand_mask, which is null on every row in every
HuRo subset: the per-hand masks were transient inputs to inpainting and were never persisted.
The stage-8 columns listed below are present but null in this table, as are n_person_det and
overlay_valid.
allex/annot/<clip_id>/<start>_<end>.parquet β Allex only
Skip this table if you are targeting your own robot; stage 8 regenerates its equivalent from
original/annot.
| Column | Type | Meaning |
|---|---|---|
state_qpos |
float32[48] | Allex joint angles, ordered by the state_qpos_joint_names metadata |
state_eef_left, state_eef_right |
float32 | wrist position (3) + rotation 6D (6) + hand joint angles |
state_mask_left, state_mask_right |
bool | false where that side's hand is missing |
cam_pose_base |
float32[4,4] | cam-to-robot-base, OpenCV |
retarget_cost, err_ddq, err_tip_mm_*, err_palm_mm_*, err_local_dir_*, err_cam_pos_mm, err_cam_rot_deg |
float32 | IK diagnostics; stage 10 filters episodes on the err_* columns |
overlay_valid |
bool | stage 9's per-frame flag saying the Allex projection was usable. Use it to drop bad frames before you spend GPU time rendering |
Schema metadata, readable via pq.read_table(...).schema.metadata and attached to the allex
tables only: robot_name, state_qpos_joint_names, state_eef_layout
(per_side:wrist_pos3_rot6d6_hand_joints), state_eef_left_hand_joint_names,
state_eef_right_hand_joint_names, state_eef_frame (robot_base_frame), cam_pose_convention
(opencv_cam_to_base).
Positions in metres, angles in radians. Robot base frame is x forward, y left, z up; camera axes are OpenCV.
Other HuRo subsets
| Subset | Repo | Pixels included? |
|---|---|---|
| Egocentric-10K | this repo | arm-removed frames only β no original pixels, no robot rendered |
| EPIC-KITCHENS-100 | Joocjun/HuRo-EPIC-KITCHENS |
arm-removed frames only β no original pixels, no robot rendered |
| Ego4D | Joocjun/HuRo-Ego4D |
annotations only β its licence forbids the video; bring your own copy |
| EgoVerse | Joocjun/HuRo-EgoVerse |
arm-removed frames only β no original pixels, no robot rendered |
| EgoDex | not released β CC BY-NC-ND 4.0 forbids sharing adapted material. regenerating it from your own copy is possible with this pipeline, but the recipe document is not published yet | β |
Licence
This dataset is not a single licence. Three layers stack, and the narrowest one wins for any given file:
| Layer | Sets |
|---|---|
| The source corpus | what may be redistributed of this subset at all β here Apache-2.0, which permits commercial use |
| The pipeline | reaches the arm-removed frames only. They are ProPainter output (S-Lab 1.0, non-commercial), so they ship CC BY-NC 4.0. The tables are not ProPainter output, and a tool's licence is not generally held to reach what the tool produces β so they ship CC BY 4.0 |
| The robot model | the ALLEX URDF is BSD 3-Clause and its notice must be kept β see ALLEX below. No ALLEX-licensed pixels are published, so no separate grant applies to any file here |
Egocentric-10K itself imposes no non-commercial term, so the tables are CC BY 4.0 β free for commercial use with attribution. Only the arm-removed video carries a non-commercial condition, and that is a conservative choice rather than a settled obligation: see Why the tables are CC BY 4.0. If you regenerate the video yourself with a permissively-licensed inpainter, the source licence places no such limit on you.
What that means file by file:
| File | Licence |
|---|---|
| Pipeline code (the HuRo repo) | Apache-2.0 |
Annotation and trajectory tables (*.parquet, segments.csv, *.done) |
CC BY 4.0 |
Arm-removed frames (*_inpainted.mp4) |
CC BY-NC 4.0 |
The reasoning behind these terms β in particular why the licences of the models inside the
pipeline are not held to reach the numbers they produced β is set out in
DATA_LICENSE.md.
ALLEX
Required attribution, to be kept with any redistribution:
ALLEX robot model and rendered appearance Β© 2026 WIRobotics Inc. Used with permission. ALLEX model information: https://github.com/wirobotics-rih/allex_model
The ALLEX model repository's BSD 3-Clause covers
its MJCF, URDF and USD; the meshes carry a separate, evaluation-only licence. This release
publishes nothing derived from the meshes. The Allex tables here come from the URDF's kinematics
β joint angles, end-effector poses and an overlay_valid flag β so BSD 3-Clause governs them, and
retaining the notice above is the whole of the obligation.
No ALLEX mesh, point cloud, depth map, surface normal or rendered image is published. If you render Allex yourself with stage 9, you obtain the meshes from the ALLEX repository directly and their evaluation-only licence applies to you, not through us.
No ALLEX mesh, point cloud, depth map or surface normal is published. What ships is kinematic:
joint angles, end-effector poses and an overlay_valid flag β the robot's configuration, not the
shape of its parts.
Source material
Source video: Egocentric-10K by Build AI, Apache-2.0. Every frame is modified from the original material.
Pipeline components retain their own licences β see THIRD_PARTY_NOTICES.md. Review it before redistributing anything derived from this release.
Citation
@misc{jeong2026huro,
title={HuRo: Robotizing Human Videos for Scalable VLA Pretraining},
author={Jinho Jeong and Se June Joo and Jaehyun Kang and Dongyun Kim and Yena Kim and Hanjung Kim and Seon Joo Kim},
year={2026},
eprint={2609.10706},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2609.10706},
}
- Downloads last month
- 9,186