HuRo
Collection
Robotized egocentric human video. One repo per source dataset, since their licences differ. • 4 items • Updated
Datasets:
Accept the terms to access HuRo-EgoVerse
Access is granted automatically once the form is submitted.
This repository holds two kinds of file under two different licences.
The Parquet tables — hand keypoints, camera parameters and poses, language instructions and retargeted ALLEX joint trajectories — are CC BY 4.0. No non-commercial component contributed to them.
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 EgoVerse frames are included. The source episodes are distributed by the Georgia Tech RL2 lab under MIT from a credentialed bucket; obtain them yourself if you need the originals.
Your username, email address and the fields below are collected solely to administer access to this dataset. 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-EgoVerse
The EgoVerse share of the HuRo dataset — 193,499 segments / 79,006,018 frames across 7,562 clips, 731.5 hours at 30 fps: annotation and ALLEX trajectory tables, plus the arm-removed video.
Per frame: camera intrinsics and metric-scaled poses, 3D hand keypoints with MANO-convention rotations, per-segment language instructions, the retargeted ALLEX joint trajectories, and the frames with the human arms inpainted away.
Produced by the HuRo robotization pipeline. Paper: arXiv:2609.10706 · Project page: 3587jjh.github.io/HuRo
Two licences in one repository. The tables are CC BY 4.0; the arm-removed video is CC BY-NC 4.0. This is the only HuRo subset whose tables carry no non-commercial term, and the reason is in Licence.
What's in this repository
The repository name says which corpus this came from, not what is inside. What is here:
| Human annotation tables | ✅ clips_chunked/original/annot/ — one Parquet table per segment, robot-independent |
| Arm-removed video | ✅ clips_chunked/original/video/ — one *_inpainted.mp4 per segment, the frames with the human arms painted out |
| Robot trajectory tables | ✅ ALLEX — clips_chunked/allex/annot/ |
| Unmodified source frames | ❌ never — download EgoVerse yourself, and tools/prepare_segments.py cuts it to these segments |
| Robot-rendered frames | ❌ never — 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 — see Licence.
What you can do with this
| You want | You need | Stages you run |
|---|---|---|
| Hand and camera annotations | this repo | none |
| Arm-removed video for robot-video pretraining | this repo | none |
| Retarget onto your own robot | this repo | 8 |
| Render your robot into the frames | this repo + your robot's URDF | 8 → 9 |
| A LeRobot dataset | this repo + your robot's URDF | 8 → 9 → 10 |
Because the arm-removed video ships, you skip the entire annotation half of the pipeline — camera calibration, hand reconstruction, SLAM, segmentation, inpainting — and go straight to retargeting and rendering. Adding a robot needs one YAML file and a URDF, no code change.
Read this first
- Access is gated. Accept the terms on the dataset page and access is granted automatically.
- No unmodified EgoVerse frames are included. Only the inpainted ones. If you need the originals, get the episodes yourself — see Getting the source episodes.
arm_maskis null on every row. The masks were computed on the fly during upstream inpainting and never persisted. You do not need them: the inpainted video is already here. If you want them anyway, they are reproducible — see Rebuilding the arm masks.- The hand keypoints are EgoVerse's own, not ours. Unlike every other HuRo subset, stage 4
(HaWoR) was not run here — EgoVerse supplies 3D hand tracking and camera intrinsics, and the
pipeline consumes them as given.
wrist_rotandfinger_rotare computed geometrically from those keypoints using the 16-joint MANO kinematic tree convention; the MANO model itself is not used. This is why the tables are free of HaWoR (CC BY-NC-ND 4.0) and MANO (research only). box,confandside_confare derived, not measured. EgoVerse ships no hand boxes (every source box is[0,0,0,0]) and no confidences (every value is1.0). We drop those placeholders:boxis recomputed as the 2D extent of the projected keypoints, andconf/side_confarenull. Do not read them as detector scores.- These are pipeline pseudo-labels, not ground truth. Camera poses, retargeting and instructions are all estimated. Use them for pre-training, not as an evaluation reference.
Contents
clips_chunked/
├── original/annot/<clip_id>/<start>_<end>.parquet human annotations CC BY 4.0
├── original/video/<clip_id>/<start>_<end>_inpainted.mp4 arm-removed video CC BY-NC 4.0
├── original/video/<clip_id>.done stage 6 marker
├── original/annot/<clip_id>.done stage 7 marker
├── allex/annot/<clip_id>/<start>_<end>.parquet ALLEX trajectory CC BY 4.0
└── allex/annot/<clip_id>.done stage 8 marker
segments.csv clip_id, start, end, fps, resolution, language
DATA_LICENSE.md
<clip_id> is the EgoVerse episode id — egoverse-i_<UTC recording start>, the episode hash
defined in EgoVerse's CONTRIBUTING_DATA.md. <start> and <end> are inclusive frame indices
into the 30 fps re-encode of that episode.
clips_chunked/original/video/<stem>.mp4 — the unmodified segment — is absent; only
<stem>_inpainted.mp4 ships. There is no allex/overlay/ tree: the stage-9 render is derived from
ALLEX meshes that are separately licensed and evaluation-only, so no HuRo repository ships one. The
overlay table differed from the stage-8 table in one column, overlay_valid, which is carried in
allex/annot/ instead.
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 — stage 9 skips every clip whose marker already exists, so shipping one would make your own
render silently produce nothing.
| Count | |
|---|---|
| Clips with human annotations | 7,562 |
| Clips with an ALLEX trajectory | 7,508 |
| Human segment tables | 193,499 |
| ALLEX segment tables | 182,888 |
| Arm-removed videos | 193,499 |
| Frame resolution | 240×320, every segment |
xi (distortion term) |
0.0 throughout — the frames are already rectified upstream, so fx/fy/cx/cy and pinhole_* agree |
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,
or run the repo's reader:
python examples/read_parquet.py clips_chunked
original/annot/<clip_id>/<start>_<end>.parquet — robot-independent
| Column | Type | Meaning |
|---|---|---|
clip_id, frame_id |
string, int32 | episode id; frame index, restarting at 0 per segment |
height, width |
int32 | frame size the annotations were made at — 240×320 throughout |
intr_model |
string | given — EgoVerse supplies its own intrinsics |
fx, fy, cx, cy, xi |
float64 | the camera model as recorded. xi is 0.0 for every EgoVerse segment |
pinhole_fx, pinhole_fy, pinhole_cx, pinhole_cy |
float64 | the rectified pinhole the annotations were computed against. Equal to the above here, because xi is 0 |
cam_pose |
float32[4,4] | cam-to-world, OpenCV, metric-scaled and gravity-aligned |
hands[] |
list of struct | one entry per hand — see below |
arm_mask |
binary | null on every row in this subset |
narr |
struct | think, left, right, bimanual — the 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] derived from the keypoints, conf null,
side int8 (0 left, 1 right), side_conf null, kpts3d float32[21,3] in MediaPipe order and
camera frame, wrist_rot float32[3] axis-angle, finger_rot float32[15,3] axis-angle,
hand_mask null.
The stage-8 columns listed below are present but null in this table.
allex/annot/<clip_id>/<start>_<end>.parquet — 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 |
overlay_valid |
bool | carried from stage 9, which is not shipped |
Schema metadata, readable via pq.read_table(...).schema.metadata on the allex tables:
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.
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.
Other HuRo subsets
| Subset | Tables | Arm-removed video |
|---|---|---|
| EPIC-KITCHENS-100 | CC BY-NC 4.0 | ✅ CC BY-NC 4.0 |
| Egocentric-10K | CC BY 4.0 | ✅ CC BY-NC 4.0 |
| Ego4D | CC BY 4.0, with arm_mask |
❌ — bring your own Ego4D licence and run stage 7 |
| EgoVerse | this repository, CC BY 4.0 | ✅ CC BY-NC 4.0 |
| EgoDex | not released in any form — CC BY-NC-ND 4.0 forbids sharing adapted material | — |
Licence
Two licences, by file type. The full text is in DATA_LICENSE.md.
| Files | Licence | Why |
|---|---|---|
**/*.parquet, segments.csv, *.done |
CC BY 4.0 | Nothing non-commercial contributed to them |
clips_chunked/original/video/**.mp4 |
CC BY-NC 4.0 | The frames are ProPainter output, and ProPainter's S-Lab 1.0 licence is non-commercial |
The source, EgoVerse, is MIT — confirmed in writing by the EgoVerse maintainers (Georgia Tech RL2) on 2026-11-08: "All of EgoVerse uses MIT license so you can redistribute it as long as you provide the citation / attribution." MIT imposes no non-commercial and no ShareAlike term, so the only inbound obligation on the tables is attribution.
Required attribution, which you must keep on any redistribution:
Derived from EgoVerse, GaTech RL2, https://github.com/GaTech-RL2/EgoVerse, licensed MIT (EgoMimic-Team). Material was modified. Cite the EgoVerse paper.
The reasoning behind these terms — in particular why the licences of the models inside the
pipeline are not held to reach the numbers they produced — is set out in
DATA_LICENSE.md.
ALLEX
The ALLEX trajectories are joint angles solved against WiRobotics' ALLEX URDF, published under
BSD 3-Clause. Retain that notice when redistributing the allex/ tables. No ALLEX mesh, texture or
rendered pixel is included in this repository; the mesh licence is separate and evaluation-only.
Source material
EgoVerse is an aggregate. Its episodes were contributed by Meta Project Aria, Mecka AI, Scale AI and Lightwheel and registered by the Georgia Tech RL2 lab. We rely on the maintainers' statement that the dataset is distributed under MIT.
Citation
@article{huro2026,
title = {HuRo: Robotizing Human Videos for Scalable VLA Pretraining},
journal = {arXiv preprint arXiv:2609.10706},
year = {2026}
}
- Downloads last month
- 2