EgoExplorer: A Vision-Motion-Touch Observatory for Human Loco-Manipulation In Situ
Paper • 2610.18872 • Published • 20
Request access to EgoExplorer
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
To download EgoExplorer, please first fill in the access form: https://forms.gle/2tT2dTxAYXgGXBkK6 (name, affiliation, verified email, and your Hugging Face username). Requests that pass the automatic checks are approved within about a minute and you will get a confirmation email. The button below alone does not grant access. This repository currently hosts a one-hour mini release; to hear about the full collection, tick the waitlist option in the same form.
Log in or Sign Up to review the conditions and access this dataset content.
🔐 Gated dataset — request access before downloadingFill in a short form with your name, affiliation, Google-verified email and Hugging Face username. Requests that pass the automatic check are approved within about a minute, and a confirmation email follows. 📦 Now available: a one-hour mini release — 18 recordings (66 min) spanning all ten task templates, every one with all four streams and refined SMPL-X. The full collection, which keeps growing, is being opened up in stages; tick “Join the waitlist for the full release” in the same form and we will get in touch as more data becomes available.
|
EgoExplorer
A Vision–Motion–Touch Observatory for Human Loco-Manipulation In Situ
| 10 task templates |
5 environments |
4 streams, one clock |
Humanoids are built for human spaces, and almost nothing useful in those spaces happens standing still. A person fetching a reagent walks to the bench, leans in, adjusts their stance so both hands can reach, and keeps the grasp while they turn away. The walking and the manipulation are not two phases. They are one behaviour.
EgoExplorer records that behaviour: a large and steadily growing observatory of human loco-manipulation, captured with a wearable rig light enough to walk between a chemistry bench, a factory bay and a kitchen. Four streams — what the person saw, how the whole body moved, what every finger did, and where the hands actually pressed — are recorded together, on one clock, in rooms we did not instrument.
Egocentric video scales well but leaves the body that produced the view unobserved. Wearable trackers recover the body and usually stop at the wrist. Glove and contact datasets recover the fingers or the touch and usually drop the walking. EgoExplorer keeps the part they all drop: the coupling between the parts.
Contents
- Ten task templates, five rooms
- Four streams, one clock
- Capture rig
- Reconstruction: Calibrate, Complete, Constrain
- Environments
- Data format
- Mini release
- Access
- Quick start
- Comparison with related datasets
- License
- Citation
Ten task templates, five rooms
Six laboratory procedures and four household / industrial tasks. Same rig, same calibration protocol, no instrumentation in the room. Every panel below is a different session.
| Environment | Task templates |
|---|---|
| 🧪 Laboratory | Weighing · Liquid–liquid extraction · Reaction setup · Rotary evaporation · Oven drying · Column chromatography |
| 🏭 Industrial workspace | Folding a cardboard box |
| 📚 Study | Bookshelf organisation |
| 🛋️ Living room | Clothes folding |
| 🥪 Food-preparation area | Sandwich making |
Besides regular episodes, the collection deliberately includes recovery episodes — responses to interruptions and failed attempts — so the data covers how people correct course, not only the clean path.
Four streams, one clock
Every recording carries all four streams at every instant, resampled onto a shared 30 Hz grid.
| Stream | Source | Native rate | What it gives you | |
|---|---|---|---|---|
| A | 🎥 Egocentric video | GoPro HERO13 + Media Mod, rigidly coupled to the head tracker | 30 Hz, 1920×970 | What the head camera saw, with a measured camera pose |
| B | 🧍 Whole-body motion | 5 × VIVE Ultimate Tracker → C³ reconstruction | 60 Hz trackers | Metric SMPL-X body, incl. the 17 joints no tracker sees |
| C | ✋ Finger articulation | MANUS Metagloves Pro or Wuji gloves | ~120 Hz | Both hands, measured by the glove — not inferred from pixels |
| D | 🔥 Tactile contact | Wuji tactile glove, 24×31 taxels per hand | ~120 Hz | Raw taxels + binary contact over 17 anatomical regions per hand |
One recording (weighing), played at 5×: egocentric video, whole-body motion, hand skeleton and tactile contact, and the step-level description, all on one timeline.
Every recording also carries step annotations: step ID, free-text description, interval bounds, skill, object roles and observed outcome, written or checked by domain experts.
Capture rig
Five anchors, two gloves, one camera. Everything mounts on the body in a few minutes, survives a day of real work, and leaves the wearer able to do their actual job. Nothing stands in the room: no tripod, no base station, no line of sight to keep clear.
 |
The head coupling turns the camera pose into a measured quantity instead of something to solve for later. Participant geometry and mount placement are solved once, before any task is recorded — a session costs a donning and a short calibration, not a re-rig. |
Reconstruction: Calibrate, Complete, Constrain
Sparse sensing moves the cost from hardware to reconstruction. A tracker pose is the product of two unknowns — where the device sits on the limb, and how the limb is articulated — and fitting them together quietly turns a mounting error into a pose error. Our pipeline, C³, refuses to solve them at the same time.
- Calibrate. A short prescribed sequence, matched by phase against a reference motion library, gives body shape, metric scale and each sensor-to-anchor transform — fixed before any motion is inferred.
- Complete. A temporal prior (AvatarPoser extended from three anchors to five, trained on AMASS) turns the calibrated anchors into a first whole-body trajectory, supplying the joints no tracker observes.
- Constrain. Three layers refine that trajectory jointly — against the measurements (skeleton), against the body's own surface (clearance), and against the floor (every posed vertex).
Fingers and contact are measured, not estimated: the calibrated wrist transform carries the glove's frame onto the solved arm, so fingers and pressure land on the same body, on the same clock.
One real recording (sandwich making, 30 Hz). The five trailing gizmos are the tracker poses, the body is the solved SMPL-X, and the heat on the hands is the glove's own per-region pressure.
| Body-point error vs. tracker-fitted references | All joints | 17 unobserved joints |
|---|---|---|
| AvatarPoser, retrained on all five anchors | 64.1 mm | — |
| EgoExplorer C³ | 18.0 mm | 20.3 mm |
The near-flat margin between the two columns is the point: a method that merely snapped to its anchors would be tight at the five trackers and loose at the elbows and knees.
Environments
Because the rig carries no room-side infrastructure, each capture space is reconstructed once, and every session in it can be replayed inside its own geometry.
Data format
The release follows the LeRobotDataset v3.0 specification. All streams are aligned to a common 30 Hz grid; each row is one episode timestamp.
EgoExplorer/
├── meta/
│ ├── info.json # fps=30, feature shapes/dtypes, shard paths
│ ├── stats.json # per-feature statistics
│ ├── tasks.parquet # task_index → task description
│ └── episodes/chunk-000/file-000.parquet # row ranges, video offsets, session / split / calibration-profile IDs
├── data/chunk-000/file-000.parquet # frame-level numeric data + text annotations
└── videos/observation.images.ego/chunk-000/file-000.mp4
| Field group | Per-row shape | Contents and conventions |
|---|---|---|
| Index and time | scalars | episode_index, frame_index, index, task_index, timestamp (s, episode-relative) |
| Ego RGB | linked video frame | observation.images.ego, frame matched to the row timestamp |
| Tracker poses | 5×3 ; 5×4 | position (m); quaternion (x, y, z, w); per-device validity |
| Body motion | 3 + 3 + 63 | world pelvis position (m), root orientation, 21 local body-joint rotations (axis-angle, rad) |
| Hand motion | 2×45 | left / right SMPL-X finger rotations, local axis-angle |
| Tactile readings | 2×24×31 | normalized taxel readings, invalid cells masked (Wuji subset only) |
| Regional contact | 2×17 | binary region state per 30 Hz bin, with separate validity |
| Step annotations | ints, text, times | step ID, description, interval bounds, skill, object roles, observed outcome |
| Validity & provenance | arrays, masks | stream support, source-time mapping, interpolation flags, review status |
Conventions. Metric, right-handed, +Y-up world frame. Tracker quaternions are active device-to-world rotations. SMPL-X rotations are local axis-angle in radians; the root orientation is world-relative. The translation field stores the world pelvis position, not the SMPL-X library's additive transl parameter. Body shape, scale and mount parameters are referenced by calibration-profile ID rather than repeated per row. Unavailable streams are marked by validity masks, never encoded as measured zeros.
Temporal alignment. Positions are linearly interpolated, rotations SLERPed, RGB uses the nearest recorded frame. Contact keeps any valid native-rate detection inside each 1/30 s bin (logical OR), so brief touches survive resampling. Interpolated values keep source-time and validity information; they are not new measurements.
Camera. 1920×970, intrinsics fx=1012.0, fy=1016.0, cx=960.0, cy=430.0.
Mini release
What is on this page today is a one-hour preview of EgoExplorer, meant to let you inspect the format and try the data in your own pipeline before the larger release.
| Recordings | 18 episodes · 118,910 frames · 66 min at 30 Hz |
| Tasks | all 10 templates — 6 laboratory procedures (weighing, extraction, reaction setup, rotary evaporation, oven drying, column chromatography) and 4 household / industrial tasks |
| Streams | egocentric video, 5 tracker poses, both hands' 21-joint skeletons, 24×31 Wuji tactile per hand, refined SMPL-X, English step captions |
It is stored as a self-contained LeRobot v3.0 dataset under dataset/. The camera stream is observation.images.camera_0; per-episode SMPL-X parameters, 17-region contact, calibration, step annotations and quality reports live in dataset/episode_meta/episode_XXXXXX/.
More data. The full collection — far larger than this preview and still growing, with both MANUS and Wuji gloves — is being made available in stages, with priority for groups that have a concrete use for it. If you would like to be considered, tick “Join the waitlist for the full release” in the access form — whether you request mini-release access now or come back later — and we will contact you at your verified email.
Access
EgoExplorer is gated. Fill in the access form — full name, affiliation, a Google-verified email and your Hugging Face username. Submissions are checked automatically, and if they pass, access to the mini release is granted to that Hugging Face account within about a minute; a confirmation email follows. The same form has an optional box to join the waitlist for the full release. Then download as usual while signed in:
hf auth login
hf download InternRobotics/EgoExplorer --repo-type dataset --local-dir EgoExplorer
Quick start
from lerobot.datasets.lerobot_dataset import LeRobotDataset
# after `hf download ... --local-dir EgoExplorer` (see Access)
ds = LeRobotDataset("InternRobotics/EgoExplorer", root="EgoExplorer/dataset")
print(ds.meta.info["fps"], ds.num_episodes, ds.num_frames) # 30 18 118910
frame = ds[0]
frame["observation.images.camera_0"] # (3, H, W) egocentric RGB
Or read the Parquet shards directly:
from datasets import load_dataset
ds = load_dataset("InternRobotics/EgoExplorer", split="train", streaming=True)
print(next(iter(ds)).keys())
Comparison with related datasets
| Dataset | Hours | Metric full body | Non-visual body anchors | Feet observed | Articulated hands | Touch | Mobile capture | Loco-manip. | Step labels |
|---|---|---|---|---|---|---|---|---|---|
| Ego4D | 3,670 | ✗ | ✗ | ✗ | ✗ | ✗ | ✓ | ✗ | ✓ |
| Ego-Exo4D | 221 | ◐ | ✗ | ✗ | ◐ | ✗ | ✓ | ✓ | ✓ |
| Nymeria | 300 | ✓ | ✓ | ✓ | ✗ | ✗ | ✓ | ✓ | ✓ |
| EgoDex | 829 | ◐ | ✗ | ✗ | ◐ | ✗ | ✗ | ✗ | ✗ |
| EgoTouch | 20 | ✗ | ◐ | ✗ | ✓ | ✓ | ✗ | ✗ | ✗ |
| OpenTouch | 5 | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ | ✗ | ✓ |
| EgoExplorer | growing | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
✓ sensor-grounded / full support · ◐ inferred from images or partial · ✗ absent. See the paper for the full table and definitions.
License
EgoExplorer is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Commercial use is not permitted.
Citation
If EgoExplorer is useful in your work, please cite:
@misc{zhou2026egoexplorer,
title={EgoExplorer: A Vision-Motion-Touch Observatory for Human Loco-Manipulation In Situ},
author={Yunsong Zhou and Xuekun Jiang and Haiqi Shen and Hangxu Liu and Chaojian Shi and Yujie Liu and Jingyuan Li and Haoran Yang and Jiafei Cao and Dongzhan Zhou and Dong Wang and Yao Mu and Chunhua Shen and Weinan Zhang},
year={2026},
eprint={2610.18872},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2610.18872},
}
- Downloads last month
- 25