Grounded Superintelligence will use the information you provide to evaluate and manage dataset access, understand the RoboCap-10k community, and contact you about your request. RoboCap-10k is released under CC BY 4.0, which permits research and commercial use with attribution.

RoboCap-10k: A Metrically Accurate Egocentric Dataset for Robot Learning

10,000 curated hours of unscripted egocentric manipulation, recorded while people did their paid jobs, with metric 3D head and hand poses and two levels of language annotations.

Data Bucket · Paper · Grounded Superintelligence

Data location: RoboCap-10k is distributed through the groundedsi/RoboCap-10k Bucket. This repository holds the dataset card and the access request form; download the data from the Bucket. Once your request is approved, you can read from the Bucket with your Hugging Face account.

Overview

RoboCap-10k is a single-device, egocentric corpus of real occupational work for training robot policies and world models. It contains 10,000 hours (1,363,915 episodes) across 21 industry verticals, 365 site types, and 619 distinct tasks, recorded in 11 countries, including India, Nigeria, Colombia, the United States, the United Kingdom, Mexico, Chile and Spain.

Every hour was recorded on the same device, the RoboCap: a 250 g hat with six synchronized global-shutter cameras and two IMUs. Every hour was also processed by the same 3D pipeline, so there is no domain shift between devices in either the images or the quality of the 3D labels.


Hours	10,000 (curated; wall-clock capture time)
Episodes	1,363,915 (10–600 s; median 17.8 s, mean 26.4 s)
Synchronized timesteps	~1.08 × 10⁹
Camera frames	~6.5 × 10⁹ (6 streams)
Industry verticals / site types / tasks	21 / 365 / 619
Skill groups	14
Devices	214
Video size	~108 TB (H.265)
Audio	None recorded

What makes it different

Calibrated stereo on every hour. Each recording session is recalibrated online, which brings every released device below 1 px of average Sampson error (median 0.63 px on the front pair, 0.60 px on the eye pair).
Metric 6-DoF head pose. Head trajectories are gravity-aligned and come from multi-camera visual-inertial SLAM with offline bundle adjustment. Median ATE against motion capture is 3.0 mm.
World-frame 3D hands. Both hands have MANO parameters and 21 keypoints from a multi-view hand tracker. PA-MPJPE against motion capture is 4.5 mm.
Two levels of language. Each interval has one sparse caption, and dense sub-task captions are segmented in time within it.
Real occupational work. Examples include bread production, industrial garment stitching, engine and motorcycle repair, PCB soldering, rebar tying and onion harvesting.
Raw streams included. The synchronized raw streams are released alongside the labels, so the labels can be audited and regenerated as methods improve.

Capture device: RoboCap

Subsystem	Optics / FoV	Baseline / Overlap	Rate / Sync
Front stereo	Fisheye, 180° × 112°	86.8 mm (157° overlap)	1080p @ 30 Hz, HW trigger
Eye stereo	Fisheye, 180° × 112°	106.2 mm (154° overlap)	1080p @ 30 Hz, HW trigger
Lateral (2×)	Wide-angle, 161° × 98°	85° off-axis, 30° down	1080p @ 30 Hz, HW trigger
IMUs (2×)	6-axis, crown sides	Cam–IMU ≤ 3 ms	200 Hz
Clock	24 MHz TCXO	1-to-6 PWM	±1.81 ppm

All cameras are global-shutter, 10-bit RGB, with f ≈ 590–630 px. One PWM signal triggers all six cameras, so their frames are captured at the same instant.

Data format

Every episode comes in two formats.

1. LeRobot format (for training)

Each camera is an H.265-encoded MP4.
Per-frame data is stored in Parquet tables indexed by frame:
- SLAM head pose (6-DoF, metric, gravity-aligned world frame)
- MANO parameters and 21 keypoints (16 joints + 5 fingertips) for both hands, in the world frame
- IMU readings

2. Rerun recording (.rrd, for inspection)

Shows the six camera streams, the head trajectory and the hand meshes in one shared 3D world frame on a common timeline.

Each episode also includes:

Factory and online-corrected intrinsics and extrinsics
A JSON sidecar with taxonomy labels, sparse and dense captions, and per-session quality statistics

All streams share one device clock, so video frames, poses and captions line up frame for frame.

Data access

The full dataset (~108 TB) is hosted in a Hugging Face Storage Bucket: hf://buckets/groundedsi/RoboCap-10k.

1. Request access with the form at the top of this page. After your request is approved, install and authenticate the Hugging Face CLI with the account you used to request access:

pip install -U huggingface_hub lerobot
hf auth login

2. Inspect the available prefixes:

hf buckets list groundedsi/RoboCap-10k -h -R

The Bucket is large and may be updated over time, so its current file listing is the source of truth for published prefixes.

3. Stream the full dataset (recommended):

from lerobot.datasets import StreamingLeRobotDataset

ds = StreamingLeRobotDataset("groundedsi/RoboCap-10k", repo_type="bucket")
for frame in ds:
    ...

Train with streaming:

lerobot-train \
  --dataset.repo_id=groundedsi/RoboCap-10k \
  --dataset.repo_type=bucket \
  --dataset.streaming=true \
  ...

4. Download a selected directory or a single file:

# Replace PREFIX and LOCAL_DIR with paths chosen from the listing
hf buckets sync hf://buckets/groundedsi/RoboCap-10k/PREFIX ./LOCAL_DIR

# Replace PATH/TO/FILE and FILE with the selected object path and filename
hf buckets cp hf://buckets/groundedsi/RoboCap-10k/PATH/TO/FILE ./FILE

To work with a downloaded directory locally, point LeRobotDataset at it with the root argument. See the Hugging Face Bucket guide for filtering, mounting, Python access and sync options.

Language annotations

Tier	Granularity	Content
Sparse caption	One per interval	A single free-text sentence describing the whole clip
Dense units	Many per interval, non-overlapping	`scene`, `high_level_task`, `low_level_task`, joined into a three-sentence `full_caption`

Example interval:

Span	Tier	Caption
0.00–14.80 s	Sparse	Tears a piece of paper towel from a roll, folds it, and wipes a glass beaker.
0.00–3.50 s	Dense	Tears a section of paper towel from the roll.
3.50–9.50 s	Dense	Folds the torn paper towel sheet.
9.50–14.80 s	Dense	Wipes the interior of a small clear container with the folded paper.

Fields stored with each dense unit:

Category	Fields
Links	`unit_id`, `parent_id`, `episode_id`
Timing	start/end frames and clip-relative seconds, `continuation_of_previous`
Text	`scene`, `high_level_task`, `low_level_task`, `full_caption`
Grounding evidence	`evidence_frames`, `evidence_views`, `grounding_note`, `review_warnings`
Provenance	`origin`, `offline_edited`, `human_visual_review`

Dense units are generated by Gemini 3.8 Flash from the two front-stereo views. The model is not given the sparse caption, so the two tiers are produced independently. Human annotators visually reviewed the captions for 1,000 hours (10% of the corpus) against the source video, and those reviewed labels were used to fit the captioning pipeline.

Taxonomy

The dataset uses a three-level taxonomy: industry vertical → site type → task. Each task also has a difficulty tier and one of 14 manual skill groups.

Difficulty	Tasks	Description	Examples
Easy	170	Single step, no training needed	Sorting groundnuts, picking products from shelves
Medium	277	Multi-step, needs coordination or planning	Shaping dough into loaves, portioning meals into trays
Hard	172	Skilled work that requires training or expertise	Soldering PCB components, tying rebar cages, weaving cane, industrial stitching

Verticals by number of distinct tasks:

Industry vertical	Site types	Tasks
Food & Beverage	39	117
Furniture & Woodworking	26	68
Automotive & Transport	36	66
Industrial Manufacturing	39	62
Agriculture & Farming	32	56
Clothing & Fashion	34	49
Construction & Hardware	33	43
Household / Domestic	15	25
Logistics & Warehousing	9	22
Retail & Consumer Goods	20	20
Creative Workshops & Crafts	13	20
Printing, Packaging & Design	14	18
Beauty & Personal Care	13	12
Cleaning, Laundry & Sanitation	13	11
Repair Services	9	11
Healthcare & Pharmacy	6	4
Laboratory / Scientific	4	4
Energy & Utilities	3	3
Administrative & Office	3	2
Hospitality & Lodging	3	0*
Sports & Recreation	2	0*
Other / Mixed Trades	5	6
Total (distinct)	365	619

* Footage for these verticals is on file and waiting for task labels.

Annotation quality

Component	Benchmark	Result
Calibration	Held-out Sampson error, 214 devices	Median 0.63 px (front) / 0.60 px (eye); all < 1 px
Head trajectory	RoboMoCap (153.4 h, 4,669 trajectories, NOKOV mocap @ 120 Hz)	Median ATE 3.0 mm, mean 3.19 mm; scale error 0.17% median
SLAM	LaMAria (63 city-scale walks, 27.5 h)	Control-point score 72.4 (Aria SLAM: 75.9)
Hand pose	RoboMoCap held-out test split (9.2 min)	PA-MPJPE 4.5 mm over 21 keypoints
Face detection	1,336 held-out faces @ 0.05 FP/frame	77.3% recall / 92.4% precision (IoU 0.3)

Two devices whose front-pair calibration error averaged above 1 px were excluded from the release, which removed 34 hours (0.6%) of footage.

Curation

An episode is one continuous clip of at least 10 s in which one person performs one recognizable task in one setting, both hands are tracked, and SLAM is available. An episode is kept only if it:

shows active bimanual manipulation with an identifiable task,
passes the calibration, SLAM and hand-tracking quality checks,
has complete, consistently timed camera streams, and
is not a duplicate of another episode by file fingerprint, content signature or recording metadata.

Trained annotators segment the recordings into episodes following written guidelines, and a second person reviews every segmentation. Taxonomy labels come from both the recording metadata and a frame-level review of the footage done without seeing that metadata. Where the two disagree, the footage review wins.

Privacy and ethics

Consent. Every recorded operator signed an informed consent form, and every partner enterprise signed a data agreement. Participation was voluntary.
Compensation. Wearers are paid on top of their regular wages. The footage is never used to evaluate individual workers' performance.
No audio was recorded.
Face anonymization. A custom face detector, trained on about 1,000 hours of human-labeled RoboCap footage, runs on every frame of all six streams. Detections are tracked, propagated forward and backward, dilated by 15%, and blurred with a Gaussian blur. The blur is permanent: released videos are re-encoded from the blurred frames, and the originals are not distributed.
Minimal metadata. The released metadata identifies only the country and site type, never the site, employer or participant.

Limitations

Labels are estimates, not ground truth. All 3D annotations are model outputs. Hand-pose errors concentrate where fingers are hidden at the point of contact during tight grasps.
Coverage and selection bias. Collection is concentrated in a few countries, and India accounts for most verticals. Curation keeps only bimanual work that passes the 3D quality checks, so one-handed work and the hardest tracking conditions (heavy occlusion, fast motion, poor lighting) are underrepresented. Household and personal activities are underrepresented, and task counts are uneven across verticals.
No tactile or force sensing. Contact must be inferred from vision alone.
Embodiment gap. Retargeting human hands to robot end-effectors has to handle joint limits and singularities. The human arms visible in the images also differ from robot arms.

License

Released under CC BY 4.0. You may use, share and adapt the data for any purpose, including commercial use, as long as you give appropriate credit by citing the paper below.

Citation

@article{hughes2026robocap10k,
  title   = {RoboCap-10k: A Metrically Accurate Egocentric Dataset for Robot Learning},
  author  = {Hughes, Kheri and Ali, Qasim and Bhardwaj, Nikash and Akhibi, Owen and
             Dravin, Niresh and Sharma, Vivek and Cho, Sam and Lee, Ethan and Luo, Paul and
             Yu, Sunny and Victor, Jonathan and Young, Amanda and Cho, Michael and
             Adeniji, Ademi and Liu, Vincent},
  year    = {2026}
}

For details on the hardware and processing pipeline, see the companion report:

@article{ali2026robocap,
  title   = {RoboCap: A New Platform for Egocentric Robot Learning},
  author  = {Ali, Qasim and Bhardwaj, Nikash and Hughes, Kheri and Cho, Sam and Lee, Ethan and
             Luo, Paul and Chong, Shinn and Yu, Sunny and Akhibi, Owen and Proshkin, Alex and
             Zhan, Haotian and Dravin, Niresh and Sharma, Vivek and Victor, Jonathan and
             Young, Amanda and Cho, Michael and Adeniji, Ademi and Liu, Vincent},
  year    = {2026}
}

Downloads last month: 2

Total file size:

17.4 kB